Metric-guided Image Reconstruction Bounds via Conformal Prediction.

Recent advancements in machine learning have led to novel imaging systems and algorithms that address ill-posed problems. Assessing their trustworthiness and understanding how to deploy them safely at test time remains an important and open problem. We propose a method that leverages conformal prediction to retrieve upper/lower bounds and statistical inliers/outliers of reconstructions based on the prediction intervals of downstream metrics. We apply our method to sparse-view CT for downstream radiotherapy planning and show 1) that metric-guided bounds have valid coverage for downstream metrics while conventional pixel-wise bounds do not and 2) anatomical differences of upper/lower bounds between metric-guided and pixel-wise methods. Our work paves the way for more meaningful reconstruction bounds. Code available at https://github.com/matthewyccheung/conformal-metric.

Evaluating automatically generated normal tissue contours for safe use in head and neck and cervical cancer treatment planning.

PURPOSE: Volumetric-modulated arc therapy (VMAT) is a widely accepted treatment method for head and neck (HN) and cervical cancers; however, creating contours and plan optimization for VMAT plans is a time-consuming process. Our group has created an automated treatment planning tool, the Radiation Planning Assistant (RPA), that uses deep learning models to generate organs at risk (OARs), planning structures and automates plan optimization. This study quantitatively evaluates the quality of contours generated by the RPA tool. METHODS: For patients with HN (54) and cervical (39) cancers, we retrospectively generated autoplans using the RPA. Autoplans were generated using deep-learning and RapidPlan models developed in-house. The autoplans were, then, applied to the original, physician-drawn contours, which were used as a ground truth (GT) to compare with the autocontours (RPA). Using a "two one-sided tests" (TOST) procedure, we evaluated whether the autocontour normal tissue dose was equivalent to that of the ground truth by a margin, δ, that we determined based on clinical judgement. We also calculated the number of plans that met established clinically accepted dosimetric criteria. RESULTS: For HN plans, 91.8% and 91.7% of structures met dosimetric criteria for automatic and manual contours, respectively; for cervical plans, 95.6% and 95.7% of structures met dosimetric criteria for automatic and manual contours, respectively. Autocontours were equivalent to the ground truth for 71% and 75% of common DVH metrics for the HN and cervix, respectively. CONCLUSIONS: This study shows that dosimetrically equivalent normal tissue contours can be created for HN and cervical cancers using deep learning techniques. In general, differences between the contours did not affect the passing or failing of clinical dose tolerances.

Artificial Intelligence-Based Radiotherapy Contouring and Planning to Improve Global Access to Cancer Care.

PURPOSE: Increased automation has been identified as one approach to improving global cancer care. The Radiation Planning Assistant (RPA) is a web-based tool offering automated radiotherapy (RT) contouring and planning to low-resource clinics. In this study, the RPA workflow and clinical acceptability were assessed by physicians around the world. METHODS: The RPA output for 75 cases was reviewed by at least three physicians; 31 radiation oncologists at 16 institutions in six countries on five continents reviewed RPA contours and plans for clinical acceptability using a 5-point Likert scale. RESULTS: For cervical cancer, RPA plans using bony landmarks were scored as usable as-is in 81% (with minor edits 93%); using soft tissue contours, plans were scored as usable as-is in 79% (with minor edits 96%). For postmastectomy breast cancer, RPA plans were scored as usable as-is in 44% (with minor edits 91%). For whole-brain treatment, RPA plans were scored as usable as-is in 67% (with minor edits 99%). For head/neck cancer, the normal tissue autocontours were acceptable as-is in 89% (with minor edits 97%). The clinical target volumes (CTVs) were acceptable as-is in 40% (with minor edits 93%). The volumetric-modulated arc therapy (VMAT) plans were acceptable as-is in 87% (with minor edits 96%). For cervical cancer, the normal tissue autocontours were acceptable as-is in 92% (with minor edits 99%). The CTVs for cervical cancer were scored as acceptable as-is in 83% (with minor edits 92%). The VMAT plans for cervical cancer were acceptable as-is in 99% (with minor edits 100%). CONCLUSION: The RPA, a web-based tool designed to improve access to high-quality RT in low-resource settings, has high rates of clinical acceptability by practicing clinicians around the world. It has significant potential for successful implementation in low-resource clinics.

Automatic end-to-end VMAT treatment planning for rectal cancers.

BACKGROUND: The treatment planning process from segmentation to producing a deliverable plan is time-consuming and labor-intensive. Existing solutions automate the segmentation and planning processes individually. The feasibility of combining auto-segmentation and auto-planning for volumetric modulated arc therapy (VMAT) for rectal cancers in an end-to-end process is not clear. PURPOSE: To create and clinically evaluate a complete end-to-end process for auto-segmentation and auto-planning of VMAT for rectal cancer requiring only the gross tumor volume contour and a CT scan as inputs. METHODS: Patient scans and data were retrospectively selected from our institutional records for patients treated for malignant neoplasm of the rectum. We trained, validated, and tested deep learning auto-segmentation models using nnU-Net architecture for clinical target volume (CTV), bowel bag, large bowel, small bowel, total bowel, femurs, bladder, bone marrow, and female and male genitalia. For the CTV, we identified 174 patients with clinically drawn CTVs. We used data for 18 patients for all structures other than the CTV. The structures were contoured under the guidance of and reviewed by a gastrointestinal (GI) radiation oncologist. The predicted results for CTV in 35 patients and organs at risk (OAR) in six patients were scored by the GI radiation oncologist using a five-point Likert scale. For auto-planning, a RapidPlan knowledge-based planning solution was modeled for VMAT delivery with a prescription of 25 Gy in five fractions. The model was trained and tested on 20 and 34 patients, respectively. The resulting plans were scored by two GI radiation oncologists using a five-point Likert scale. Finally, the end-to-end pipeline was evaluated on 16 patients, and the resulting plans were scored by two GI radiation oncologists. RESULTS: In 31 of 35 patients, CTV contours were clinically acceptable without necessary modifications. The CTV achieved a Dice similarity coefficient of 0.85 (±0.05) and 95% Hausdorff distance of 15.25 (±5.59) mm. All OAR contours were clinically acceptable without edits, except for large and small bowel which were challenging to differentiate. However, contours for total, large, and small bowel were clinically acceptable. The two physicians accepted 100% and 91% of the auto-plans. For the end-to-end pipeline, the two physicians accepted 88% and 62% of the auto-plans. CONCLUSIONS: This study demonstrated that the VMAT treatment planning technique for rectal cancer can be automated to generate clinically acceptable and safe plans with minimal human interventions.

SC-GAN: Structure-completion generative adversarial network for synthetic CT generation from MR images with truncated anatomy.

Creating synthetic CT (sCT) from magnetic resonance (MR) images enables MR-based treatment planning in radiation therapy. However, the MR images used for MR-guided adaptive planning are often truncated in the boundary regions due to the limited field of view and the need for sequence optimization. Consequently, the sCT generated from these truncated MR images lacks complete anatomic information, leading to dose calculation error for MR-based adaptive planning. We propose a novel structure-completion generative adversarial network (SC-GAN) to generate sCT with full anatomic details from the truncated MR images. To enable anatomy compensation, we expand input channels of the CT generator by including a body mask and introduce a truncation loss between sCT and real CT. The body mask for each patient was automatically created from the simulation CT scans and transformed to daily MR images by rigid registration as another input for our SC-GAN in addition to the MR images. The truncation loss was constructed by implementing either an auto-segmentor or an edge detector to penalize the difference in body outlines between sCT and real CT. The experimental results show that our SC-GAN achieved much improved accuracy of sCT generation in both truncated and untruncated regions compared to the original cycleGAN and conditional GAN methods.

Clinical acceptability of automatically generated lymph node levels and structures of deglutition and mastication for head and neck radiation therapy.

Background and Purpose: Auto-contouring of complex anatomy in computed tomography (CT) scans is a highly anticipated solution to many problems in radiotherapy. In this study, artificial intelligence (AI)-based auto-contouring models were clinically validated for lymph node levels and structures of swallowing and chewing in the head and neck. Materials and Methods: CT scans of 145 head and neck radiotherapy patients were retrospectively curated. One cohort (n = 47) was used to analyze seven lymph node levels and the other (n = 98) used to analyze 17 swallowing and chewing structures. Separate nnUnet models were trained and validated using the separate cohorts. For the lymph node levels, preference and clinical acceptability of AI vs human contours were scored. For the swallowing and chewing structures, clinical acceptability was scored. Quantitative analyses of the test sets were performed for AI vs human contours for all structures using overlap and distance metrics. Results: Median Dice Similarity Coefficient ranged from 0.77 to 0.89 for lymph node levels and 0.86 to 0.96 for chewing and swallowing structures. The AI contours were superior to or equally preferred to the manual contours at rates ranging from 75% to 91%; there was not a significant difference in clinical acceptability for nodal levels I-V for manual versus AI contours. Across all AI-generated lymph node level contours, 92% were rated as usable with stylistic to no edits. Of the 340 contours in the chewing and swallowing cohort, 4% required minor edits. Conclusions: An accurate approach was developed to auto-contour lymph node levels and chewing and swallowing structures on CT images for patients with intact nodal anatomy. Only a small portion of test set auto-contours required minor edits.

Random Forest Modeling of Acute Toxicity in Anal Cancer: Effects of Peritoneal Cavity Contouring Approaches on Model Performance.

PURPOSE: Our purpose was to analyze the effect on gastrointestinal (GI) toxicity models when their dose-volume metrics predictors are derived from segmentations of the peritoneal cavity after different contouring approaches. METHODS AND MATERIALS: A random forest machine learning approach was used to predict acute grade ≥3 GI toxicity from dose-volume metrics and clinicopathologic factors for 246 patients (toxicity incidence = 9.5%) treated with definitive chemoradiation for squamous cell carcinoma of the anus. Three types of random forest models were constructed based on different bowel bag segmentation approaches: (1) physician-delineated after Radiation Therapy Oncology Group (RTOG) guidelines, (2) autosegmented by a deep learning model (nnU-Net) following RTOG guidelines, and (3) autosegmented but spanning the entire bowel space. Each model type was evaluated using repeated cross-validation (100 iterations; 50%/50% training/test split). The performance of the models was assessed using area under the precision-recall curve (AUPRC) and the receiver operating characteristic curve (AUROCC), as well as optimal F1 score. RESULTS: When following RTOG guidelines, the models based on the nnU-Net auto segmentations (mean values: AUROCC, 0.71 ± 0.07; AUPRC, 0.42 ± 0.09; F1 score, 0.46 ± 0.08) significantly outperformed (P < .001) those based on the physician-delineated contours (mean values: AUROCC, 0.67 ± 0.07; AUPRC, 0.34 ± 0.08; F1 score, 0.36 ± 0.07). When spanning the entire bowel space, the performance of the autosegmentation models improved considerably (mean values: AUROCC, 0.87 ± 0.05; AUPRC, 0.70 ± 0.09; F1 score, 0.68 ± 0.09). CONCLUSIONS: Random forest models were superior at predicting acute grade ≥3 GI toxicity when based on RTOG-defined bowel bag autosegmentations rather than physician-delineated contours. Models based on autosegmentations spanning the entire bowel space show further considerable improvement in model performance. The results of this study should be further validated using an external data set.

Monitoring Variations in the Use of Automated Contouring Software.

PURPOSE: Our purpose was to identify variations in the clinical use of automatically generated contours that could be attributed to software error, off-label use, or automation bias. METHODS AND MATERIALS: For 500 head and neck patients who were contoured by an in-house automated contouring system, Dice similarity coefficient and added path length were calculated between the contours generated by the automated system and the final contours after editing for clinical use. Statistical process control was used and control charts were generated with control limits at 3 standard deviations. Contours that exceeded the thresholds were investigated to determine the cause. Moving mean control plots were then generated to identify dosimetrists who were editing less over time, which could be indicative of automation bias. RESULTS: Major contouring edits were flagged for: 1.0% brain, 3.1% brain stem, 3.5% left cochlea, 2.9% right cochlea, 4.8% esophagus, 4.1% left eye, 4.0% right eye, 2.2% left lens, 4.9% right lens, 2.5% mandible, 11% left optic nerve, 6.1% right optic nerve, 3.8% left parotid, 5.9% right parotid, and 3.0% of spinal cord contours. Identified causes of editing included unexpected patient positioning, deviation from standard clinical practice, and disagreement between dosimetrist preference and automated contouring style. A statistically significant (P < .05) difference was identified between the contour editing practice of dosimetrists, with 1 dosimetrist editing more across all organs at risk. Eighteen percent (27/150) of moving mean control plots created for 5 dosimetrists indicated the amount of contour editing was decreasing over time, possibly corresponding to automation bias. CONCLUSIONS: The developed system was used to detect statistically significant edits caused by software error, unexpected clinical use, and automation bias. The increased ability to detect systematic errors that occur when editing automatically generated contours will improve the safety of the automatic treatment planning workflow.

Quantifying the Effect of 4-Dimensional Computed Tomography-Based Deformable Dose Accumulation on Representing Radiation Damage for Patients with Locally Advanced Non-Small Cell Lung Cancer Treated with Standard-Fractionated Intensity-Modulated Radiation Therapy.

PURPOSE: The aim of this study was to investigate the dosimetric and clinical effects of 4-dimensional computed tomography (4DCT)-based longitudinal dose accumulation in patients with locally advanced non-small cell lung cancer treated with standard-fractionated intensity-modulated radiation therapy (IMRT). METHODS AND MATERIALS: Sixty-seven patients were retrospectively selected from a randomized clinical trial. Their original IMRT plan, planning and verification 4DCTs, and ∼4-month posttreatment follow-up CTs were imported into a commercial treatment planning system. Two deformable image registration algorithms were implemented for dose accumulation, and their accuracies were assessed. The planned and accumulated doses computed using average-intensity images or phase images were compared. At the organ level, mean lung dose and normal-tissue complication probability (NTCP) for grade ≥2 radiation pneumonitis were compared. At the region level, mean dose in lung subsections and the volumetric overlap between isodose intervals were compared. At the voxel level, the accuracy in estimating the delivered dose was compared by evaluating the fit of a dose versus radiographic image density change (IDC) model. The dose-IDC model fit was also compared for subcohorts based on the magnitude of NTCP difference (|ΔNTCP|) between planned and accumulated doses. RESULTS: Deformable image registration accuracy was quantified, and the uncertainty was considered for the voxel-level analysis. Compared with planned doses, accumulated doses on average resulted in <1-Gy lung dose increase and <2% NTCP increase (up to 8.2 Gy and 18.8% for a patient, respectively). Volumetric overlap of isodose intervals between the planned and accumulated dose distributions ranged from 0.01 to 0.93. Voxel-level dose-IDC models demonstrated a fit improvement from planned dose to accumulated dose (pseudo-R2 increased 0.0023) and a further improvement for patients with ≥2% |ΔNTCP| versus for patients with <2% |ΔNTCP|. CONCLUSIONS: With a relatively large cohort, robust image registrations, multilevel metric comparisons, and radiographic image-based evidence, we demonstrated that dose accumulation more accurately represents the delivered dose and can be especially beneficial for patients with greater longitudinal response.

Deep learning-based automatic segmentation of cardiac substructures for lung cancers.

PURPOSE: Accurate and comprehensive segmentation of cardiac substructures is crucial for minimizing the risk of radiation-induced heart disease in lung cancer radiotherapy. We sought to develop and validate deep learning-based auto-segmentation models for cardiac substructures. MATERIALS AND METHODS: Nineteen cardiac substructures (whole heart, 4 heart chambers, 6 great vessels, 4 valves, and 4 coronary arteries) in 100 patients treated for non-small cell lung cancer were manually delineated by two radiation oncologists. The valves and coronary arteries were delineated as planning risk volumes. An nnU-Net auto-segmentation model was trained, validated, and tested on this dataset with a split ratio of 75:5:20. The auto-segmented contours were evaluated by comparing them with manually drawn contours in terms of Dice similarity coefficient (DSC) and dose metrics extracted from clinical plans. An independent dataset of 42 patients was used for subjective evaluation of the auto-segmentation model by 4 physicians. RESULTS: The average DSCs were 0.95 (+/- 0.01) for the whole heart, 0.91 (+/- 0.02) for 4 chambers, 0.86 (+/- 0.09) for 6 great vessels, 0.81 (+/- 0.09) for 4 valves, and 0.60 (+/- 0.14) for 4 coronary arteries. The average absolute errors in mean/max doses to all substructures were 1.04 (+/- 1.99) Gy and 2.20 (+/- 4.37) Gy. The subjective evaluation revealed that 94% of the auto-segmented contours were clinically acceptable. CONCLUSION: We demonstrated the effectiveness of our nnU-Net model for delineating cardiac substructures, including coronary arteries. Our results indicate that this model has promise for studies regarding radiation dose to cardiac substructures.

Fully-automated, CT-only GTV contouring for palliative head and neck radiotherapy.

Planning for palliative radiotherapy is performed without the advantage of MR or PET imaging in many clinics. Here, we investigated CT-only GTV delineation for palliative treatment of head and neck cancer. Two multi-institutional datasets of palliative-intent treatment plans were retrospectively acquired: a set of 102 non-contrast-enhanced CTs and a set of 96 contrast-enhanced CTs. The nnU-Net auto-segmentation network was chosen for its strength in medical image segmentation, and five approaches separately trained: (1) heuristic-cropped, non-contrast images with a single GTV channel, (2) cropping around a manually-placed point in the tumor center for non-contrast images with a single GTV channel, (3) contrast-enhanced images with a single GTV channel, (4) contrast-enhanced images with separate primary and nodal GTV channels, and (5) contrast-enhanced images along with synthetic MR images with separate primary and nodal GTV channels. Median Dice similarity coefficient ranged from 0.6 to 0.7, surface Dice from 0.30 to 0.56, and 95th Hausdorff distance from 14.7 to 19.7 mm across the five approaches. Only surface Dice exhibited statistically-significant difference across these five approaches using a two-tailed Wilcoxon Rank-Sum test (p ≤ 0.05). Our CT-only results met or exceeded published values for head and neck GTV autocontouring using multi-modality images. However, significant edits would be necessary before clinical use in palliative radiotherapy.

ARCHERY: a prospective observational study of artificial intelligence-based radiotherapy treatment planning for cervical, head and neck and prostate cancer - study protocol.

INTRODUCTION: Fifty per cent of patients with cancer require radiotherapy during their disease course, however, only 10%-40% of patients in low-income and middle-income countries (LMICs) have access to it. A shortfall in specialised workforce has been identified as the most significant barrier to expanding radiotherapy capacity. Artificial intelligence (AI)-based software has been developed to automate both the delineation of anatomical target structures and the definition of the position, size and shape of the radiation beams. Proposed advantages include improved treatment accuracy, as well as a reduction in the time (from weeks to minutes) and human resources needed to deliver radiotherapy. METHODS: ARCHERY is a non-randomised prospective study to evaluate the quality and economic impact of AI-based automated radiotherapy treatment planning for cervical, head and neck, and prostate cancers, which are endemic in LMICs, and for which radiotherapy is the primary curative treatment modality. The sample size of 990 patients (330 for each cancer type) has been calculated based on an estimated 95% treatment plan acceptability rate. Time and cost savings will be analysed as secondary outcome measures using the time-driven activity-based costing model. The 48-month study will take place in six public sector cancer hospitals in India (n=2), Jordan (n=1), Malaysia (n=1) and South Africa (n=2) to support implementation of the software in LMICs. ETHICS AND DISSEMINATION: The study has received ethical approval from University College London (UCL) and each of the six study sites. If the study objectives are met, the AI-based software will be offered as a not-for-profit web service to public sector state hospitals in LMICs to support expansion of high quality radiotherapy capacity, improving access to and affordability of this key modality of cancer cure and control. Public and policy engagement plans will involve patients as key partners.

Synthetic Megavoltage Cone Beam Computed Tomography Image Generation for Improved Contouring Accuracy of Cardiac Pacemakers.

In this study, we aimed to enhance the contouring accuracy of cardiac pacemakers by improving their visualization using deep learning models to predict MV CBCT images based on kV CT or CBCT images. Ten pacemakers and four thorax phantoms were included, creating a total of 35 combinations. Each combination was imaged on a Varian Halcyon (kV/MV CBCT images) and Siemens SOMATOM CT scanner (kV CT images). Two generative adversarial network (GAN)-based models, cycleGAN and conditional GAN (cGAN), were trained to generate synthetic MV (sMV) CBCT images from kV CT/CBCT images using twenty-eight datasets (80%). The pacemakers in the sMV CBCT images and original MV CBCT images were manually delineated and reviewed by three users. The Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and mean surface distance (MSD) were used to compare contour accuracy. Visual inspection showed the improved visualization of pacemakers on sMV CBCT images compared to original kV CT/CBCT images. Moreover, cGAN demonstrated superior performance in enhancing pacemaker visualization compared to cycleGAN. The mean DSC, HD95, and MSD for contours on sMV CBCT images generated from kV CT/CBCT images were 0.91 ± 0.02/0.92 ± 0.01, 1.38 ± 0.31 mm/1.18 ± 0.20 mm, and 0.42 ± 0.07 mm/0.36 ± 0.06 mm using the cGAN model. Deep learning-based methods, specifically cycleGAN and cGAN, can effectively enhance the visualization of pacemakers in thorax kV CT/CBCT images, therefore improving the contouring precision of these devices.

Radiation Planning Assistant - A Web-based Tool to Support High-quality Radiotherapy in Clinics with Limited Resources.

Access to radiotherapy worldwide is limited. The Radiation Planning Assistant (RPA) is a fully automated, web-based tool that is being developed to offer fully automated radiotherapy treatment planning tools to clinics with limited resources. The goal is to help clinical teams scale their efforts, thus reaching more patients with cancer. The user connects to the RPA via a webpage, completes a Service Request (prescription and information about the radiotherapy targets), and uploads the patient's CT image set. The RPA offers two approaches to automated planning. In one-step planning, the system uses the Service Request and CT scan to automatically generate the necessary contours and treatment plan. In two-step planning, the user reviews and edits the automatically generated contours before the RPA continues to generate a volume-modulated arc therapy plan. The final plan is downloaded from the RPA website and imported into the user's local treatment planning system, where the dose is recalculated for the locally commissioned linac; if necessary, the plan is edited prior to approval for clinical use.

Validation of an automated contouring and treatment planning tool for pediatric craniospinal radiation therapy.

Purpose: Treatment planning for craniospinal irradiation (CSI) is complex and time-consuming, especially for resource-constrained centers. To alleviate demanding workflows, we successfully automated the pediatric CSI planning pipeline in previous work. In this work, we validated our CSI autosegmentation and autoplanning tool on a large dataset from St. Jude Children's Research Hospital. Methods: Sixty-three CSI patient CT scans were involved in the study. Pre-planning scripts were used to automatically verify anatomical compatibility with the autoplanning tool. The autoplanning pipeline generated 15 contours and a composite CSI treatment plan for each of the compatible test patients (n=51). Plan quality was evaluated quantitatively with target coverage and dose to normal tissue metrics and qualitatively with physician review, using a 5-point Likert scale. Three pediatric radiation oncologists from 3 institutions reviewed and scored 15 contours and a corresponding composite CSI plan for the final 51 test patients. One patient was scored by 3 physicians, resulting in 53 plans scored total. Results: The algorithm automatically detected 12 incompatible patients due to insufficient junction spacing or head tilt and removed them from the study. Of the 795 autosegmented contours reviewed, 97% were scored as clinically acceptable, with 92% requiring no edits. Of the 53 plans scored, all 51 brain dose distributions were scored as clinically acceptable. For the spine dose distributions, 92%, 100%, and 68% of single, extended, and multiple-field cases, respectively, were scored as clinically acceptable. In all cases (major or minor edits), the physicians noted that they would rather edit the autoplan than create a new plan. Conclusions: We successfully validated an autoplanning pipeline on 51 patients from another institution, indicating that our algorithm is robust in its adjustment to differing patient populations. We automatically generated 15 contours and a comprehensive CSI treatment plan for each patient without physician intervention, indicating the potential for increased treatment planning efficiency and global access to high-quality radiation therapy.

A real-time contouring feedback tool for consensus-based contour training.

Purpose: Variability in contouring structures of interest for radiotherapy continues to be challenging. Although training can reduce such variability, having radiation oncologists provide feedback can be impractical. We developed a contour training tool to provide real-time feedback to trainees, thereby reducing variability in contouring. Methods: We developed a novel metric termed localized signed square distance (LSSD) to provide feedback to the trainee on how their contour compares with a reference contour, which is generated real-time by combining trainee contour and multiple expert radiation oncologist contours. Nine trainees performed contour training by using six randomly assigned training cases that included one test case of the heart and left ventricle (LV). The test case was repeated 30 days later to assess retention. The distribution of LSSD maps of the initial contour for the training cases was combined and compared with the distribution of LSSD maps of the final contours for all training cases. The difference in standard deviations from the initial to final LSSD maps, ΔLSSD, was computed both on a per-case basis and for the entire group. Results: For every training case, statistically significant ΔLSSD were observed for both the heart and LV. When all initial and final LSSD maps were aggregated for the training cases, before training, the mean LSSD ([range], standard deviation) was -0.8 mm ([-37.9, 34.9], 4.2) and 0.3 mm ([-25.1, 32.7], 4.8) for heart and LV, respectively. These were reduced to -0.1 mm ([-16.2, 7.3], 0.8) and 0.1 mm ([-6.6, 8.3], 0.7) for the final LSSD maps during the contour training sessions. For the retention case, the initial and final LSSD maps of the retention case were aggregated and were -1.5 mm ([-22.9, 19.9], 3.4) and -0.2 mm ([-4.5, 1.5], 0.7) for the heart and 1.8 mm ([-16.7, 34.5], 5.1) and 0.2 mm ([-3.9, 1.6],0.7) for the LV. Conclusions: A tool that uses real-time contouring feedback was developed and successfully used for contour training of nine trainees. In all cases, the utility was able to guide the trainee and ultimately reduce the variability of the trainee's contouring.

Identifying the optimal deep learning architecture and parameters for automatic beam aperture definition in 3D radiotherapy.

PURPOSE: Two-dimensional radiotherapy is often used to treat cervical cancer in low- and middle-income countries, but treatment planning can be challenging and time-consuming. Neural networks offer the potential to greatly decrease planning time through automation, but the impact of the wide range of hyperparameters to be set during training on model accuracy has not been exhaustively investigated. In the current study, we evaluated the effect of several convolutional neural network architectures and hyperparameters on 2D radiotherapy treatment field delineation. METHODS: Six commonly used deep learning architectures were trained to delineate four-field box apertures on digitally reconstructed radiographs for cervical cancer radiotherapy. A comprehensive search of optimal hyperparameters for all models was conducted by varying the initial learning rate, image normalization methods, and (when appropriate) convolutional kernel size, the number of learnable parameters via network depth and the number of feature maps per convolution, and nonlinear activation functions. This yielded over 1700 unique models, which were all trained until performance converged and then tested on a separate dataset. RESULTS: Of all hyperparameters, the choice of initial learning rate was most consistently significant for improved performance on the test set, with all top-performing models using learning rates of 0.0001. The optimal image normalization was not consistent across architectures. High overlap (mean Dice similarity coefficient = 0.98) and surface distance agreement (mean surface distance < 2 mm) were achieved between the treatment field apertures for all architectures using the identified best hyperparameters. Overlap Dice similarity coefficient (DSC) and distance metrics (mean surface distance and Hausdorff distance) indicated that DeepLabv3+ and D-LinkNet architectures were least sensitive to initial hyperparameter selection. CONCLUSION: DeepLabv3+ and D-LinkNet are most robust to initial hyperparameter selection. Learning rate, nonlinear activation function, and kernel size are also important hyperparameters for improving performance.