深度学习或强化学习用于CT出束角度规划问题

Radiotherapy treatment planning (TP) aims to maximize radiation dose delivered to tumors while minimizing exposure to surrounding healthy tissues. Beam angle optimization (BAO) is a crucial component of TP, characterized by high dimensionality and non‐convexity, and is traditionally solved via heuristic or manual iterative approaches. These conventional methods are time‐consuming and often yield suboptimal solutions due to incomplete exploration of the vast solution space.

This study proposes an automated approach to radiotherapy treatment planning by integrating a reinforcement-learning-style iterative framework with a multimodal Large Language Model (LLM). We specifically investigate the problem of Beam Angle Optimization, a high-dimensional and non-convex subproblem of Treatment Planning. Our system employs GPT-4V to select candidate beam angles and analyze three-dimensional dose distributions generated by Monte Carlo simulations within the MatRAD environment. Iterative plan refinement is guided by a reward function that encourages target dose conformity and penalizes excessive dose to organs at risk. We incorporate exploration-exploitation principles to strike a balance between investigating diverse action proposals and refining promising solutions. Experimental results on prostate cancer cases demonstrate that our LLM-based framework offers superior performance compared to random beam selection and can outperform the quality of deep reinforcement learning baselines, indicating the potential for LLMs to assist in complex radiotherapy treatment planning tasks.Clinical relevance—This approach is designed to alleviate the significant effort of manual treatment planning by assisting medical physicists in exploring beam configurations and systematically refining plans to improve dose coverage and protect healthy tissues.

Abstract Background Deep learning (DL)‐based auto‐planning has emerged as a powerful tool for optimizing radiotherapy treatment plans, reducing variability, and improving efficiency. However, current approaches often rely on predefined beam angles and arc spans, which may not be optimal for individual patients. Automated beam angle optimization can further enhance plan quality, particularly in early‐stage breast cancer radiotherapy, where precise beam configurations are crucial for balancing target coverage and organ‐at‐risk (OAR) sparing. Purpose This study presents an automated segment reduction‐based beam angle optimization technique to improve DL‐based auto‐planning for radiotherapy in early‐stage breast cancer. The method optimizes arc spans for volumetric‐modulated‐arc‐therapy (VMAT) and beam configurations for intensity‐modulated‐radiation‐therapy (IMRT) to improve dose distribution while reducing OAR exposure. Methods Plans using three different irradiation strategies—partial arc VMAT (PA‐VMAT), complex IMRT (C‐IMRT), and simple IMRT (S‐IMRT)—were generated using two full arcs for dose mimicking of the predicted dose, followed by the segment reduction performed using a stepwise PAMU (Product of segment Area and Monitor Units) thresholding approach to determine optimal arc spans and beam angles. These strategies were compared against the standard continuous partial arc VMAT (CPA‐VMAT) technique currently used in our clinical practice. Twenty left‐sided breast cancer patients treated under deep inspiration breath‐hold (DIBH) conditions were included for evaluation. Plan quality was assessed using dosimetric criteria, conformity indices, dose mimicking index (DMI), and statistical comparisons. Results PA‐VMAT exhibited superior OAR sparing and the best overall dose mimicking performance, reducing the heart, left lung, and right lung mean doses by 27%, 11%, and 50%, respectively, compared to CPA‐VMAT. C‐IMRT provided the best target coverage but required higher monitor units, while S‐IMRT showed suboptimal dose homogeneity. The automated segment reduction method significantly improved plan efficiency, optimizing beam angles without requiring manual intervention. Conclusion This study demonstrates the feasibility of an automated segment reduction‐based optimization technique for DL auto‐planning in early‐stage breast cancer. PA‐VMAT emerged as the preferred strategy, balancing plan quality, delivery efficiency, and OAR sparing. The proposed approach enhances treatment planning flexibility and will be incorporated into future clinical practice.

No abstract available

Objective. A good boron neutron capture therapy (BNCT) treatment plan, which can deliver higher tumor dose and lower doses to organs at risk (OAR), critically depends on the accuracy of dose prediction and optimization strategy. Existing clinical treatment planning systems mainly use Monte Carlo (MC) simulations. These simulations offer high dosimetric accuracy but are computationally costly and compromise planning efficiency. To overcome this limitation, we aim to develop an improved neural network model that can efficiently and accurately predict BNCT dose distributions under different beam angles, thereby facilitating treatment plan optimization. Approach. We propose a deep learning framework that integrates dose prediction with Bayesian optimization (BO) for beam angle selection. A 3D Vision Transformer backbone captures long-range spatial dependencies, while a Mamba module enhances local feature extraction. A region of interest–guided attention mechanism further directs the model’s focus toward gross tumor volume (GTV) and skin. Predicted doses are incorporated into BO to identify the optimal beam conditions. Main results. On a clinical dataset, the proposed model achieved a mean absolute error (MAE) below 0.6 Gy and mean absolute percentage error (MAPE) below 2% for GTV; for skin, MAE was under 0.15 Gy and MAPE below 3.5%. The average gamma passing rates exceeded 90% (2 mm/2%) and 97% (3 mm/3%). After optimization, the minimum voxel dose of the GTV increased by an average of 1.8 Gy, while the maximum voxel dose of the OARs did not increase. Significance. The proposed method has accurate dose prediction and efficient optimization ability, with results validated by MC simulations. It offers a potential application for clinical automated BNCT treatment planning design and optimization.

Proton pencil beam scanning (PBS) treatment planning for head and neck (H&N) cancers is a time‐consuming and experience‐demanding task where a large number of potentially conflicting planning objectives are involved. Deep reinforcement learning (DRL) has recently been introduced to the planning processes of intensity‐modulated radiation therapy (IMRT) and brachytherapy for prostate, lung, and cervical cancers. However, existing DRL planning models are built upon the Q‐learning framework and rely on weighted linear combinations of clinical metrics for reward calculation. These approaches suffer from poor scalability and flexibility, that is, they are only capable of adjusting a limited number of planning objectives in discrete action spaces and therefore fail to generalize to more complex planning problems.

Objective. To develop and evaluate a deep reinforcement learning (RL) framework for rapid and automatic machine parameter optimization of volumetric modulated arc therapy (VMAT) treatment plans for localized prostate cancer. Approach. A multi-task policy network combining convolution and long short-term memory was trained to sequentially predict the set of actions on the dose rate and multi-leaf collimator positions over the range of two arcs. The network uses as input the cumulative dose grid at the current gantry angle, contours of the planning target volume (PTV) and organs at risk, and the set of machine parameters at all preceding gantry angles. The method was evaluated on a set of 15 localized prostate cancer patients for a prescription dose of 60 Gy in 20 fractions. For each case, the final state dose distribution was compared against clinical plans. For seamless integration with the clinical workflow, the proposed model was integrated into a clinical treatment planning system (TPS), enabling dosimetric review and final plan adjustments. Main results. The RL framework produced deliverable dual-arc VMAT plans in an average of 20.7 ± 5.0 s over the test set. Dosimetric comparison to clinical plans showed no statistically significant differences for the mean rectum dose as well as for the bladder V6160 Gy, indicating that the RL model was as efficient in sparing these structures as human planners. While the approach showed limitations in terms of PTV coverage and maximum body dose, our proposed integration to TPS showed the RL plans could be automatically refined to clinical quality in an additional 83.8 ± 7.2 s. Significance. The accuracy and fast run time of the approach show the potential of the framework to significantly streamline VMAT treatment planning and enable adaptive radiation therapy.

Objective. Despite the high-quality treatment, the long treatment time of the Cyberknife system is believed to be a drawback. The high flexibility of its robotic arm requires meticulous path-finding algorithms to deliver the prescribed dose in the shortest time. Approach. We proposed a Deep Q-learning based on Graph Neural Networks to find the subset of the beams and the order to traverse them. A complex reward function is defined to minimize the distance covered by the robotic arm while avoiding the selection of close beams. Individual beam scores are also generated based on their effect on the beam intensity and are incorporated in the reward function. Main results. The performance of the presented method is evaluated on three clinical cases suffering from lung cancer. Applying this approach leads to an average of 35% reduction in the treatment time while delivering the prescribed dose provided by the physicians. Significance. Shorter treatment times result in a better treatment experience for individual patients, reduces discomfort and the sides effects of inadvertent movements for them. Additionally, it creates the opportunity to treat a higher number of patients in a given time period at the radiation therapy centers.

To automatically identify optimal beam angles for proton therapy configured with the double-scattering delivery technique, a beam angle optimization method based on a convolutional neural network (BAODS-Net) is proposed. Fifty liver plans were used for training in BAODS-Net. To generate a sequence of input data, 25 rays on the eye view of the beam were determined per angle. Each ray collects nine features, including the normalized Hounsfield unit and the position information of eight structures per 2° of gantry angle. The outputs are a set of beam angle ranking scores (S beam) ranging from 0° to 359°, with a step size of 1°. Based on these input and output designs, BAODS-Net consists of eight convolution layers and four fully connected layers. To evaluate the plan qualities of deep-learning, equi-spaced, and clinical plans, we compared the performances of three types of loss functions and performed K-fold cross-validation (K = 5). For statistical analysis, the volumes V27Gy and V30Gy as well as the mean, minimum, and maximum doses were calculated for organs-at-risk by using a paired-samples t-test. As a result, smooth-L1 loss showed the best optimization performance. At the end of the training procedure, the mean squared errors between the reference and predicted S beam were 0.031, 0.011, and 0.004 for L1, L2, and smooth-L1 loss, respectively. In terms of the plan quality, statistically, PlanBAO has no significant difference from PlanClinic (P >.05). In our test, a deep-learning based beam angle optimization method for proton double-scattering treatments was developed and verified. Using Eclipse API and BAODS-Net, a plan with clinically acceptable quality was created within 5 min.

PURPOSE/OBJECTIVE(S) Conventional intensity modulated radiation therapy (IMRT) with a typical 5-20 fixed beams often does not provide sufficient angular sampling required for conformal dose shaping, whereas current volumetric modulated arc therapy (VMAT) discretizes the angular space into equally spaced control points without considering the differential need for intensity modulation of different angles, leading to undersampling at some angles while oversampling at some other angles. Our goal is to develop a node or station parameter optimized radiation therapy (SPORT) strategy with simultaneously optimized angular sampling and beam modulation by leveraging state-of-the-art reinforcement learning and the unique capability of modern digital LINACs in dose delivery through programmable nodal points. MATERIALS/METHODS We developed a SPORT optimization framework, in which, the process of programming control points (or station parameters) was formulated as a stochastic dynamic programming problem, which was solved by a reinforcement learning-based algorithm. On-policy reinforcement learning method, namely, state-action-reward-state-action (SARSA) was integrated with deep convolutional neural network to predict station parameters by utilizing the patient's anatomical structures meanwhile considering the delivery capability of a typical digital LINAC machine. Here, the deep convolutional neural network estimated the state-action value by using the quality of the plan with current station parameters when a next potential station parameter was selected. The state-action value was then updated by SARSA learning. The quality of the plan was quantified by dosimetry constraints. The model was assessed by a retrospective study on a cohort of patients underwent head-and-neck radiation therapy. Dosimetric analysis and delivery efficiency comparisons were used to evaluate the performance of the proposed framework. RESULTS Our model was used to generate 16 plans unseen in the original training set. All the plans predicted by our model achieved better dose distributions without violating clinical planning constraints. Moreover, instead of using 4 full standard arcs in the original clinically used plans obtained via manual optimization, the predicted plans only used one full standard arc (about 178 control points) plus boost from a few sub-arcs (less than 30 degrees of gantry angles), which significantly improved the efficiency of the beam delivery. We are in the process of integrating the sub-arcs into the full arc by considering the programmable capability of modern LINACs. CONCLUSION We demonstrated that a machine learning-based SPORT framework capable of optimizing the spatial sampling and beam modulation simultaneously for modern radiation therapy. The framework not only significantly improves the quality and efficiency of beam delivery, but also has the potential to be incorporated into current clinical workflow to improve the efficiency of dose planning and delivery.

Proton pencil beam scanning (PBS) treatment planning for head and neck (H&N) cancers is a time-consuming and experience-demanding task where a large number of planning objectives are involved. Deep reinforcement learning (DRL) has recently been introduced to the planning processes of intensity-modulated radiation therapy and brachytherapy for prostate, lung, and cervical cancers. However, existing approaches are built upon the Q-learning framework and weighted linear combinations of clinical metrics, suffering from poor scalability and flexibility and only capable of adjusting a limited number of planning objectives in discrete action spaces. We propose an automatic treatment planning model using the proximal policy optimization (PPO) algorithm and a dose distribution-based reward function for proton PBS treatment planning of H&N cancers. Specifically, a set of empirical rules is used to create auxiliary planning structures from target volumes and organs-at-risk (OARs), along with their associated planning objectives. These planning objectives are fed into an in-house optimization engine to generate the spot monitor unit (MU) values. A decision-making policy network trained using PPO is developed to iteratively adjust the involved planning objective parameters in a continuous action space and refine the PBS treatment plans using a novel dose distribution-based reward function. Proton H&N treatment plans generated by the model show improved OAR sparing with equal or superior target coverage when compared with human-generated plans. Moreover, additional experiments on liver cancer demonstrate that the proposed method can be successfully generalized to other treatment sites. To the best of our knowledge, this is the first DRL-based automatic treatment planning model capable of achieving human-level performance for H&N cancers.

PURPOSE To develop and evaluate a volumetric modulated arc therapy (VMAT) machine parameter optimization (MPO) approach based on deep-Q reinforcement learning (RL) capable of finding an optimal machine control policy using previous prostate cancer patient CT scans and contours, and applying the policy to new cases to rapidly produce deliverable VMAT plans in a simplified beam model. METHODS A convolutional deep-Q network was employed to control the dose rate and multi-leaf collimator of a C-arm linear accelerator model using the current dose distribution and machine parameter state as input. A Q-value was defined as the discounted cumulative cost based on dose objectives, and experience-replay RL was performed to determine a policy to minimize the Q-value. A two-dimensional network design was employed which optimized each opposing leaf pair independently while monitoring the corresponding dose plane blocked by those leaves. This RL approach was applied to CT and contours from 40 retrospective prostate cancer patients. The dataset was split into training (15 patients) and validation (5 patients) groups to optimize the network, and its performance was tested in an independent cohort of 20 patients by comparing RL-based dose distributions to conformal arcs and clinical intensity modulated radiotherapy (IMRT) delivering a prescription dose of 78 Gy in 40 fractions. RESULTS Mean±SD execution time of the RL VMAT optimization was 1.5±0.2 s per slice. In the test cohort, mean±SD (p-value) planning target volume (PTV), bladder, and rectum dose were 80.5±2.0 Gy (p <.001), 44.2±14.6 Gy (p <.001), and 43.7±11.1 Gy (p <.001) for RL VMAT compared to 81.6±1.1 Gy, 51.6±12.9 Gy, and 36.0±12.3 Gy for clinical IMRT. CONCLUSIONS RL was applied to VMAT MPO using clinical patient contours without independently optimized treatment plans for training and achieved comparable target and normal tissue dose to clinical plans despite the application of a relatively simple network design originally developed for video-game control. These results suggest that extending a RL approach to a full three-dimensional (3D) beam model could enable rapid artificial intelligence-based optimization of deliverable treatment plans, reducing the time required for radiotherapy planning without requiring previous plans for training.

Current beam orientation optimization algorithms for radiotherapy, such as column generation (CG), are typically heuristic or greedy in nature because of the size of the combinatorial problem, which leads to suboptimal solutions. We propose a reinforcement learning strategy using a Monte Carlo Tree Search (MCTS) that can find a better beam orientation set in less time than CG. We utilize a reinforcement learning structure involving a supervised learning network to guide the MCTS and to explore the decision space of beam orientation selection problems. We previously trained a deep neural network (DNN) that takes in the patient anatomy, organ weights, and current beams, then approximates beam fitness values to indicate the next best beam to add. Here, we use this DNN to probabilistically guide the traversal of the branches of the Monte Carlo decision tree to add a new beam to the plan. To assess the feasibility of the algorithm, we used a test set of 13 prostate cancer patients, distinct from the 57 patients originally used to train and validate the DNN, to solve five-beam plans. To show the strength of the guided MCTS (GTS) compared to other search methods, we also provided the performances of Guided Search, Uniform Tree Search and Random Search algorithms. On average, GTS outperformed all the other methods. It found a better solution than CG in 237 s on average, compared to 360 s for CG, and outperformed all other methods in finding a solution with a lower objective function value in less than 1000 s. Using our GTS method, we could maintain planning target volume (PTV) coverage within 1% error similar to CG, while reducing the organ-at-risk mean dose for body, rectum, left and right femoral heads; the mean dose to bladder was 1% higher with GTS than with CG.

Anatomical changes in head-and-neck cancer (HNC) patients during intensity-modulated proton therapy (IMPT) can shift the Bragg Peak of proton beams, risking tumor underdosing and organ-at-risk (OAR) overdosing. As a result, treatment replanning is often required to maintain clinically acceptable treatment quality. However, current manual replanning processes are often resource intensive and time consuming. In this work, we propose a patient-specific deep reinforcement learning (DRL) framework for automated IMPT replanning, with a reward-shaping mechanism based on a 150-point plan quality score designed to handle competing clinical objectives in radiotherapy planning. We formulate the planning process as a reinforcement learning (RL) problem where agents learn high-dimensional control policies to adjust plan optimization priorities to maximize plan quality. Unlike population-based approaches, our framework trains personalized agents for each patient using their planning Computed Tomography (CT) and augmented anatomies simulating anatomical changes (tumor progression and regression). This patient-specific approach leverages anatomical similarities along the treatment course, enabling effective plan adaptation. We implemented and compared two DRL algorithms, Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), using dose-volume histograms (DVHs) as state representations and a 22-dimensional action space of priority adjustments. Evaluation on eight HNC patients using actual replanning CT data showed that both DRL agents improved initial plan scores from 120.78 ± 17.18 to 139.59 ± 5.50 (DQN) and 141.50 ± 4.69 (PPO), surpassing the replans manually generated by a human planner (136.32±4.79). Further comparison of dosimetric endpoints confirms these improvements translate to better tumor coverage and OAR sparing across diverse anatomical changes. This work highlights the potential of DRL in addressing the geometric and dosimetric complexities of adaptive proton therapy, offering a promising solution for efficient offline adaptation and paving the way for online adaptive proton therapy.

Ultrasound (US) is a widely popular imaging technique for the diagnosis of tumors and associated soft tissue pathology. Traditionally, excised tumor masses are manually sliced for microscopic examination, which is a resource-intensive, time-consuming process, and prone to human error. The proposed work addresses these challenges by developing a cost-effective US gantry system integrated with a deep learning algorithm to automate the tissue slice selection process. This system scans the entire tumor and by integrating a deep learning algorithm predicts the optimal slice to assist its preparation for microscopic analysis. Automating this process reduces the time and resources required while minimizing the risk of human error. Optimal tissue slice reduces sampling associated uncertainty in diagnosis and treatment planning. Thereby determining tumor grade and type, and helping to reduce the treatment risks. The initial development focused on a linear US gantry that moves in one direction to acquire B-mode images. However, this design is limited, as it cannot fully capture the tumor’s structural complexity. In order to overcome this, we developed an angular US gantry that can maneuver along multiple angles, acquiring a broader range of images for comprehensive geometric analysis. The angular gantry demonstrated significant improvement, achieving 98% accuracy in selecting the optimal tissue slice.

BACKGROUND Dose deposition characteristics of proton radiation can be advantageous over photons. Proton treatment planning however poses additional challenges for the planners. Proton therapy is usually delivered with only a small number of beam angles, and the quality of a proton treatment plan is largely determined by the beam angles employed. Finding the optimal beam angles for a proton treatment plan requires time and experience, motivating the investigation of automatic beam angle selection methods. PURPOSE A deep learning-based approach to automatic beam angle selection is proposed for proton pencil-beam scanning treatment planning of liver lesions. METHODS We cast beam-angle selection as a multi-label classification problem. To account for angular boundary discontinuity, the underlying convolution neural network is trained with the proposed Circular Earth Mover's Distance based regularization and multi-label circular-smooth label technique. Furthermore, an analytical algorithm emulating proton treatment planners' clinical practice is employed in post-processing to improve the output of the model. Forty-nine patients that received proton liver treatments between 2017 and 2020 were randomly divided into training (n = 31), validation (n = 7), and test sets (n = 11). AI-selected beam angles were compared with those angles selected by human planners, and the dosimetric outcome was investigated by creating plans using knowledge-based treatment planning. RESULTS For 7 of the 11 cases in the test set, AI-selected beam angles agreed with those chosen by human planners to within 20 degrees (median angle difference = 10°; mean = 18.6°). Moreover, out of the total 22 beam angles predicted by the model, 15 (68%) were within 10 degrees of the human-selected angles. The high correlation in beam angles resulted in comparable dosimetric statistics between proton treatment plans generated using AI- and human-selected angles. For the cases with beam angle differences exceeding 20°, the dosimetric analysis showed similar plan quality although with different emphases on organ-at-risk sparing. CONCLUSIONS This pilot study demonstrated the feasibility of a novel deep learning-based beam angle selection technique. Testing on liver cancer patients showed that the resulting plans were clinically viable with comparable dosimetric quality to those using human-selected beam angles. In tandem with auto-contouring and knowledge-based treatment planning tools, the proposed model could represent a pathway for nearly fully automated treatment planning in proton therapy. This article is protected by copyright. All rights reserved.

Proton arc therapy (PAT) is an emerging and promising modality in radiotherapy, offering improved dose distribution and treatment robustness over intensity-modulated proton therapy. Yet, identifying the optimal energy layer (EL) sequence remains challenging due to the intensive computational demand and prolonged treatment delivery time. This study proposes an unsupervised deep learning model for fast EL pre-selection that minimizes EL switch (ELS) time while maintaining high plan quality. We introduce a novel data representation method, spot-count representation, which encodes the number of proton spots intersecting the target and organs at risk (OAR) in a matrix structured by sorted gantry angles and energy layers. This representation serves as the input of an U-Net style architecture, SPArc_dl, which is trained using a tri-objective function: maximizing spot-counts on target, minimizing spot-counts on OAR, and reducing ELS time. The model is evaluated on 35 nasopharyngeal cancer cases, and its performance is compared to SPArc_particle_swarm (SPArc_ps). SPArc_dl produces EL pre-selection that significantly improves both plan quality and delivery efficiency. Compared to SPArc_ps, it enhances the conformity index by 0.1 (p<0.01), reduces the homogeneity index by 0.71 (p<0.01), lowers the brainstem mean dose by 0.25 (p<0.01), and shortens the ELS time by 37.2% (p<0.01). The results unintentionally reveal employing unchanged ELS is more time-wise efficient than descended ELS. SPArc_dl's inference time is within 1 second. However, SPArc_dl plan demonstrates limitation in robustness. The proposed spot-count representation lays a foundation for incorporating unsupervised deep learning approaches into EL pre-selection task. SPArc_dl is a fast tool for generating high-quality PAT plans by strategically pre-selecting EL to reduce delivery time while maintaining excellent dosimetric performance.

BACKGROUND Real-time liver imaging is challenged by the short imaging time (within hundreds of milliseconds) to meet the temporal constraint posted by rapid patient breathing, resulting in extreme under-sampling for desired 3D imaging. Deep learning (DL)-based real-time imaging/motion estimation techniques are emerging as promising solutions, which can use a single X-ray projection to estimate 3D moving liver volumes by solved deformable motion. However, such techniques were mostly developed for a specific, fixed X-ray projection angle, thereby impractical to verify and guide arc-based radiotherapy with continuous gantry rotation. PURPOSE To enable deformable motion estimation and 3D liver imaging from individual X-ray projections acquired at arbitrary X-ray scan angles, and to further improve the accuracy of single X-ray-driven motion estimation. METHODS We developed a DL-based method, X360, to estimate the deformable motion of the liver boundary using an X-ray projection acquired at an arbitrary gantry angle (angle-agnostic). X360 incorporated patient-specific prior information from planning 4D-CTs to address the under-sampling issue, and adopted a deformation-driven approach to deform a prior liver surface mesh to new meshes that reflect real-time motion. The liver mesh motion is solved via motion-related image features encoded in the arbitrary-angle X-ray projection, and through a sequential combination of rigid and deformable registration modules. To achieve the angle agnosticism, a geometry-informed X-ray feature pooling layer was developed to allow X360 to extract angle-dependent image features for motion estimation. As a liver boundary motion solver, X360 was also combined with priorly-developed, DL-based optical surface imaging and biomechanical modeling techniques for intra-liver motion estimation and tumor localization. RESULTS With geometry-aware feature pooling, X360 can solve the liver boundary motion from an arbitrary-angle X-ray projection. Evaluated on a set of 10 liver patient cases, the mean (± s.d.) 95-percentile Hausdorff distance between the solved liver boundary and the "ground-truth" decreased from 10.9 (±4.5) mm (before motion estimation) to 5.5 (±1.9) mm (X360). When X360 was further integrated with surface imaging and biomechanical modeling for liver tumor localization, the mean (± s.d.) center-of-mass localization error of the liver tumors decreased from 9.4 (± 5.1) mm to 2.2 (± 1.7) mm. CONCLUSION X360 can achieve fast and robust liver boundary motion estimation from arbitrary-angle X-ray projections for real-time imaging guidance. Serving as a surface motion solver, X360 can be integrated into a combined framework to achieve accurate, real-time, and marker-less liver tumor localization.

Purpose. Gamma analysis serves as a critical safety assurance tool in radiotherapy, yet its broader clinical implementation remains constrained by insufficient error cause determination. To address this limitation, this study proposes a gamma passing rate (GPR) prediction method with error classification capabilities by integrating dosimetric feature engineering with a dose prediction model.Method. The study cohort comprised 26 clinical cases, with 6 cases (1,515 static segments generated from volumetric modulated arc therapy (VMAT) plans) allocated for model training and 20 cases (10 step-and-shoot plans with 415 segments and 10 VMAT plans) allocated for testing. Measurements were performed using a MatriXX chamber array at a gantry angle of 0 degrees. Data was acquired segment-by-segment for step-and-shoot plans and via integration for VMAT plans, respectively. A dosimetric feature engineering protocol was used to partition each static segment into five distinct regions (Region 1–5) on the basis of physical characteristics and error susceptibility patterns. These regional dose distributions served as both model inputs and independent variables for error analysis. A GAN-based model was trained to predict segment doses, which were subsequently aggregated for plan-level GPR calculations. Model accuracy was first validated by statistically analyzing GPR differences between measurements and predictions across various plan types in the test set, followed by assessing dose discrepancies at failure and passing points for both measured and predicted values. Results. The predicted GPR was 70.26% ± 13.07% for segments, 93.53% ± 2.06% for step-and-shoot plans, and 92.61% ± 4.67% for VMAT plans, with corresponding measured values of 74.47% ± 10.06%, 96.35% ± 1.82%, and 91.60% ± 4.05%, respectively. Regional dose analysis revealed statistically significant differences (p < 0.05) in the measured values for Regions 2–5, with classification AUC values of 0.69, 0.64, 0.65, and 0.63, respectively. The predicted values showed comparable performance for Region 2 with AUC of 0.66, whereas Regions 3–5 demonstrated AUCs of 0.50, 0.59, and 0.57 respectively. Conclusions. The integrated approach enables accurate GPR prediction while providing actionable error localization at the control point level. The quantitative error source analysis offers valuable guidance for modifying high-risk treatment plans and demonstrates significant potential for enhancing clinical radiotherapy quality assurance protocols.

Sparse-angle X-ray Computed Tomography (CT) plays a vital role in industrial quality control but leads to an inherent trade-off between scan time and reconstruction quality. Adaptive angle selection strategies try to improve upon this based on the idea that the geometry of the object under investigation leads to an uneven distribution of the information content over the projection angles. Deep Reinforcement Learning (DRL) has emerged as an effective approach for adaptive angle selection in X-ray CT. While previous studies focused on optimizing generic image quality measures using a fixed number of angles, our work extends them by considering a specific downstream task, namely image-based defect detection, and introducing flexibility in the number of angles used. By leveraging prior knowledge about typical defect characteristics, our task-adaptive angle selection method, adaptable in terms of angle count, enables easy detection of defects in the reconstructed images.

BACKGROUND Recently, patient rotating devices for gantry-free radiotherapy, a new approach to implement external beam radiotherapy, have been introduced. When a patient is rotated in the horizontal position, gravity causes anatomic deformation. For treatment planning, one feasible method is to acquire simulation images at different horizontal rotation angles. PURPOSE This study aimed to investigate the feasibility of synthesizing magnetic resonance (MR) images at patient rotation angles of 180° (prone position) and 90° (lateral position) from those at a rotation angle of 0° (supine position) using deep learning. METHODS This study included 23 healthy male volunteers. They underwent MR imaging in the supine position and then in the prone (23 volunteers) and lateral (16 volunteers) positions. T1-weighted fast spin echo was performed for all positions with the same parameters. Two two-dimensional deep learning networks, pix2pix Generative Adversarial Network (pix2pix GAN) and CycleGAN, were developed for synthesizing MR images in the prone and lateral positions from those in the supine position, respectively. For the evaluation of the models, leave-one-out cross-validation was performed. The mean absolute error (MAE), Dice similarity coefficient (DSC), and Hausdorff distance (HD) were used to determine the agreement between the prediction and ground truth for the entire body and four specific organs. RESULTS For pix2pix GAN, the synthesized images were visually bad, and no quantitative evaluation was performed. The quantitative evaluation metrics of the body outlines calculated for the synthesized prone and lateral images using CycleGAN were as follows: MAE, 35.63 ± 3.98 and 40.45 ± 5.83, respectively; DSC, 0.97 ± 0.01 and 0.94 ± 0.01, respectively; and HD (in pixels), 16.74 ± 3.55 and 31.69 ± 12.03, respectively. The quantitative metrics of the bladder and prostate performed were also promising for both the prone and lateral images, with mean values > 0.90 in DSC (P>0.05). The mean DSC and HD values of the bilateral femur for the prone images were 0.96 and 3.63 (in pixels), respectively, and 0.78 and 12.65 (in pixels) for the lateral images, respectively (P<0.05). CONCLUSIONS The CycleGAN could synthesize the MRI at lateral and prone positions using images at supine position, and it could benefit gantry-free radiation therapy. This article is protected by copyright. All rights reserved.

PURPOSE C-arm cone-beam computed tomography (CBCT) in-suite imaging is often used in a brachytherapy suite. However, due to the limited rotation angle of the C-arm gantry and the dimension of the flat panel imager (FPI), CBCT images are often truncated and not suitable for treatment planning. In this simulation study, we present the design of a novel ultra-compact mobile dual-source CBCT (dCBCT) that can scan large field of view with half system rotation. Enabled by deep learning image reconstruction, it can perform ultra-short scans and stereoscopic imaging before and during high dose rate (HDR) treatments. MATERIAL AND METHODS The dCBCT comprises two x-ray sources and a flat panel imager mounted on a C-arm gantry. The dual-sources configuration enables real-time stereoscopic imaging, also avoids data truncation problem of conventional C-arm CBCT. Simulation studies were performed to prove the concept of ultra-short scan of dCBCT. Deep Image Prior (DIP) image reconstruction without and with a Prior was also developed to reduce the scan angle. RESULTS The simulation studies of dCBCT show that it can achieve a sufficient reconstruction field of view with 180° rotation. DIP reconstruction reduces scanning angle to 135° without sacrificing image quality. With body profile as constraint, ultra-short scan with merely 90° system rotation can be achieved. CONCLUSIONS Powered by deep-learning based limited-angle image reconstruction, dCBCT can scan full body with a short scan, allowing rapid 3D and real-time planar imaging in brachytherapy suite.

BACKGROUND This study presents a physics-informed, feature-engineered machine learning (ML) framework to predict multileaf collimator (MLC) and gantry positional errors in volumetric modulated arc therapy (VMAT) METHODS: Data from 32 VMAT trajectory logs (TrueBeam linac, HD120 MLC) were synchronized with DICOMRT plans to extract delivery dynamics. Novel physics-based parameters were introduced: a friction factor, an enhanced gravity vector, and MLC speed-normalized features. Three ML models XGBoost, LightGBM, and deep neural networks (DNNs) were optimized using Optuna and trained on trajectory log and DICOM-RT-derived datasets. Feature importance was evaluated via Spearman correlation, mutual information, and SHapley Additive Explanations (SHAP). RESULTS Systematic discrepancies between DICOM-RT and trajectory log data were identified, with mean absolute deviations of 7.0 % (MLC speed), 8.0 % (gantry speed), and 8.5 % (dose rate). MLC speed emerged as the dominant predictor (Spearman: rs = 0.891), while friction and gravity features exhibited significant correlations (rs = 0.46 and 0.33, respectively). Mutual information revealed non-monotonic dependencies between gantry error and gantry angle (score: 0.34). LightGBM and XGBoost achieved superior MLC error prediction (MAE: 0.0019 mm, RMSE: 0.0027 mm), capturing > 90 % of observed errors, while DNNs lagged by 30 %. Engineered features reduced residual errors by 30 %. Gantry error predictions showed lower accuracy (MAE: 0.012°-0.015°). SHAP analysis highlighted physics-driven features as top contributors. CONCLUSION This work underscores the critical role of domain knowledge in ML for radiotherapy, achieving a 30% error reduction through physics-based feature engineering. The findings advocate for prioritizing feature space exploration alongside model optimization to enhance VMAT quality assurance.

INTRODUCTION New radiotherapy machines such as Halcyon are capable of delivering dose-rate of 600 monitor-units per minute, allowing large numbers of patients treated per day. However, patient-specific quality assurance (QA) is still required, which dramatically decrease machine availability. Innovative artificial intelligence (AI) algorithms could predict QA result based on complexity metrics. However, no AI solution exists for Halcyon machines and the complexity metrics to be used have not been definitively determined. The aim of this study was to develop an AI solution capable of firstly determining the complexity indices to be obtained and secondly predicting patient-specific QA in a routine clinical setting. METHODS Three hundred and eighteen beams from 56 patients with breast cancer were used. The seven complexity indices named Modulation-Complexity-Score (MCS), Small-Aperture-Score (SAS10), Beam-Area (BA), Beam-Irregularity (BI), Beam-Modulation (BM), Gantry and Collimator angles were used as input to the AI model. Machine learning (ML) and deep learning (DL) models using tensorflow were set up to predict DreamDose QA conformance. RESULTS MCS, BI, gantry and collimator angle are not correlated with QA compliance. Therefore, ML and DL models were trained using SAS10, BA and BM complexity indices. ROC analyses enabled to find best predicted probability threshold to increase specificity and sensitivity. ML models did not show satisfactory performance with an area under-the-curve (AUC) of 0.75 and specificity and sensitivity of 0.88 and 0.86. However, optimised DL model showed better performance with an AUC of 0.95 and specificity and sensitivity of 0.98 and 0.97. CONCLUSION The DL model demonstrated a high degree of accuracy in its predictions of the quality assurance (QA) results. Our online predictive QA-platform offers significant time savings in terms of accelerator occupancy and working time.

PURPOSE The feasibility of a deep learning-based markerless real-time tumor tracking (RTTT) method was retrospectively studied with orthogonal kV X-ray images and clinical tracking records acquired during lung cancer treatment. METHODS Ten patients with lung cancer treated with marker-implanted RTTT were included. The prescription dose was 50 Gy in four fractions, using seven- to nine-port non-coplanar static beams. This corresponds to 14-18 X-ray tube angles for an orthogonal X-ray imaging system rotating with the gantry. All patients underwent 10 respiratory phases four-dimensional computed tomography. After a data augmentation approach, for each X-ray tube angle of a patient, 2250 digitally reconstructed radiograph (DRR) images with gross tumor volume (GTV) contour labeled were obtained. These images were adopted to train the patient and X-ray tube angle-specific GTV contour prediction model. During the testing, the model trained with DRR images predicted GTV contour on X-ray projection images acquired during treatment. The predicted three-dimensional (3D) positions of the GTV were calculated based on the centroids of the contours in the orthogonal images. The 3D positions of GTV determined by the marker-implanted RTTT during the treatment were considered as the ground truth. The 3D deviations between the prediction and the ground truth were calculated to evaluate the performance of the model. RESULTS The median GTV volume and motion range were 7.42 (range, 1.18-25.74) cm3 and 22 (range, 11-28) mm, respectively. In total, 8993 3D position comparisons were included. The mean calculation time was 85 ms per image. The overall median value of the 3D deviation was 2.27 (interquartile range: 1.66-2.95) mm. The probability of the 3D deviation smaller than 5 mm was 93.6%. CONCLUSIONS The evaluation results and calculation efficiency show the proposed deep learning-based markerless RTTT method may be feasible for patients with lung cancer.

No abstract available

Objective. Beam orientation optimization (BOO) in intensity-modulated radiation therapy (IMRT) is a complex, non-convex problem traditionally addressed with heuristic methods. Approach. This work demonstrates the potential improvement of the proposed BOO, providing a mathematically grounded benchmark that can guide and validate heuristic BOO methods, while also offering a computationally efficient workflow suitable for clinical application. A novel framework integrating second-order cone programming (SOCP) relaxation, sparse mixed integer programming (SMIP), and deep inverse optimization is proposed. Nonconvex dose-volume constraints were managed via SOCP relaxation, ensuring convexity while maintaining sparsity. BOO was formulated as an SMIP problem with binary beam selection, solved using an augmented Lagrange method. To accelerate optimization, a neural network approximated optimal solution, improving computational efficiency eightfold. A retrospective analysis of 12 locally advanced non-small cell lung cancer (NSCLC) patients (60 Gy prescription) compared automated BOO-selected beam angles with expert selections, evaluating dosimetric metrics such as planning target volume (PTV) maximum dose, D98%, lung V20, and mean heart and esophagus dose. Main results. In 12 retrospective study, the automated BOO demonstrated superior dose conformity and sparing of critical structures. Specifically, the BOO plans achieved comparable PTV coverage (maximum: 61.7 ± 1.4 Gy vs. 62.1 ± 1.5 Gy, D98%: 59.5 ± 0.7 Gy vs. 59.5 ± 0.6 Gy, D2%: 61.2 ± 1.3 Gy vs. 61.4 ± 1.4 Gy with p-values >0.5) but demonstrated improved sparing for lungs (V20: 9.8 ± 2.2% vs. 11.5 ± 2.3%, p-value: 0.01), heart (mean: 3.3 ± 0.6 Gy vs. 4.3 ± 0.5 Gy, p-value: 0.04), esophagus (mean: 0.5 ± 1.3 Gy vs. 1.8 ± 1.5 Gy, p-value: 0.02), and spinal cord (max: 7.2 ± 3.4 Gy vs. 9.0 ± 3.2 Gy, p-value < 0.01) compared to human-selected plans. Significance. This approach highlighted the potential of BOO to enhance treatment efficacy by optimizing beam angles more effectively than manual selection. This framework establishes a benchmark for BOO in IMRT, enhancing heuristic methods through a hybrid framework that combines mathematical optimization with targeted heuristics to improve solution quality and computational efficiency. The integration of SMIP and deep inverse optimization significantly improves computational efficiency and plan quality.

PURPOSE Beam orientation selection, whether manual or protocol-based, is the current clinical standard in radiation therapy treatment planning, but it is tedious and can yield suboptimal results. Many algorithms have been designed to optimize beam orientation selection because of its impact on treatment plan quality, but these algorithms suffer from slow calculation of the dose influence matrices of all candidate beams. We propose a fast beam orientation selection method, based on deep learning neural networks (DNN), capable of developing a plan comparable to those developed by the state-of-the-art column generation method. Our model's novelty lies in its supervised learning structure (using column generation to teach the network), DNN architecture, and ability to learn from anatomical features to predict dosimetrically suitable beam orientations without using dosimetric information from the candidate beams. This may save hours of computation. METHODS A supervised DNN is trained to mimic the column generation algorithm, which iteratively chooses beam orientations one-by-one by calculating beam fitness values based on Karush-Kush-Tucker optimality conditions at each iteration. The DNN learns to predict these values. The dataset contains 70 prostate cancer patients-50 training, 7 validation, and 13 test patients-to develop and test the model. Each patient's data contains 6 contours: PTV, body, bladder, rectum, and left and right femoral heads. Column generation was implemented with a GPU-based Chambolle-Pock algorithm, a first-order primal-dual proximal-class algorithm, to create 6270 plans. The DNN trained over 400 epochs, each with 2500 steps and a batch size of 1, using the Adam optimizer at a learning rate of 1×10-5 and a 6-fold cross-validation technique. RESULTS The average and standard deviation of training, validation, and testing loss functions among the 6-folds were 0.62±0.09%, 1.04±0.06%, and 1.44±0.11%, respectively. Using column generation and supervised DNN, we generated two sets of plans for each scenario in the test set. The proposed method took at most 1.5 seconds to select a set of five beam orientations and 300 second to calculate the dose influence matrices for 5 beams and finally 20 seconds to solve the fluence map optimization. However, column generation needed around 15 hours to calculate the dose influence matrices of all beams and at least 400 seconds to solve both the beam orientation selection and fluence map optimization problems. The differences in the dose coverage of PTV between plans generated by column generation and by DNN were 0.2%. The average dose differences received by organs at risk were between 1 and 6 percent: Bladder had the smallest average difference in dose received (0.956±1.184%), then Rectum (2.44±2.11%), Left Femoral Head (6.03±5.86%), and Right Femoral Head (5.885±5.515%). The dose received by Body had an average difference of 0.10± 0.1% between the generated treatment plans. CONCLUSIONS We developed a fast beam orientation selection method based on a DNN that selects beam orientations in seconds and is therefore suitable for clinical routines. In the training phase of the proposed method, the model learns the suitable beam orientations based on patients' anatomical features and omits time intensive calculations of dose influence matrices for all possible candidate beams. Solving the fluence map optimization to get the final treatment plan requires calculating dose influence matrices only for the selected beams.

No abstract available

No abstract available

Proton beam radiotherapy is an advanced cancer treatment utilizing high-energy protons to destroy tumor matter. This treatment requires precise Bragg-peak localization, but Compton-camera image reconstructions are often unusable due to mischaracterized scattering sequences and excessive image noise. We present machine learning models to classify the scattering events. Multiple novel robust-volume datasets simulating particle interactions with human tissue were generated using Duke University CT scans and Geant4 and Monte-Carlo Detector Effects (MCDE) software. Novel implementations of an Event Classifier Transformer and a 1D Convolutional Neural Network (CNN) were developed, while prior models, such as Fully-Connected Neural Networks (FCN) and Long Short-Term Memory Neural Network (LSTM), were optimized through large-scale hyperparameter studies using a novel automated tuning framework built into the Big-Data REU Integrated Development and Experimentation (BRIDE) machine learning pipeline. Fully-connected neural networks and convolutional neural networks show significant improvements in model accuracy over prior work on simulated patient data and demonstrate that relatively shallow, regularized models generalize best.

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

No abstract available

Background Stereotactic body radiation therapy (SBRT) for liver cancer has shown promising therapeutic effects. Effective treatment relies not only on the precise delivery provided by image-guided radiation therapy (IGRT) but also high dose gradient formed around the treatment volume to spare functional liver tissue, which is highly dependent on the beam/arc angle selection. In this study, we aim to develop a decision support model to learn human planner's beam navigation approach for beam angle/arc angle selection for liver SBRT. Methods A total of 27 liver SBRT/HIGRT patients (10 IMRT, 17 VMAT/DCA) were included in this study. A dosimetric budget index was defined for each beam angle/control point considering dose penetration through the patient body and liver tissue. Optimal beam angle setting (beam angles for IMRT and start/terminal angle for VMAT/DCA) was determined by minimizing the loss function defined as the sum of total dosimetric budget index and beam span penalty function. Leave-one-out validation was exercised on all 27 cases while weighting coefficients in the loss function was tuned in nested cross validation. To compare the efficacy of the model, a model plan was generated using automatically generated beam setting while retaining the original optimization constraints in the clinical plan. Model plan was normalized to the same planning target volume (PTV) V100% as the clinical plans. Dosimetric endpoints including PTV D98%, D2%, liver V20Gy and total MU were compared between two plan groups. Wilcoxon Signed-Rank test was performed with the null hypothesis being that no difference exists between two plan groups. Results Beam setting prediction was instantaneous. Mean PTV D98% was 91.3% and 91.3% (P=0.566), while mean PTV D2% was 107.9% and 108.1% (P=0.164) for clinical plan and model plan respectively. Liver V20Gy showed no significant difference (P=0.590) with 23.3% for clinical plan and 23.4% for the model plan. Total MU is comparable (P=0.256) between the clinical plan (avg. 2,389.6) and model plan (avg. 2,319.6). Conclusions The evidence driven beam setting model yielded similar plan quality as hand-crafted clinical plan. It is capable of capturing human's knowledge in beam selection decision making. This model could facilitate decision making for beam angle selection while eliminating lengthy trial-and-error process of adjusting beam setting during liver SBRT treatment planning.

PURPOSE Cone beam computed tomography (CBCT)-based online adaptive radiotherapy (ART) is especially beneficial for patients with large inter-fractional anatomical changes. However, treatment planning and review decisions need to be made at the treatment console in real-time and may be delegated to clinical staff whose conventional scope of practice does not include making such decisions. Therefore, implementation can create new safety risks and inefficiencies. The objective of this work is to systematically analyze the safety and efficiency implications of human decision-making during the treatment session for CBCT-based online ART. METHODS AND MATERIALS The analysis was performed by applying the Systems-Theoretic Process Analysis technique and its extension for human decision-making. Four centers of different CBCT-based online ART practice models comprised the analysis team. RESULTS The general radiotherapy control structure was refined to model the interactions between routine treatment delivery staff and in-person or remote support staff. The treatment delivery staff perform six key control actions. Eighteen undesirable states of those control actions were identified as impacting safety and/or efficiency. In turn, 97 hazardous clinical scenarios were identified, with the control action prepare and position patient having the least number of scenarios and delineate/edit influencer and target structures having the most. Five of these are specific to either in-person or remote support during the treatment session, and 12 arise from staff support in general. CONCLUSIONS An optimally safe and efficient online ART program should require little to no support staff at the treatment console and thus reducing staff coordination. Uptraining of the staff already at the treatment console is needed to achieve this goal. Beyond the essential knowledge and skills such as contour editing and the selection of an optimal plan, uptraining should also target the specific cognitive biases identified in this work and the cognitive strategies to overcome these biases. Additionally, technological and organizational changes are necessary.

Purpose/Objective(s): The additional personnel and imaging procedures required for Adaptive Radiation Therapy (ART) pose a challenge for a broad implementation. We hypothesize that a change in transit fluence during the treatment course is correlated with the change of quality of life and thus can be used as a replanning trigger. Materials/Methods: Twenty-one head and neck cancer (HNC) patients filled out an MD Anderson Dysphagia Inventory (MDADI) questionnaire, before-and-after the radiotherapy treatment course. The transit fluence was measured by the Watchdog (WD) in-vivo portal dosimetry system. The patients were monitored with daily WD and weekly CBCTs. The region of interest (ROI) of each patient was defined as the outer contour of the patient between approximate spine levels C1 to C4, essentially the neck and mandible inside the beam’s eye view. The nth day integrated transit fluence change, Δϕn, and the volume change, ΔVROI, of the ROI of each patient was calculated from the corresponding WD and CBCT measurements. The correlation between MDADI scores and age, gender, planning mean dose to salivary glands , weight change ΔW, ΔVROI, and Δϕn, were analyzed using the ranked-Pearson correlation. Results: No statistically significant correlation was found for age, gender and ΔW. was found to have clinically important correlation with functional MDADI (ρ = −0.39, P = 0.081). ΔVROI was found to have statistically significant correlation of 0.44, 0.47 and 0.44 with global, physical and functional MDADI (P-value < 0.05). Δϕn was found to have statistically significant ranked-correlation (−0.46, −0.46 and −0.45) with physical, functional and total MDADI (P-value < 0.05). Conclusion: A transit fluence based decision support metric (DSM) is statistically correlated with the dysphagia risk. It can not only be used as an early signal in assisting clinicians in the ART patient selection for replanning, but also lowers the resource barrier of ART implementation.

No abstract available

PURPOSE/OBJECTIVE(S) Spot-scanning proton arc (SPArc) is a novel technique that employs a planning optimization algorithm to select the energies and positions of spots along a dynamic rotational arc trajectory. The SPArc technique has the potential to achieve superior dose conformality and treatment delivery efficiency over intensity-modulated proton therapy. However, creating such a SPArc plan using existing approaches is time-consuming and computationally extensively. This study investigated the feasibility of using the deep learning (DL) technique to predict the 3D dose distribution of the SPArc treatment plan, leveraging the prior knowledge acquired from conventional intensity-modulated radiation therapy (IMRT) plans. MATERIALS/METHODS A DL model, 3D-Unet with residual connections and attention gates, was trained using an open-source database of CT images, critical structures, and IMRT plans from 340 head and neck cancer patients (HNC) as the base model. Transfer learning technique was applied to fine-tune the model parameters using the SPArc treatment plans created on the HNC patients from an in-house dataset, where the SPArc treatment plans (including control point sampling, energy layer distribution, arc trajectory, etc.,) were optimized using a previously developed iterative approach. The performance of the DL model was evaluated by comparing predicted and planned doses over 17 SPArc treatment plans by using 4-fold cross-validation. RESULTS The SPArc planning time per patient was 8∼12 hours, while the dose prediction time was reduced to 2∼3 minutes using the proposed DL model. The deviation of D95 in the target was (-1.8±1.6) %. The deviation of the mean dose in the parotids, cord, mandible, and brainstem were (2.5±6.5) %, (-0.5±4.3) %, (1.4±3.9) %, and (3.4±8) % of the prescription, respectively. The dice similarity coefficients of the 80%, 70%, and 60% isodose lines were (0.9±0.09), (0.93±0.01), and (0.94±0.01), respectively. CONCLUSION Our results demonstrate that a DL-based dose prediction model can be created with a limited number of SPArc treatment plans through transfer learning. The DL model can directly predict the 3D dose distribution in minutes for automated planning. This study paves the roadmap to develop a quick clinical decision platform for the optimal selection among the multi-treatment modalities.

Simple Summary Pancreatic cancer is a devastating disease with more than 60,000 new cases each year and a less than 10 percent 3-year overall survival rate. Radiation therapy is an effective treatment for locally advanced pancreatic cancer. The current clinical RT workflow, however, is lengthy and involves separate image acquisition for diagnostic CT and planning CT, which imposes a huge burden on patients and their caretakers. Moreover, studies have shown a reduction in mortality rate from expeditious radiotherapy treatment courses. Here, we proposed an innovative deep learning solution to adapt the shape of a patient’s body in diagnostic CT to the treatment delivery setup, and consequently, reduce the time to treatment initiation by half. As a result, our method also reduces the time to surgery and greatly decreases the risk of progression for pancreatic cancer. Abstract Major sources of delay in the standard of care RT workflow are the need for multiple appointments and separate image acquisition. In this work, we addressed the question of how we can expedite the workflow by synthesizing planning CT from diagnostic CT. This idea is based on the theory that diagnostic CT can be used for RT planning, but in practice, due to the differences in patient setup and acquisition techniques, separate planning CT is required. We developed a generative deep learning model, deepPERFECT, that is trained to capture these differences and generate deformation vector fields to transform diagnostic CT into preliminary planning CT. We performed detailed analysis both from an image quality and a dosimetric point of view, and showed that deepPERFECT enabled the preliminary RT planning to be used for preliminary and early plan dosimetric assessment and evaluation.

No abstract available

BACKGROUND AND PURPOSE Accurate delineation of the clinical target volume (CTV) and planning target volume (PTV) is essential for effective radiotherapy in uterine malignancies. Manual contouring is laborious, time-consuming, and subjective, and current automatic methods often focus on a single cancer type with limited external validation. To address this, we developed a deep-learning model capable of accurately delineating both CTV and PTV across multiple uterine malignancies using CT imaging. MATERIALS AND METHODS We retrospectively collected 602 contrast-enhanced CT scans, comprising 302 cases (cervical and endometrial cancers) from our institution and an additional 300 cervical cancer scans from external centers. Expert radiation oncologists manually delineated the CTV and PTV on each image. Among the 302 internal cancer cases, 177 cervical cancer cases were used for model training with five-fold cross-validation. Additionally, 41 cervical cancer cases were reserved as an internal testing cohort, while 84 endometrial cancer cases constituted the first external testing cohort to assess the model's generalizability across cancer types. The remaining 300 cervical cancer scans from external centers formed a second external testing cohort to assess model robustness across institutions. We evaluated three segmentation architectures-2D, full-resolution 3D, and cascaded 3D networks-and measured their performance using three standard metrics: Dice Similarity Coefficient (DSC), 95 % Hausdorff Distance (HD95), and Average Surface Distance (ASD). RESULTS The model-generated segmentations demonstrated strong concordance with the expert contours. In the internal testing cohort with the same cancer type, performance metrics (DSC, HD95, ASD) were consistently high. Similarly, the external testing cohort with different cancer types showed robust performance, indicating effective generalizability. On the internal testing cohort, the model demonstrated strong performance, achieving mean DSCs of 83.42 % for PTV and 81.23 % for CTV, with low spatial errors (PTV: ASD 2.01 mm, HD95 5.71 mm; CTV: ASD 1.35 mm, HD95 4.75 mm). In the endometrial cancer cohort, PTV segmentation achieved a DSC of 82.88 %, while CTV segmentation yielded an HD95 of 5.85  mm and an ASD of 1.34  mm. Additionally, clinical evaluation revealed that approximately 90 % of the model-generated contours required no or only minor revision. CONCLUSIONS We present a multicenter-validated deep-learning based framework for automatic CTV and PTV delineation across diverse uterine malignancies on CT. Our model offers a scalable, generalized solution with the potential to reduce the workload in radiation oncology, improve consistency, and streamline clinical workflows.

To evaluate the precision of automated segmentation facilitated by deep learning (DL) and dose calculation in adaptive radiotherapy (ART) for nasopharyngeal cancer (NPC), leveraging synthetic CT (sCT) images derived from cone-beam CT (CBCT) scans on a conventional C-arm linac. Sixteen NPC patients undergoing a two-phase offline ART were analyzed retrospectively. The initial (pCT1) and adaptive (pCT2) CT scans served as gold standard alongside weekly acquired CBCT scans. Patient data, including manually delineated contours and dose information, were imported into ArcherQA. Using a cycle-consistent generative adversarial network (cycle-GAN) trained on an independent dataset, sCT images (sCT1, sCT4, sCT4*) were generated from weekly CBCT scans (CBCT1, CBCT4, CBCT4) paired with corresponding planning CTs (pCT1, pCT1, pCT2). Auto-segmentation was performed on sCTs, followed by GPU-accelerated Monte Carlo dose recalculation. Auto-segmentation accuracy was assessed via Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95). Dose calculation fidelity on sCTs was evaluated using dose-volume parameters. Dosimetric consistency between recalculated sCT and pCT plans was analyzed via Spearman’s correlation, while volumetric changes were concurrently evaluated to quantify anatomical variations. Most anatomical structures demonstrated high pCT-sCT agreement, with mean values of DSC > 0.85 and HD95 < 5.10 mm. Notable exceptions included the primary Gross Tumor Volume (GTVp) in the pCT2-sCT4 comparison (DSC: 0.75, HD95: 6.03 mm), involved lymph node (GTVn) showing lower agreement (DSC: 0.43, HD95: 16.42 mm), and submandibular glands with moderate agreement (DSC: 0.64–0.73, HD95: 4.45–5.66 mm). Dosimetric analysis revealed the largest mean differences in GTVn D99: -1.44 Gy (95% CI: [-3.01, 0.13] Gy) and right parotid mean dose: -1.94 Gy (95% CI: [-3.33, -0.55] Gy, p < 0.05). Anatomical variations, quantified via sCTs measurements, correlated significantly with offline adaptive plan adjustments in ART. This correlation was strong for parotid glands (ρ > 0.72, p < 0.001), a result that aligned with sCT-derived dose discrepancy analysis (ρ > 0.57, p < 0.05). The proposed method exhibited minor variations in volumetric and dosimetric parameters compared to prior treatment data, suggesting potential efficiency improvements for ART in NPC through reduced human dependency.

BACKGROUND AND PURPOSE To investigate the performance of head-and-neck (HN) organs-at-risk (OAR) automatic segmentation (AS) using four atlas-based (ABAS) and two deep learning (DL) solutions. MATERIAL AND METHODS All patients underwent iodine contrast-enhanced planning CT. Fourteen OAR were manually delineated. DL.1 and DL.2 solutions were trained with 63 mono-centric patients and >1000 multi-centric patients, respectively. Ten and 15 patients with varied anatomies were selected for the atlas library and for testing, respectively. The evaluation was based on geometric indices (DICE coefficient and 95th percentile-Hausdorff Distance (HD95%)), time needed for manual corrections and clinical dosimetric endpoints obtained using automated treatment planning. RESULTS Both DICE and HD95% results indicated that DL algorithms generally performed better compared with ABAS algorithms for automatic segmentation of HN OAR. However, the hybrid-ABAS (ABAS.3) algorithm sometimes provided the highest agreement to the reference contours compared with the 2 DL. Compared with DL.2 and ABAS.3, DL.1 contours were the fastest to correct. For the 3 solutions, the differences in dose distributions obtained using AS contours and AS+manually corrected contours were not statistically significant. High dose differences could be observed when OAR contours were at short distances to the targets. However, this was not always interrelated. CONCLUSION DL methods generally showed higher delineation accuracy compared with ABAS methods for AS segmentation of HN OAR. Most ABAS contours had high conformity to the reference but were more time consuming than DL algorithms, especially when considering the computing time and the time spent on manual corrections.

Proton dose deposition results are influenced by various factors, such as irradiation angle, beamlet energy and other parameters. The calculation of the proton dose deposition matrix (DDM) can be highly complex but is crucial in intensity-modulated proton therapy (IMPT). In this work, we present a novel deep learning (DL) approach using multi-source features for proton DDM prediction. The DL5 proton DDM prediction method involves five input features containing beamlet geometry, dosimetry and treatment machine information like patient CT data, beamlet energy, distance from voxel to beamlet axis, distance from voxel to body surface, and pencil beam (PB) dose. The dose calculated by Monte Carlo (MC) method was used as the ground truth dose label. A total of 40 000 features, corresponding to 8000 beamlets, were obtained from head patient datasets and used for the training data. Additionally, seventeen head patients not included in the training process were utilized as testing cases. The DL5 method demonstrates high proton beamlet dose prediction accuracy, with an average determination coefficient R 2 of 0.93 when compared to the MC dose. Accurate beamlet dose estimation can be achieved in as little as 1.5 milliseconds for an individual proton beamlet. For IMPT plan dose comparisons to the dose calculated by the MC method, the DL5 method exhibited gamma pass rates of γ(2 mm, 2%) and γ(3 mm, 3%) ranging from 98.15% to 99.89% and 98.80% to 99.98%, respectively, across all 17 testing cases. On average, the DL5 method increased the gamma pass rates to γ(2 mm, 2%) from 82.97% to 99.23% and to γ(3 mm, 3%) from 85.27% to 99.75% when compared with the PB method. The proposed DL5 model enables rapid and precise dose calculation in IMPT plan, which has the potential to significantly enhance the efficiency and quality of proton radiation therapy.

Background Existing deep learning methods, such as generative adversarial network (GAN) technology, face challenges when dealing with mixed datasets, which involve a combination of Intensity Modulated Radiotherapy (IMRT) and Volumetric Modulated Arc Therapy (VMAT). This issue significantly complicates the application of dose prediction in the field of radiotherapy. In this study, we propose a novel approach called beam channel GAN (Bc-GAN) to address the task of radiation dose prediction for mixed datasets. Bc-GAN introduces a dose prediction calculation method that requires less precision. By defining an approximate range for dose prediction, Bc-GAN limits the physical range of GAN prediction, resulting in more reasonable dose distribution predictions. Methods We adopt a beam angle weighting method to determine the beam angle in the dose calculation. The dose of the beam with the highest weight is calculated using medical images and is then inputted into the artificial intelligence dose prediction model as the input channel. Additionally, we collect data from a total of 346 patients with Cervical Cancer (CC) for dataset. After cleaning the data, we exclude 51 cases with incomplete organ delineation, leaving us with 295 cases (IMRT: VMAT = 137:158) randomly divided into three sets: the training set, the validation set, and the test set, with proportions of 205:60:30, respectively. The assessment of model predictions was conducted via an analysis of dose distributions on the tomographic plane, dose volume histogram (DVH), and dosimetric parameters within the target zones and organs at risk (OAR). Results After DVH analysis, minimal discrepancy was found between predicted and actual dose distributions in PTV and OAR. The predicted distribution aligned with clinical standards. Dosimetric parameters for PTV were generally lower in the predicted model, except for homogeneity index (HI) (0.238 ± 0.024, P = 0.017) and Dmax (53.599 ± 0.710 Gy, P = 1.8e-05). The prediction model varied in estimating doses for six organs. Specifically, small intestine showed higher V20 (67.92 ± 51.64 %, P = 0.019) and V30 (57.171 ± 1.213 %, P = 0.024) than manual planning. A similar trend was seen in colon's V30 (37.13 ± 61.14 %, P = 0.016). However, predicted bladder V30 (87.51 ± 41.44 %, P = 2.03e-16) was lower, indicating significant dosimetric differences. Conclusion Overall, this study presents an innovative prediction method for CC in radiotherapy using the Bc-GAN model, addressing the challenges posed by different radiotherapy techniques. The proposed approach allows IMRT and VMAT in radiotherapy to be used as training sets, enabling the potential for large-scale engineering and commercialization applications of artificial intelligence (AI). The Bc-GAN-based prediction method for CC in radiotherapy not only reduces the amount of data needed for the training set but also expedites the model generation process. This approach can be applied to guide the development of clinical radiation therapy plans. Furthermore, future studies should consider extending the dose prediction method to encompass other types of tumors.

BACKGROUND Radiation therapy is one of the crucial treatment modalities for cancer. An excellent radiation therapy plan relies heavily on an outstanding dose distribution map, which is traditionally generated through repeated trials and adjustments by experienced physicists. However, this process is both time-consuming and labor-intensive, and it comes with a degree of subjectivity. Now, with the powerful capabilities of deep learning, we are able to predict dose distribution maps more accurately, effectively overcoming these challenges. METHODS In this study, we propose a novel Swin-UMamba-Channel prediction model specifically designed for predicting the dose distribution of patients with left breast cancer undergoing radiotherapy after total mastectomy. This model integrates anatomical position information of organs and ray angle information, significantly enhancing prediction accuracy. Through iterative training of the generator (Swin-UMamba) and discriminator, the model can generate images that closely match the actual dose, assisting physicists in quickly creating DVH curves and shortening the treatment planning cycle. Our model exhibits excellent performance in terms of prediction accuracy, computational efficiency, and practicality, and its effectiveness has been further verified through comparative experiments with similar networks. RESULTS The results of the study indicate that our model can accurately predict the clinical dose of breast cancer patients undergoing intensity-modulated radiation therapy (IMRT). The predicted dose range is from 0 to 50 Gy, and compared with actual data, it shows a high accuracy with an average Dice similarity coefficient of 0.86. Specifically, the average dose change rate for the planning target volume ranges from 0.28 % to 1.515 %, while the average dose change rates for the right and left lungs are 2.113 % and 0.508 %, respectively. Notably, due to their small sizes, the heart and spinal cord exhibit relatively higher average dose change rates, reaching 3.208 % and 1.490 %, respectively. In comparison with similar dose studies, our model demonstrates superior performance. Additionally, our model possesses fewer parameters, lower computational complexity, and shorter processing time, further enhancing its practicality and efficiency. These findings provide strong evidence for the accuracy and reliability of our model in predicting doses, offering significant technical support for IMRT in breast cancer patients. CONCLUSION This study presents a novel Swin-UMamba-Channel dose prediction model, and its results demonstrate its precise prediction of clinical doses for the target area of left breast cancer patients undergoing total mastectomy and IMRT. These remarkable achievements provide valuable reference data for subsequent plan optimization and quality control, paving a new path for the application of deep learning in the field of radiation therapy.

Deep learning has been utilized in knowledge-based radiotherapy planning in which a system trained with a set of clinically approved plans is employed to infer a three-dimensional dose map for a given new patient. However, previous deep methods are primarily limited to simple scenarios, e.g., a fixed planning type or a consistent beam angle configuration. This in fact limits the usability of such approaches and makes them not generalizable over a larger set of clinical scenarios. Herein, we propose a novel conditional generative model, Flexible-Cm GAN, utilizing additional information regarding planning types and various beam geometries. A miss-consistency loss is proposed to deal with the challenge of having a limited set of conditions on the input data, e.g., incomplete training samples. To address the challenges of including clinical preferences, we derive a differentiable shift-dose-volume loss to incorporate the well-known dose-volume histogram constraints. During inference, users can flexibly choose a specific planning type and a set of beam angles to meet the clinical requirements. We conduct experiments on an illustrative face dataset to show the motivation of Flexible-Cm GAN and further validate our model's potential clinical values with two radiotherapy datasets. The results demonstrate the superior performance of the proposed method in a practical heterogeneous radiotherapy planning application compared to existing deep learning-based approaches.

Abstract Purpose Balancing quality and efficiency has been a challenge for online adaptive therapy. Most systems start the online re‐optimization with the original planning goals. While some systems allow planners to modify the planning goals, achieving a high‐quality plan within time constraints remains a common barrier. This study aims to bolster plan quality by leveraging a deep‐learning dose prediction model to predict new planning goals that account for inter‐fractional anatomical changes. Methods Fine‐tuned patient‐specific (FT‐PS) models were clinically evaluated to accurately predict dose for 23 adaptive fractions of 15 head‐and‐neck (H&N) patients treated with Ethos ART. The original adapted plan from the adaptive treatment session was used as the quality baseline. Based on physician‐approved adaptive treatment contours, the FT‐PS model predicted subsequent planning goals for high‐impact organs at risk (OARs). These goals were retrospectively re‐optimized in Ethos to compare the original adapted plan (IOE‐Auto Plan) with the newly re‐optimized plan (AI‐guided IOE Plan). A physician blindly selected the preferred plan. Results Dose savings were observed for nine high impact OAR's including the constrictor, ipsilateral/contralateral parotid, ipsilateral/contralateral submandibular gland, oral cavity, and esophagus, mandible and larynx with a maximum value of 5.47 Gy. Of the 23 plans reviewed in the blind observer study, 19 re‐optimized plans were chosen over the original adapted session plan. Conclusions Our preliminary results demonstrate the feasibility of utilizing an AI dose predictor to predict optimal planning goals with anatomical changes, thereby improving adaptive plan quality. This method is feasible for both online and offline adaptive radiotherapy (ART) and has the potential to significantly enhance treatment outcomes for head‐and‐neck (H&N) cancer patients.

Abstract Background Automatic radiotherapy (RT) planning based on deep learning (DL) has been extensively researched. However, it is challenging to import the predicted dose distribution into mainstream treatment planning systems (TPSs) and generate clinically deliverable plans. Purpose To investigate the feasibility and accuracy of an automatic volumetric modulated arc therapy (VMAT) and intensity‐modulated radiation therapy (IMRT) planning method for generation of universally deliverable plans based on DL dose prediction and dose rings optimization. Methods First, dose distributions were predicted using a three‐dimensional (3D) Fusion Residual Unet (F‐ResUnet) DL network with data from two hospitals, which included 230 and 210 gynecological cancer (GC) patients underwent VMAT and IMRT, respectively. Then, the predicted dose distributions were discretized into dose rings to optimize the plans automatically in two mainstream TPSs based on the dose rings. Finally, the deliverability of generated plans was verified with patient‐specific quality assurance (PSQA). Results The predicted dose distributions were clinically acceptable with a target coverage over 95%. Compared with the clinical plans, the automatic plans optimized with dose rings achieved a similar dose coverage on planning target volumes (PTV) with an average target coverage over 96.5%. For organs at risk (OARs) sparing, automatic VMAT plans markedly decreased the V30Gy of left femoral head (p = 0.05), right femoral head (p = 0.004), and small intestine (p = 0.04). The V45Gy of bladder and rectum in the automatic IMRT plans were reduced by approximately 7% and 9%, respectively. Deliverability verification with PSQA achieved a mean gamma passing rate of 99.1%, 97.1% and 98.3%, 95.0% under the criteria of 3%/3 mm and 3%/2 mm for VMAT and IMRT plans, respectively. Conclusions The proposed automatic planning method combining DL dose prediction and dose rings optimization was feasible to generate universally deliverable VMAT and IMRT plans for gynecological cancer (GC) patients.

Objective. To evaluate the feasibility of using a deep learning dose prediction approach to identify patients who could benefit most from proton therapy based on the normal tissue complication probability (NTCP) model. Approach. Two 3D UNets were established to predict photon and proton doses. A dataset of 95 patients with localized prostate cancer was randomly partitioned into 55, 10, and 30 for training, validation, and testing, respectively. We selected NTCP models for late rectum bleeding and acute urinary urgency of grade 2 or higher to quantify the benefit of proton therapy. Propagated uncertainties of predicted ΔNTCPs resulting from the dose prediction errors were calculated. Patient selection accuracies for a single endpoint and a composite evaluation were assessed under different ΔNTCP thresholds. Main results. Our deep learning-based dose prediction technique can reduce the time spent on plan comparison from approximately 2 days to as little as 5 seconds. The expanded uncertainty of predicted ΔNTCPs for rectum and bladder endpoints propagated from the dose prediction error were 0.0042 and 0.0016, respectively, which is less than one-third of the acceptable tolerance. The averaged selection accuracies for rectum bleeding, urinary urgency, and composite evaluation were 90%, 93.5%, and 93.5%, respectively. Significance. Our study demonstrates that deep learning dose prediction and NTCP evaluation scheme could distinguish the NTCP differences between photon and proton treatment modalities. In addition, the dose prediction uncertainty does not significantly influence the decision accuracy of NTCP-based patient selection for proton therapy. Therefore, automated deep learning dose prediction and NTCP evaluation schemes can potentially be used to screen large patient populations and to avoid unnecessary delays in the start of prostate cancer radiotherapy in the future.

BACKGROUND AND PURPOSE Fast, high-quality deep learning (DL) prediction of patient-specific 3D dose distributions can enable instantaneous treatment planning (IP), in which the treating physician can evaluate the dose and approve the plan immediately after contouring, rather than days later. This would greatly benefit clinical workload, patient waiting times and treatment quality. IP requires that predicted dose distributions closely match the ground truth. This study examines how training dataset size and model size affect dose prediction accuracy for Erasmus-iCycle GT plans to enable IP. MATERIALS AND METHODS For 1250 prostate patients, dose distributions were automatically generated using Erasmus-iCycle. Hierarchically Densely Connected U-Nets with 2/3/4/5/6 pooling layers were trained with datasets of 50/100/250/500/1000 patients, using a validation set of 100 patients. A fixed test set of 150 patients was used for evaluations. RESULTS For all model sizes, prediction accuracy increased with the number of training patients, without levelling off at 1000 patients. For 4-6 level models with 1000 training patients, prediction accuracies were high and comparable. For 6 levels and 1000 training patients, the median prediction errors and interquartile ranges for PTV V95%, rectum V75Gy and bladder V65Gy were 0.01 [-0.06,0.15], 0.01 [-0.20,0.29] and -0.02 [-0.27,0.27] %-point. Dose prediction times were around 1.2 s. CONCLUSION Although even for 1000 training patients there was no convergence in obtained prediction accuracy yet, the accuracy for the 6-level model with 1000 training patients may be adequate for the pursued instantaneous planning, which is subject of further research.

Purpose: Deep learning (DL) is widely used in dose prediction for radiation oncology, multiple DL techniques comparison is often lacking in the literature. To compare the performance of 4 state-of-the-art DL models in predicting the voxel-level dose distribution for cervical cancer volumetric modulated arc therapy (VMAT). Methods and Materials: A total of 261 patients’ plans for cervical cancer were retrieved in this retrospective study. A three-channel feature map, consisting of a planning target volume (PTV) mask, organs at risk (OARs) mask, and CT image was fed into the three-dimensional (3D) U-Net and its 3 variants models. The data set was randomly divided into 80% as training-validation and 20% as testing set, respectively. The model performance was evaluated on the 52 testing patients by comparing the generated dose distributions against the clinical approved ground truth (GT) using mean absolute error (MAE), dose map difference (GT-predicted), clinical dosimetric indices, and dice similarity coefficients (DSC). Results: The 3D U-Net and its 3 variants DL models exhibited promising performance with a maximum MAE within the PTV 0.83% ± 0.67% in the UNETR model. The maximum MAE among the OARs is the left femoral head, which reached 6.95% ± 6.55%. For the body, the maximum MAE was observed in UNETR, which is 1.19 ± 0.86%, and the minimum MAE was 0.94 ± 0.85% for 3D U-Net. The average error of the Dmean difference for different OARs is within 2.5 Gy. The average error of V40 difference for the bladder and rectum is about 5%. The mean DSC under different isodose volumes was above 90%. Conclusions: DL models can predict the voxel-level dose distribution accurately for cervical cancer VMAT treatment plans. All models demonstrated almost analogous performance for voxel-wise dose prediction maps. Considering all voxels within the body, 3D U-Net showed the best performance. The state-of-the-art DL models are of great significance for further clinical applications of cervical cancer VMAT.

No abstract available

Abstract Background Intensity‐Modulated Radiation Therapy (IMRT) has been the standard of care for many types of tumors. However, treatment planning for IMRT is a time‐consuming and labor‐intensive process. Purpose To alleviate this tedious planning process, a novel deep learning based dose prediction algorithm (TrDosePred) was developed for head and neck cancers. Methods The proposed TrDosePred, which generated the dose distribution from a contoured CT image, was a U‐shape network constructed with a convolutional patch embedding and several local self‐attention based transformers. Data augmentation and ensemble approach were used for further improvement. It was trained based on the dataset from Open Knowledge‐Based Planning Challenge (OpenKBP). The performance of TrDosePred was evaluated with two mean absolute error (MAE) based scores utilized by OpenKBP challenge (i.e., Dose score and DVH score) and compared to the top three approaches of the challenge. In addition, several state‐of‐the‐art methods were implemented and compared to TrDosePred. Results The TrDosePred ensemble achieved the dose score of 2.426 Gy and the DVH score of 1.592 Gy on the test dataset, ranking at 3rd and 9th respectively in the leaderboard on CodaLab as of writing. In terms of DVH metrics, on average, the relative MAE against the clinical plans was 2.25% for targets and 2.17% for organs at risk. Conclusions A transformer‐based framework TrDosePred was developed for dose prediction. The results showed a comparable or superior performance as compared to the previous state‐of‐the‐art approaches, demonstrating the potential of transformer to boost the treatment planning procedures.

BACKGROUND In cancer care, determining the most beneficial treatment technique is a key decision affecting the patient's survival and quality of life. Patient selection for proton therapy (PT) over conventional radiotherapy (XT) currently entails comparing manually generated treatment plans, which requires time and expertise. PURPOSE We developed an automatic and fast tool, AI-PROTIPP (Artificial Intelligence Predictive Radiation Oncology Treatment Indication to Photons/Protons), that assesses quantitatively the benefits of each therapeutic option. Our method uses deep learning (DL) models to directly predict the dose distributions for a given patient for both XT and PT. By using models that estimate the Normal Tissue Complication Probability (NTCP), namely the likelihood of side effects to occur for a specific patient, AI-PROTIPP can propose a treatment selection quickly and automatically. METHODS A database of 60 patients presenting oropharyngeal cancer, obtained from the Cliniques Universitaires Saint Luc in Belgium, was used in this study. For every patient, a PT plan and an XT plan were generated. The dose distributions were used to train the two dose DL prediction models (one for each modality). The model is based on U-Net architecture, a type of convolutional neural network currently considered as the state of the art for dose prediction models. A NTCP protocol used in the Dutch model-based approach, including grades II and III xerostomia and grades II and III dysphagia, was later applied in order to perform automatic treatment selection for each patient. The networks were trained using a nested cross-validation approach with 11-folds. We set aside three patients in an outer set and each fold consists of 47 patients in training, five in validation and five for testing. This method allowed us to assess our method on 55 patients (five patients per test times the number of folds). RESULTS The treatment selection based on the DL-predicted doses reached an accuracy of 87.4% for the threshold parameters set by the Health Council of the Netherlands. The selected treatment is directly linked with these threshold parameters as they express the minimal gain brought by the PT treatment for a patient to be indicated to PT. To validate the performance of AI-PROTIPP in other conditions, we modulated these thresholds, and the accuracy was above 81% for all the considered cases. The difference in average cumulative NTCP per patient of predicted and clinical dose distributions is very similar (less than 1% difference). CONCLUSIONS AI-PROTIPP shows that using DL dose prediction in combination with NTCP models to select PT for patients is feasible and can help to save time by avoiding the generation of treatment plans only used for the comparison. Moreover, DL models are transferable, allowing, in the future, experience to be shared with centers that would not have PT planning expertise.

Abstract Purpose The purpose of this study is to investigate the use of a deep learning architecture for automated treatment planning for proton pencil beam scanning (PBS). Methods A 3‐dimensional (3D) U‐Net model has been implemented in a commercial treatment planning system (TPS) that uses contoured regions of interest (ROI) binary masks as model inputs with a predicted dose distribution as the model output. Predicted dose distributions were converted to deliverable PBS treatment plans using a voxel‐wise robust dose mimicking optimization algorithm. This model was leveraged to generate machine learning (ML) optimized plans for patients receiving proton PBS irradiation of the chest wall. Model training was carried out on a retrospective set of 48 previously‐treated chest wall patient treatment plans. Model evaluation was carried out by generating ML‐optimized plans on a hold‐out set of 12 contoured chest wall patient CT datasets from previously treated patients. Clinical goal criteria and gamma analysis were used to compare dose distributions of the ML‐optimized plans against the clinically approved plans across the test patients. Results Statistical analysis of mean clinical goal criteria indicates that compared to the clinical plans, the ML optimization workflow generated robust plans with similar dose to the heart, lungs, and esophagus while achieving superior dosimetric coverage to the PTV chest wall (clinical mean V95 = 97.6% vs. ML mean V95 = 99.1%, p < 0.001) across the 12 test patients. Conclusions ML‐based automated treatment plan optimization using the 3D U‐Net model can generate treatment plans of similar clinical quality compared to human‐driven optimization.

No abstract available

Background and purpose Treatment planning of radiotherapy is a time-consuming and planner dependent process that can be automated by dose prediction models. The purpose of this study was to evaluate the performance of two machine learning models for breast cancer radiotherapy before possible clinical implementation. Materials and methods An in-house developed model, based on U-net architecture, and a contextual atlas regression forest (cARF) model integrated in the treatment planning software were trained. Obtained dose distributions were mimicked to create clinically deliverable plans. For training and validation, 90 patients were used, 15 patients were used for testing. Treatment plans were scored on predefined evaluation criteria and percent errors with respect to clinical dose were calculated for doses to planning target volume (PTV) and organs at risk (OARs). Results The U-net plans before mimicking met all criteria for all patients, both models failed one evaluation criterion in three patients after mimicking. No significant differences (p < 0.05) were found between clinical and predicted U-net plans before mimicking. Doses to OARs in plans of both models differed significantly from clinical plans, but no clinically relevant differences were found. After mimicking, both models had a mean percent error within 1.5% for the average dose to PTV and OARs. The mean errors for maximum doses were higher, within 6.6%. Conclusions Differences between predicted doses to OARs of the models were small when compared to clinical plans, and not found to be clinically relevant. Both models show potential in automated treatment planning for breast cancer.

BACKGROUND AND PURPOSE This study aims to investigate how accurate our deep learning (DL) dose prediction models for intensity modulated radiotherapy (IMRT) and pencil beam scanning (PBS) treatments, when chained with normal tissue complication probability (NTCP) models, are at identifying esophageal cancer patients who are at high risk of toxicity and should be switched to proton therapy (PT). MATERIALS AND METHODS Two U-Net were created, for photon (XT) and proton (PT) plans, respectively. To estimate the dose distribution for each patient, they were trained on a database of 40 uniformly planned patients using cross validation and a circulating test set. These models were combined with a NTCP model for postoperative pulmonary complications. The NTCP model used the mean lung dose, age, histology type, and body mass index as predicting variables. The treatment choice is then done by using a ΔNTCP threshold between XT and PT plans. Patients with ΔNTCP ≥ 10% were referred to PT. RESULTS Our DL models succeed in predicting dose distributions with a mean error on the mean dose to the lungs (MLD) of 1.14 ± 0.93 % for XT and 0.66 ± 0.48 % for PT. The complete automated workflow (DL chained with NTCP) achieved 100 % accuracy in patient referral. The average residual (ΔNTCP ground truth - ΔNTCP predicted) is 1.43 ± 1.49 %. CONCLUSION This study evaluates our DL dose prediction models in a broader patient referral context and demonstrates their ability to support clinical decisions.

Abstract Purpose To develop a 3D‐Unet dose prediction model to predict the three‐dimensional dose distribution of volumetric modulated arc therapy (VMAT) for cervical cancer and test the dose prediction performance of the model in endometrial cancer to explore the feasibility of model generalization. Methods One hundred and seventeen cases of cervical cancer and 20 cases of endometrial cancer treated with VMAT were used for the model training, validation, and test. The prescribed dose was 50.4 Gy in 28 fractions. Eight independent channels of contoured structures were input to the model, and the dose distribution was used as the output of the model. The 3D‐Unet prediction model was trained and validated on the training set (n = 86) and validation set (n = 11), respectively. Then the model was tested on the test set (n = 20) of cervical cancer and endometrial cancer, respectively. The results between clinical dose distribution and predicted dose distribution were compared in the following aspects: (a) the mean absolute error (MAE) within the body, (b) the Dice similarity coefficients (DSCs) under different isodose volumes, (c) the dosimetric indexes including the mean dose (D mean), the received dose of 2 cm3 (D 2cc), the percentage volume of receiving 40 Gy dose of organs‐at‐risk (V 40), planning target volume (PTV) D 98%, and homogeneity index (HI), (d) dose–volume histograms (DVHs). Results The model can accurately predict the dose distribution of the VMAT plan for cervical cancer and endometrial cancer. The overall average MAE and maximum MAE for cervical cancer were 2.43 ± 3.17% and 3.16 ± 4.01% of the prescribed dose, respectively, and for endometrial cancer were 2.70 ± 3.54% and 3.85 ± 3.11%. The average DSCs under different isodose volumes is above 0.9. The predicted dosimetric indexes and DVHs are equivalent to the clinical dose for both cervical cancer and endometrial cancer, and there is no statistically significant difference. Conclusion A 3D‐Unet dose prediction model was developed for VMAT of cervical cancer, which can predict the dose distribution accurately for cervical cancer. The model can also be generalized for endometrial cancer with good performance.

Simple Summary This work describes the development of a fast and accurate machine learning (ML) 3D U-Net dose engine, trained with Monte Carlo (MC) radiation transport simulations, to calculate the dose in rat patients treated in Microbeam Radiation Therapy (MRT) preclinical studies at the Imaging and Medical Beamline at the Australian Synchrotron. Digital phantoms are created based on CT scans of sixteen rats and are augmented to obtain enough anatomical data. Augmented variations of the digital phantoms are then used to simulate with Geant4 the energy depositions of an MRT beam inside the phantoms with 15% (high-noise) and 2% (low-noise) statistical uncertainty. The high-noise MC simulations are used for ML model training and validation, while the low-noise ones for testing. The results show that the ML dose engine provides a satisfactory dose description in the tumor target and generates the dose maps in less than one second. Abstract Microbeam radiation therapy (MRT) utilizes coplanar synchrotron radiation beamlets and is a proposed treatment approach for several tumor diagnoses that currently have poor clinical treatment outcomes, such as gliosarcomas. Monte Carlo (MC) simulations are one of the most used methods at the Imaging and Medical Beamline, Australian Synchrotron to calculate the dose in MRT preclinical studies. The steep dose gradients associated with the 50μm-wide coplanar beamlets present a significant challenge for precise MC simulation of the dose deposition of an MRT irradiation treatment field in a short time frame. The long computation times inhibit the ability to perform dose optimization in treatment planning or apply online image-adaptive radiotherapy techniques to MRT. Much research has been conducted on fast dose estimation methods for clinically available treatments. However, such methods, including GPU Monte Carlo implementations and machine learning (ML) models, are unavailable for novel and emerging cancer radiotherapy options such as MRT. In this work, the successful application of a fast and accurate ML dose prediction model for a preclinical MRT rodent study is presented for the first time. The ML model predicts the peak doses in the path of the microbeams and the valley doses between them, delivered to the tumor target in rat patients. A CT imaging dataset is used to generate digital phantoms for each patient. Augmented variations of the digital phantoms are used to simulate with Geant4 the energy depositions of an MRT beam inside the phantoms with 15% (high-noise) and 2% (low-noise) statistical uncertainty. The high-noise MC simulation data are used to train the ML model to predict the energy depositions in the digital phantoms. The low-noise MC simulations data are used to test the predictive power of the ML model. The predictions of the ML model show an agreement within 3% with low-noise MC simulations for at least 77.6% of all predicted voxels (at least 95.9% of voxels containing tumor) in the case of the valley dose prediction and for at least 93.9% of all predicted voxels (100.0% of voxels containing tumor) in the case of the peak dose prediction. The successful use of high-noise MC simulations for the training, which are much faster to produce, accelerates the production of the training data of the ML model and encourages transfer of the ML model to different treatment modalities for other future applications in novel radiation cancer therapies.

Accurate pre-treatment dose prediction is essential for efficient radiotherapy planning. Although deep learning models have advanced automated dose distribution, comprehensive multi-tumor analyses remain scarce. This study assesses deep learning models for dose prediction across diverse tumor types, combining objective and subjective evaluation methods. We included 622 patients with planning data across various tumor sites: nasopharyngeal carcinoma (n = 29), esophageal carcinoma (n = 82), left-sided breast carcinoma (n = 107), right-sided breast carcinoma (n = 95), cervical carcinoma treated with radical radiotherapy (n = 84), postoperative cervical carcinoma (n = 122), and rectal carcinoma (n = 103). Dose predictions were generated using U-Net, Flex-Net, and Highres-Net models, with data split into training (60%), validation (20%), and testing (20%) sets. Quantitative comparisons used normalized dose difference (NDD) and dose-volume histogram (DVH) metrics, and qualitative assessments by radiation oncologists were performed on the testing set. Predicted and clinical doses correlated well, with NDD values under 3% for tumor targets in nasopharyngeal, breast, and postoperative cervical cancer. Qualitative assessments revealed that U-Net, Flex-Net, and Highres-Net achieved the highest accuracy in cervical radical, breast/rectal/postoperative cervical, and nasopharyngeal/esophageal cancers, respectively. Among the test cases (n = 123), 53.7% were deemed clinically acceptable and 32.5% required minor adjustments. The “Best Selection” approach, combining strengths of all three models, raised clinical acceptance to 62.6%. This study demonstrates that automated dose prediction can provide a robust starting point for rapid plan generation. Leveraging model-specific strengths through the “Best Selection” approach enhances prediction accuracy and shows potential to improve clinical efficiency across multiple tumor types.

BACKGROUND Recent studies have shown deep learning techniques are able to predict three-dimensional (3D) dose distributions of radiotherapy treatment plans. However, their use in dose prediction for treatments with varied prescription doses including simultaneous integrated boost (SIB), that is, using multiple prescription doses within the same plan, and benefit in improving plan quality should be validated. PURPOSE To investigate the feasibility and potential benefit of using deep learning to predict dose distribution of volumetric modulated arc therapy (VMAT) including SIB techniques and improve treatment planning for patients with lung cancer. METHODS The dose prediction model was trained with 93 retrospective clinical VMAT plans for patients with lung cancer from our institutional patient database. The prescription doses of these plans ranged from 35 to 72 Gy, with various fractionation schemes. We used a 3D U-Net architecture to predict 3D dose distributions with 75 plans for training and 18 plans for testing. Model input consisted of computed tomography (CT) images, target and normal tissue contours and prescription doses. We first evaluated model accuracy by comparing the predicted and clinical plan doses for the test set, and then performed replanning according to predicted dose distributions. Furthermore, we evaluated the model prospectively in an additional set of 10 patients from our institution by two approaches where dose prediction was either blinded or provided to treatment planners. We then assessed whether dose prediction could identify suboptimal plan quality and how it affects plan quality if adopted in clinical planning workflow. RESULTS The dose prediction model achieved good agreement between the predicted and clinical plan dose distributions, with a mean dose difference of -0.49 ± 0.54 Gy across the test set. The replanning study guided by dose prediction showed that a small subset of the original plans could benefit from improvements regarding sparing of the spinal cord and esophagus. The analysis of the prospective dataset, with initial and final clinical plans generated in the absence of dose prediction, showed that the predicted doses were able to identify possible improvements of target coverage and normal tissue sparing in the initial plans similar to those made by the final plans for majority of the patients, but in varied magnitudes. Moreover, the plans generated with dose prediction guidance were able to consistently improve normal tissue sparing compared to the plans generated without dose prediction guidance. CONCLUSIONS We demonstrated that our deep learning model can consistently predict high quality VMAT lung plans for a variety of prescription doses. The dose prediction tool was also effective in identifying suboptimal plan quality, suggesting its potential benefit in automated treatment planning and evaluation.

Online adaptive radiation therapy (ART) personalizes treatment plans by accounting for daily anatomical changes, requiring workflows distinct from conventional radiotherapy. Deep learning-based dose prediction models can enhance treatment planning efficiency by rapidly generating accuracy dose distributions, reducing manual trial-and-error and accelerating the overall workflow; however, most existing approaches overlook critical pre-treatment plan information—specifically, physician-defined clinical objectives tailored to individual patients. To address this limitation, we introduce the multi-headed U-Net (MHU-Net), a novel architecture that explicitly incorporates physician intent from pre-treatment plans to improve dose prediction accuracy in adaptive head and neck cancer treatments. Our dataset comprised 43 patients, each with pre-treatment plans, adaptive treatment plans, structure sets, and CT images. MHU-Net builds upon the widely adopted Stander U-Net architecture, extending it with a dual-head design: the primary head processes adaptive session contours and their corresponding signed distance maps, while the secondary head integrates pre-treatment contours, signed distance maps, and dose distributions. The features are merged within a primary U-Net framework to enhance dose prediction accuracy for adaptive treatment sessions. To evaluate the effectiveness of MHU-Net, we conducted a comparative analysis against U-Net. On average, MHU-Net reduced organ-at-risk dose prediction errors, achieving 1.78% lower maximum dose error and 1.22% lower mean dose error compared to U-Net. For the planning target volume, MHU-Net demonstrated significantly improved accuracy, with maximum and mean dose errors of 3.54  ±  2.75% and 1.07 ± 0.88%, respectively, outperforming U-Net’s corresponding errors of 5.36 ± 4.19% and 2.76 ± 2.22% (P < 0.05). Taken together, these findings demonstrate that the proposed MHU-Met advances DL-based dose prediction for ART by effectively integrating both pre-treatment and adaptive session data. This approach facilitates the generation of dose distributions that more closely resemble the clinical ground truth, supporting personalization in ART planning and improving alignment with physician intent and treatment objectives.

Objective. This study investigates key factors influencing deep learning-based dose prediction models for head and neck cancer radiation therapy. The goal is to evaluate model accuracy, robustness, and computational efficiency, and to identify key components necessary for optimal performance. Approach. We systematically analyze the impact of input and dose grid resolution, input type, loss function, model architecture, and noise on model performance. Two datasets are used: a public dataset (OpenKBP) and an in-house clinical dataset. Model performance is primarily evaluated using two metrics: dose score and dose–volume histogram (DVH) score. Main results. High-resolution inputs improve prediction accuracy (dose score and DVH score) by 8.6%–13.5% compared to low resolution. Using a combination of CT, planning target volumes, and organs-at-risk as input significantly enhances accuracy, with improvements of 57.4%–86.8% over using CT alone. Integrating mean absolute error (MAE) loss with value-based and criteria-based DVH loss functions further boosts DVH score by 7.2%–7.5% compared to MAE loss alone. In the robustness analysis, most models show minimal degradation under Poisson noise (0–0.3 Gy) but are more susceptible to adversarial noise (0.2–7.8 Gy). Notably, certain models, such as SwinUNETR, demonstrate superior robustness against adversarial perturbations. Significance. These findings highlight the importance of optimizing deep learning models and provide valuable guidance for achieving more accurate and reliable radiotherapy dose prediction.

Radiation therapy (RT) is a cornerstone in the management of localized and locally advanced prostate cancer, traditionally delivered with a full bladder (FB) protocol to reduce radiation exposure to surrounding organs. However, consistent bladder filling is difficult to maintain, leading to workflow delays, anatomical inconsistencies, and variable toxicity outcomes. Recent evidence, including the ongoing RELIEF trial at Mayo Clinic, suggests that an empty bladder (EB) protocol provides comparable toxicity outcomes to FB while improving patient comfort and treatment consistency. To address the increased anatomical variability associated with EB protocols, we developed a deep learning (DL)-based dose prediction model tailored to EB patients. A conditional generative adversarial network (cGAN) with a modified 3D U-Net architecture was trained on 90 FB cases and fine-tuned on 20 EB cases stratified into stereotactic body radiotherapy (SBRT) and intensity-modulated radiotherapy (IMRT). Model performance was evaluated against clinical manual plans using mean absolute percentage error (MAPE) and dose-volume histogram (DVH) metrics. The EB Fine-tuning model(SBRT/IMRT) achieved superior accuracy compared with the general FB-trained model, with an average MAPE of 3.53 ± 0.40% versus 4.87 ± 0.86%. DVH analyses demonstrated improved agreement with manual plans for planning target volumes and organs at risk, with discrepancies consistently within 2.5 Gy or 3%. These results demonstrate that fine-tuning with EB-specific data enhances prediction accuracy and clinical relevance of the DL-based model. The proposed framework supports efficient EB treatment planning, provides reference dose distributions for quality assurance, and offers educational value to clinicians adopting EB protocols. By combining automation with clinical applicability, this approach facilitates broader adoption of EB radiotherapy in prostate cancer while improving workflow reproducibility and patient-centered care.

Purpose. The dose distribution of lung cancer patients treated with the CyberKnife (CK) system is influenced by various factors, including tumor location and the direction of CK beams. The objective of this study is to present a deep learning approach that integrates CK beam dose characteristics into CK planning dose calculations. Methods. The inputs utilized for the geometry and dosimetry method (GDM) include the patient’s CT, the PTV structure, and multiple CK noncoplanar beam dose deposition features. The dose distributions were calculated using the Monte Carlo (MC) algorithm provided with the CK system and served as the ground truth dose label. Additionally, dose prediction was conducted through the geometry method (GM) for comparative analysis. The gamma pass rate γ(1 mm,1%), γ(2 mm,2%) and γ(3 mm,3%) were calculated between the predicted model and the MC method. Results. Compared to the GDM, the GM shows a significant dose difference from the MC approach in the low-dose region (<5 Gy) outside the target created by the various CK noncoplanar beams. The GDM increased the γ(1 mm, 1%) from 49.55% to 81.69%, γ(2 mm, 2%) from 73.24% to 98.11% and the γ(3 mm, 3%) from 81.69% to 99.37% when compared with the GM’s results. Conclusions. This work proposed a deep learning dose calculation method by using patient geometry and dosimetry features in CK plans. The proposed method extends the geometric and dosimetric feature-driven deep learning dose calculation method to CK application scenarios, which has a great potential to accelerate the CK planning dose calculation and improve the planning efficiency.

Online adaptive radiotherapy (OART) and rapid quality assurance (QA) are essential for effective heavy ion therapy (HIT). However, there is a shortage of deep learning (DL) models and workflows for predicting Monte Carlo (MC) doses in such treatments.

Objective. To evaluate the impact of beam mask implementation and data aggregation on artificial intelligence-based dose prediction accuracy in proton therapy, with a focus on scenarios involving limited or highly heterogeneous datasets. Approach. In this study, 541 prostate and 632 head and neck (H&N) proton therapy plans were used to train and evaluate convolutional neural networks designed for the task of dose prediction. Datasets were grouped by anatomical site and beam configuration to assess the impact of beam masks—graphical depictions of radiation paths—as a model input. We also evaluated the effect of combining datasets. Model performance was measured using dose-volume histograms (DVHs) scores, mean absolute error, mean absolute percent error, dice similarity coefficients (DSCs), and gamma passing rates. Main results. DSC analysis revealed that the inclusion of beam masks improved dose prediction accuracy, particularly in low-dose regions and for datasets with diverse beam configurations. Data aggregation alone produced mixed results, with improvements in high-dose regions but potential degradation in low-dose areas. Notably, combining beam masks and data aggregation yielded the best overall performance, effectively leveraging the strengths of both strategies. Additionally, the magnitude of the improvements was larger for datasets with greater heterogeneity, with the combined approach increasing the DSC score by as much as 0.2 for a subgroup of H&N cases characterized by small size and heterogeneity in beam arrangement. DVH scores reflected these benefits, showing statistically significant improvements (p < 0.05) for the more heterogeneous H&N datasets. Significance. Artificial intelligence-based dose prediction models incorporating beam masks and data aggregation significantly improve accuracy in proton therapy planning, especially for complex cases. This technique could accelerate the planning process, enabling more efficient and effective cancer treatment strategies.

Accurate surface dose calculation is crucial in superficial low‐energy electron beam radiotherapy owing to shallow treatment depths and the risk of skin toxicity. Traditional Monte Carlo (MC) simulations are precise but computationally expensive and time‐consuming.

Purpose Difficulties remain in dose optimization and evaluation of cervical cancer radiotherapy that combines external beam radiotherapy (EBRT) and brachytherapy (BT). This study estimates and improves the accumulated dose distribution of EBRT and BT with deep learning–based dose prediction. Materials and methods A total of 30 patients treated with combined cervical cancer radiotherapy were enrolled in this study. The dose distributions of EBRT and BT plans were accumulated using commercial deformable image registration. A ResNet-101–based deep learning model was trained to predict pixel-wise dose distributions. To test the role of the predicted accumulated dose in clinic, each EBRT plan was designed using conventional method and then redesigned referencing the predicted accumulated dose distribution. Bladder and rectum dosimetric parameters and normal tissue complication probability (NTCP) values were calculated and compared between the conventional and redesigned accumulated doses. Results The redesigned accumulated doses showed a decrease in mean values of V50, V60, and D2cc for the bladder (−3.02%, −1.71%, and −1.19 Gy, respectively) and rectum (−4.82%, −1.97%, and −4.13 Gy, respectively). The mean NTCP values for the bladder and rectum were also decreased by 0.02‰ and 0.98%, respectively. All values had statistically significant differences (p < 0.01), except for the bladder D2cc (p = 0.112). Conclusion This study realized accumulated dose prediction for combined cervical cancer radiotherapy without knowing the BT dose. The predicted dose served as a reference for EBRT treatment planning, leading to a superior accumulated dose distribution and lower NTCP values.

AIMS Accurate dose delivery is crucial for cervical cancer volumetric modulated arc therapy (VMAT). We aimed to develop a robust deep-learning (DL) algorithm for fast and accurate dose prediction of cervical cancer VMAT in multicenter datasets and then explore the feasibility of the DL algorithm to endometrial cancer VMAT with different prescriptions. MATERIALS AND METHODS We proposed the AtTranNet algorithm for three-dimensional dose prediction. A total of 367 cervical patients were enrolled in this study. Three hundred twenty-two cervical patients from 3 centers were randomly divided into 70%, 10%, and 20% as training, validation, and testing sets, respectively. Forty-five cervical patients from another center were selected for external testing. Moreover, 70 patients of endometrial cancer with different prescriptions were further selected to test the model. Prediction precision was evaluated by dosimetric difference, dose map, and dose-volume histogram metrics. RESULTS The prediction results were all clinically acceptable. The mean absolute error within the body in internal testing was 0.66 ± 0.63%. The maximum |δD| for planning target volume was observed in D98, which is 1.24 ± 2.73 Gy. The maximum |δD| for organs at risk was observed in Dmean of bladder, which is 4.79 ± 3.14 Gy. The maximum |δV| were observed in V40 of pelvic bones, which is 4.77 ± 4.48%. CONCLUSION AtTranNet showed the feasibility and reasonable accuracy in the dose prediction for cervical cancer in multiple centers. The model can also be generalized for endometrial cancer with different prescriptions without any transfer learning.

Objective. Monte Carlo (MC) simulations are the benchmark for accurate radiotherapy dose calculations, notably in patient-specific high dose rate brachytherapy (HDR BT), in cases where considering tissue heterogeneities is critical. However, the lengthy computational time limits the practical application of MC simulations. Prior research used deep learning (DL) for dose prediction as an alternative to MC simulations. While accurate dose predictions akin to MC were attained, graphics processing unit limitations constrained these predictions to large voxels of 3 mm × 3 mm × 3 mm. This study aimed to enable dose predictions as accurate as MC simulations in 1 mm × 1 mm × 1 mm voxels within a clinically acceptable timeframe. Approach. Computed tomography scans of 98 breast cancer patients treated with Iridium-192-based HDR BT were used: 70 for training, 14 for validation, and 14 for testing. A new cropping strategy based on the distance to the seed was devised to reduce the volume size, enabling efficient training of 3D DL models using 1 mm × 1 mm × 1 mm dose grids. Additionally, novel DL architecture with layer-level fusion were proposed to predict MC simulated dose to medium-in-medium (D m,m ). These architectures fuse information from TG-43 dose to water-in-water (D w,w ) with patient tissue composition at the layer-level. Different inputs describing patient body composition were investigated. Main results. The proposed approach demonstrated state-of-the-art performance, on par with the MC D m,m maps, but 300 times faster. The mean absolute percent error for dosimetric indices between the MC and DL-predicted complete treatment plans was 0.17% ± 0.15% for the planning target volume V 100, 0.30% ± 0.32% for the skin D 2cc , 0.82% ± 0.79% for the lung D 2cc , 0.34% ± 0.29% for the chest wall D 2cc and 1.08% ± 0.98% for the heart D 2cc . Significance. Unlike the time-consuming MC simulations, the proposed novel strategy efficiently converts TG-43 D w,w maps into precise D m,m maps at high resolution, enabling clinical integration.

BACKGROUND Daily adaptive radiotherapy, as performed with the Elekta Unity MR-Linac, requires choosing between different adaptation methods, namely ATP (Adapt to Position) and ATS (Adapt to Shape), where the latter requires daily re-contouring to obtain a dose plan tailored to the daily anatomy. These steps are inherently resource-intensive, and quickly predicting the dose distribution and the dosimetric evaluation criteria while the patient is on the table could facilitate a fast selection of adaptation method and decrease the treatment times. PURPOSE In this work, we aimed to develop a deep-learning-based dose-prediction pipeline for prostate MR-Linac treatments. METHODS Two hundred twelve MR-images, structure sets, and dose distributions from 35 prostate patients treated with 6.1 Gy for 7 or 6 fractions at our MR-Linac were included, split into train/test partitions of 152/60 images, respectively. A deep-learning segmentation network was trained to segment the CTV (prostate), bladder, and rectum. A second network was trained to predict the dose distribution based on manually delineated structures. At inference, the predicted segmentations acted as input to the dose prediction network, and the predicted dose was compared to the true (optimized in the treatment planning system) dose distribution. RESULTS Median DSC values from the segmentation network were 0.90/0.94/0.87 for CTV/bladder/rectum. Predicted segmentations as input to the dose prediction resulted in mean differences between predicted and true doses of 0.7%/0.7%/1.7% (relative to the prescription dose) for D98%/D95%/D2% for the CTV. For the bladder, the difference was 0.7%/0.3% for Dmean/D2% and for the rectum 0.1/0.2/0.2 pp (percentage points) for V33Gy/V38Gy/V41Gy. In comparison, true segmentations as input resulted in differences of 1.1%/0.9%/1.6% for CTV, 0.5%/0.4% for bladder, and 0.7/0.4/0.3 pp for the rectum. Only D2% for CTV and Dmean/D2% for bladder were found to be statistically significantly better when using true structures instead of predicted structures as input to the dose prediction. CONCLUSIONS Small differences in the fulfillment of clinical dose-volume constraints are seen between utilizing deep-learning predicted structures as input to a dose prediction network and manual structures. Overall mean differences <2% indicate that the dose-prediction pipeline is useful as a decision support tool where differences are >2%.

Abstract Purpose This study evaluates deep learning (DL) based dose prediction methods in head and neck cancer (HNC) patients using two types of input contours. Materials and methods Seventy‐five HNC patients undergoing two‐step volumetric‐modulated arc therapy were included. Dose prediction was performed using the AIVOT prototype (AiRato.Inc, Sendai, Japan), a commercial software with an HD U‐net‐based dose distribution prediction system. Models were developed for the initial plan (46 Gy/23Fr) and boost plan (24 Gy/12Fr), trained with 65 cases and tested with 10 cases. The 8‐channel model used one target (PTV) and seven organs at risk (OARs), while the 10‐channel model added two dummy contours (PTV ring and spinal cord PRV). Predicted and deliverable doses, obtained through dose mimicking on another radiation treatment planning system, were evaluated using dose‐volume indices for PTV and OARs. Results For the initial plan, both models achieved approximately 2% prediction accuracy for the target dose and maintained accuracy within 3.2 Gy for OARs. The 10‐channel model outperformed the 8‐channel model for certain dose indices. For the boost plan, both models exhibited prediction accuracies of approximately 2% for the target dose and 1 Gy for OARs. The 10‐channel model showed significantly closer predictions to the ground truth for D50% and Dmean. Deliverable plans based on prediction doses showed little significant difference compared to the ground truth, especially for the boost plan. Conclusion DL‐based dose prediction using the AIVOT prototype software in HNC patients yielded promising results. While additional contours may enhance prediction accuracy, their impact on dose mimicking is relatively small.

For acute ischemic stroke (AIS) patients with large vessel occlusions, clinicians must decide if the benefit of mechanical thrombectomy (MTB) outweighs the risks and potential complications following an invasive procedure. Pre-treatment computed tomography (CT) and angiography (CTA) are widely used to characterize occlusions in the brain vasculature. If a patient is deemed eligible, a modified treatment in cerebral ischemia (mTICI) score will be used to grade how well blood flow is reestablished throughout and following the MTB procedure. An estimation of the likelihood of successful recanalization can support treatment decision-making. In this study, we proposed a fully automated prediction of a patient's recanalization score using pre-treatment CT and CTA imaging. We designed a spatial cross attention network (SCANet) that utilizes vision transformers to localize to pertinent slices and brain regions. Our top model achieved an average cross-validated ROC-AUC of 77.33 $\pm$ 3.9\%. This is a promising result that supports future applications of deep learning on CT and CTA for the identification of eligible AIS patients for MTB.

The recent advances in deep learning (DL) have been accelerated by access to large-scale data and compute. These large-scale resources have been used to train progressively larger models which are resource intensive in terms of compute, data, energy, and carbon emissions. These costs are becoming a new type of entry barrier to researchers and practitioners with limited access to resources at such scale, particularly in the Global South. In this work, we take a comprehensive look at the landscape of existing DL models for medical image analysis tasks and demonstrate their usefulness in settings where resources are limited. To account for the resource consumption of DL models, we introduce a novel measure to estimate the performance per resource unit, which we call the PePR score. Using a diverse family of 131 unique DL architectures (spanning 1M to 130M trainable parameters) and three medical image datasets, we capture trends about the performance-resource trade-offs. In applications like medical image analysis, we argue that small-scale, specialized models are better than striving for large-scale models. Furthermore, we show that using existing pretrained models that are fine-tuned on new data can significantly reduce the computational resources and data required compared to training models from scratch. We hope this work will encourage the community to focus on improving AI equity by developing methods and models with smaller resource footprints.

In this paper, we introduce ChainerRL, an open-source deep reinforcement learning (DRL) library built using Python and the Chainer deep learning framework. ChainerRL implements a comprehensive set of DRL algorithms and techniques drawn from state-of-the-art research in the field. To foster reproducible research, and for instructional purposes, ChainerRL provides scripts that closely replicate the original papers' experimental settings and reproduce published benchmark results for several algorithms. Lastly, ChainerRL offers a visualization tool that enables the qualitative inspection of trained agents. The ChainerRL source code can be found on GitHub: https://github.com/chainer/chainerrl.

The selection of beam orientations, which is a key step in radiation treatment planning, is particularly challenging for non-coplanar radiotherapy systems due to the large number of candidate beams. In this paper, we report progress on the group sparsity approach to beam orientation optimization, wherein beam angles are selected by solving a large scale fluence map optimization problem with an additional group sparsity penalty term that encourages most candidate beams to be inactive. The optimization problem is solved using an accelerated proximal gradient method, the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA). We derive a closed-form expression for a relevant proximal operator which enables the application of FISTA. The proposed algorithm is used to create non-coplanar treatment plans for four cases (including head and neck, lung, and prostate cases), and the resulting plans are compared with clinical plans. The dosimetric quality of the group sparsity treatment plans is superior to that of the clinical plans. Moreover, the runtime for the group sparsity approach is typically about 5 minutes. Problems of this size could not be handled using the previous group sparsity method for beam orientation optimization, which was slow to solve much smaller coplanar cases. This work demonstrates for the first time that the group sparsity approach, when combined with an accelerated proximal gradient method such as FISTA, works effectively for non-coplanar cases with 500-800 candidate beams.

In the modern healthcare system, rapidly expanding costs/complexity, the growing myriad of treatment options, and exploding information streams that often do not effectively reach the front lines hinder the ability to choose optimal treatment decisions over time. The goal in this paper is to develop a general purpose (non-disease-specific) computational/artificial intelligence (AI) framework to address these challenges. This serves two potential functions: 1) a simulation environment for exploring various healthcare policies, payment methodologies, etc., and 2) the basis for clinical artificial intelligence - an AI that can think like a doctor. This approach combines Markov decision processes and dynamic decision networks to learn from clinical data and develop complex plans via simulation of alternative sequential decision paths while capturing the sometimes conflicting, sometimes synergistic interactions of various components in the healthcare system. It can operate in partially observable environments (in the case of missing observations or data) by maintaining belief states about patient health status and functions as an online agent that plans and re-plans. This framework was evaluated using real patient data from an electronic health record. Such an AI framework easily outperforms the current treatment-as-usual (TAU) case-rate/fee-for-service models of healthcare (Cost per Unit Change: $189 vs. $497) while obtaining a 30-35% increase in patient outcomes. Tweaking certain model parameters further enhances this advantage, obtaining roughly 50% more improvement for roughly half the costs. Given careful design and problem formulation, an AI simulation framework can approximate optimal decisions even in complex and uncertain environments. Future work is described that outlines potential lines of research and integration of machine learning algorithms for personalized medicine.

Recently two approximate Newton methods were proposed for the optimisation of Markov Decision Processes. While these methods were shown to have desirable properties, such as a guarantee that the preconditioner is negative-semidefinite when the policy is $\log$-concave with respect to the policy parameters, and were demonstrated to have strong empirical performance in challenging domains, such as the game of Tetris, no convergence analysis was provided. The purpose of this paper is to provide such an analysis. We start by providing a detailed analysis of the Hessian of a Markov Decision Process, which is formed of a negative-semidefinite component, a positive-semidefinite component and a remainder term. The first part of our analysis details how the negative-semidefinite and positive-semidefinite components relate to each other, and how these two terms contribute to the Hessian. The next part of our analysis shows that under certain conditions, relating to the richness of the policy class, the remainder term in the Hessian vanishes in the vicinity of a local optimum. Finally, we bound the behaviour of this remainder term in terms of the mixing time of the Markov chain induced by the policy parameters, where this part of the analysis is applicable over the entire parameter space. Given this analysis of the Hessian we then provide our local convergence analysis of the approximate Newton framework.

In this work we present a multi-armed bandit framework for online expert selection in Markov decision processes and demonstrate its use in high-dimensional settings. Our method takes a set of candidate expert policies and switches between them to rapidly identify the best performing expert using a variant of the classical upper confidence bound algorithm, thus ensuring low regret in the overall performance of the system. This is useful in applications where several expert policies may be available, and one needs to be selected at run-time for the underlying environment.

In this paper knowledge based planning has been revolutionized via a novel mathematical model which converts three dimensional dose distribution (3D3) prediction to a clinical utilizable IMRT treatment plan. Presented model has benefited from both prescribed dose as well as predicted dose and its objective function includes both quadratic and linear phrases, so it was called QuadLin model. The model has been run on the data of 30 patients with head and neck cancer randomly selected from the Open KBP dataset. For each patient, there are 19 sets of dose prediction data in this database. Therefore, a total of 570 problems have been solved in the CVX framework and the results have been evaluated by two plan quality approaches: 1- DVH points differences, and 2- satisfied clinical criteria. The results of the current study indicate a strong significant improvement in clinical indicators compared to the reference plan of the dataset, 3D3 predictions, as well as the results of previous researches. Accordingly, on average for 570 problems and total ROIs, clinical indicators have improved by more than 21% and 15% compared to the predicted dose and previous research, respectively.

In this work, we present a novel application of an uncertainty-quantification framework called Deep Evidential Learning in the domain of radiotherapy dose prediction. Using medical images of the Open Knowledge-Based Planning Challenge dataset, we found that this model can be effectively harnessed to yield uncertainty estimates that inherited correlations with prediction errors upon completion of network training. This was achieved only after reformulating the original loss function for a stable implementation. We found that (i)epistemic uncertainty was highly correlated with prediction errors, with various association indices comparable or stronger than those for Monte-Carlo Dropout and Deep Ensemble methods, (ii)the median error varied with uncertainty threshold much more linearly for epistemic uncertainty in Deep Evidential Learning relative to these other two conventional frameworks, indicative of a more uniformly calibrated sensitivity to model errors, (iii)relative to epistemic uncertainty, aleatoric uncertainty demonstrated a more significant shift in its distribution in response to Gaussian noise added to CT intensity, compatible with its interpretation as reflecting data noise. Collectively, our results suggest that Deep Evidential Learning is a promising approach that can endow deep-learning models in radiotherapy dose prediction with statistical robustness. Towards enhancing its clinical relevance, we demonstrate how we can use such a model to construct the predicted Dose-Volume-Histograms' confidence intervals.