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.

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.

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.

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.

Deep learning-based dose prediction to improve the plan quality of volumetric modulated arc therapy for gynecologic cancers.

BACKGROUND: In recent years, deep-learning models have been used to predict entire three-dimensional dose distributions. However, the usability of dose predictions to improve plan quality should be further investigated. PURPOSE: To develop a deep-learning model to predict high-quality dose distributions for volumetric modulated arc therapy (VMAT) plans for patients with gynecologic cancer and to evaluate their usability in driving plan quality improvements. METHODS: A total of 79 VMAT plans for the female pelvis were used to train (47 plans), validate (16 plans), and test (16 plans) 3D dense dilated U-Net models to predict 3D dose distributions. The models received the normalized CT scan, dose prescription, and target and normal tissue contours as inputs. Three models were used to predict the dose distributions for plans in the test set. A radiation oncologist specializing in the treatment of gynecologic cancers scored the test set predictions using a 5-point scale (5, acceptable as-is; 4, prefer minor edits; 3, minor edits needed; 2, major edits needed; and 1, unacceptable). The clinical plans for which the dose predictions indicated that improvements could be made were reoptimized with constraints extracted from the predictions. RESULTS: The predicted dose distributions in the test set were of comparable quality to the clinical plans. The mean voxel-wise dose difference was -0.14 ± 0.46 Gy. The percentage dose differences in the predicted target metrics of D 1 % ${D}_{1{\mathrm{\% }}}$ and D 98 % ${D}_{98{\mathrm{\% }}}$ were -1.05% ± 0.59% and 0.21% ± 0.28%, respectively. The dose differences in the predicted organ at risk mean and maximum doses were -0.30 ± 1.66 Gy and -0.42 ± 2.07 Gy, respectively. A radiation oncologist deemed all of the predicted dose distributions clinically acceptable; 12 received a score of 5, and four received a score of 4. Replanning of flagged plans (five plans) showed that the original plans could be further optimized to give dose distributions close to the predicted dose distributions. CONCLUSIONS: Deep-learning dose prediction can be used to predict high-quality and clinically acceptable dose distributions for VMAT female pelvis plans, which can then be used to identify plans that can be improved with additional optimization.

A deep-learning-based dose verification tool utilizing fluence maps for a cobalt-60 compensator-based intensity-modulated radiation therapy system.

Background and purpose: A novel cobalt-60 compensator-based intensity-modulated radiation therapy (IMRT) system was developed for a resource-limited environment but lacked an efficient dose verification algorithm. The aim of this study was to develop a deep-learning-based dose verification algorithm for accurate and rapid dose predictions. Materials and methods: A deep-learning network was employed to predict the doses from static fields related to beam commissioning. Inputs were a cube-shaped phantom, a beam binary mask, and an intersecting volume of the phantom and beam binary mask, while output was a 3-dimensional (3D) dose. The same network was extended to predict patient-specific doses for head and neck cancers using two different approaches. A field-based method predicted doses for each field and combined all calculated doses into a plan, while the plan-based method combined all nine fluences into a plan to predict doses. Inputs included patient computed tomography (CT) scans, binary beam masks, and fluence maps truncated to the patient's CT in 3D. Results: For static fields, predictions agreed well with ground truths with average deviations of less than 0.5% for percent depth doses and profiles. Even though the field-based method showed excellent prediction performance for each field, the plan-based method showed better agreement between clinical and predicted dose distributions. The distributed dose deviations for all planned target volumes and organs at risk were within 1.3 Gy. The calculation speed for each case was within two seconds. Conclusions: A deep-learning-based dose verification tool can accurately and rapidly predict doses for a novel cobalt-60 compensator-based IMRT system.

Commissioning and dosimetric validation of a novel compensator-based Co-60 IMRT system for evaluating suitability to automated treatment planning.

PURPOSE: A novel compensator-based system has been proposed which delivers intensity-modulated radiation therapy (IMRT) with cobalt-60 beams. This could improve access to advanced radiotherapy in low- and middle-income countries. For this system to be clinically viable and to be adapted into the Radiation Planning Assistant (RPA), being developed to offer automated planning services in low- and middle-income countries, it is necessary to commission and validate it in a commercial treatment planning system (TPS). METHODS: The novel treatment device considered here employs a cobalt-60 source and nine compensators. Each compensator is produced by 3-D printing a thin plastic mold which is then filled on-demand within the machine with reusable 2-mm-diameter spherical tungsten balls. This system was commissioned in the Eclipse TPS and validation tests were conducted with Monte Carlo using Geant4 Application for Tomographic Emission for percentage depth dose, in-plane profiles, penumbra, and IMRT dose validation. And the American Association of Physicists in Medicine Task Group 119 benchmarking testing was performed. Additionally, compensator-based cobalt-60 IMRT plans were created for 46 head-and-neck cancer cases and compared to the linac-based volumetric modulated arc therapy (VMAT) plans used clinically, then dosimetric parameters were evaluated. Beam-on time for each field was calculated. In addition, the measurement was also performed in a limited environment and compared with the Monte Carlo simulations. RESULTS: The differences in percent depth doses and in-plane profiles between the Eclipse and Monte Carlo simulations were 0.65% ± 0.41% and 1.02% ± 0.99%, respectively, and the 80%-20% penumbra agreed within 0.46 ± 0.27 mm. For the Task Group 119 validation plans, all treatment planning goals were met and gamma passing rates were >95% (3%/3 mm criteria). In 46 clinical head-and-neck cases, the cobalt-60 compensator-based IMRT plans had planning target volume (PTV) coverages similar to linac-based VMAT plans: all dosimetric values for PTV were within 1.5%. The organs at risk dose parameters were somewhat higher in cobalt-60 compensator-based IMRT plans versus linac-based VMAT plans. The mean dose differences for the spinal cord, brain, and brainstem were 4.43 ± 1.92, 3.39 ± 4.67, and 2.40 ± 3.71 Gy, while those for the rest of the organs were <1 Gy. The average beam-on time per field was 0.42 ± 0.10 min for the 6 MV multi-leaf-collimator plans while those for the cobalt-60 compensator plans were 0.17 ± 0.01 and 0.31 ± 0.01 min at the dose rates of 350 and 175 cGy/min. There was a good agreement between in-plane profiles from measurements and Monte Carlo simulations, which differences are 1.34 ± 1.90% and 0.13 ± 2.16% for two different fields. CONCLUSIONS: A novel compensator-based IMRT system using cobalt-60 beams was commissioned and validated in a commercial TPS. Plan quality with this system was comparable to that of linac-based plans in all test cases with shorter estimated beam-on times. This system enables reliable, high-quality plans with reduced cost and complexity and may have benefits for underserved regions of the world. This system is being integrated into the RPA, a web-based platform for auto-contouring and auto-planning.

Automated Contouring and Planning in Radiation Therapy: What Is 'Clinically Acceptable'?

Developers and users of artificial-intelligence-based tools for automatic contouring and treatment planning in radiotherapy are expected to assess clinical acceptability of these tools. However, what is 'clinical acceptability'? Quantitative and qualitative approaches have been used to assess this ill-defined concept, all of which have advantages and disadvantages or limitations. The approach chosen may depend on the goal of the study as well as on available resources. In this paper, we discuss various aspects of 'clinical acceptability' and how they can move us toward a standard for defining clinical acceptability of new autocontouring and planning tools.

Outcomes of Complex Abdominal Wall Reconstruction After Oncologic Resection: 14-Year Experience at an NCI-Designated Cancer Center.

BACKGROUND: Outcomes studies for abdominal wall reconstruction (AWR) in the setting of previous oncologic extirpation are lacking. We sought to evaluate long-term outcomes of AWR using acellular dermal matrix (ADM) after extirpative resection, compare them to primary herniorrhaphy, and report the rates and predictors of postoperative complications. METHODS: We conducted a retrospective cohort study of patients who underwent AWR after oncologic resection from March 2005 to June 2019 at a tertiary cancer center. The primary outcome was hernia recurrence (HR). Secondary outcomes included surgical site occurrences (SSOs), surgical site infection (SSIs), length of hospital stay (LOS), reoperation, and 30-day readmission. RESULTS: Of 720 consecutive patients who underwent AWR during the study period, 194 (26.9%) underwent AWR following resection of abdominal wall tumors. In adjusted analyses, patients who had AWR after extirpative resection were more likely to have longer LOS (ß, 2.57; 95%CI, 1.27 to 3.86, p < 0.001) than those with primary herniorrhaphy, but the risk of HR, SSO, SSI, 30-day readmission, and reoperation did not differ significantly. In the extirpative cohort, obesity (Hazard ratio, 6.48; p = 0.003), and bridged repair (Hazard ratio, 3.50; p = 0.004) were predictors of HR. Radiotherapy (OR, 2.23; p = 0.017) and diabetes mellites (OR, 3.70; p = 0.005) were predictors of SSOs. Defect width (OR, 2.30; p < 0.001) and mesh length (OR, 3.32; p = 0.046) were predictors of SSIs. Concomitant intra-abdominal surgery for active disease was not associated with worse outcomes. CONCLUSIONS: AWR with ADM following extirpative resection demonstrated outcomes comparable with primary herniorrhaphy. Preoperative risk assessment and optimization are imperative for improving outcomes.

Multi-organ segmentation of abdominal structures from non-contrast and contrast enhanced CT images.

Manually delineating upper abdominal organs at risk (OARs) is a time-consuming task. To develop a deep-learning-based tool for accurate and robust auto-segmentation of these OARs, forty pancreatic cancer patients with contrast-enhanced breath-hold computed tomographic (CT) images were selected. We trained a three-dimensional (3D) U-Net ensemble that automatically segments all organ contours concurrently with the self-configuring nnU-Net framework. Our tool's performance was assessed on a held-out test set of 30 patients quantitatively. Five radiation oncologists from three different institutions assessed the performance of the tool using a 5-point Likert scale on an additional 75 randomly selected test patients. The mean (± std. dev.) Dice similarity coefficient values between the automatic segmentation and the ground truth on contrast-enhanced CT images were 0.80 ± 0.08, 0.89 ± 0.05, 0.90 ± 0.06, 0.92 ± 0.03, 0.96 ± 0.01, 0.97 ± 0.01, 0.96 ± 0.01, and 0.96 ± 0.01 for the duodenum, small bowel, large bowel, stomach, liver, spleen, right kidney, and left kidney, respectively. 89.3% (contrast-enhanced) and 85.3% (non-contrast-enhanced) of duodenum contours were scored as a 3 or above, which required only minor edits. More than 90% of the other organs' contours were scored as a 3 or above. Our tool achieved a high level of clinical acceptability with a small training dataset and provides accurate contours for treatment planning.

An Automated Treatment Planning Framework for Spinal Radiation Therapy and Vertebral-Level Second Check.

PURPOSE: Complicating factors such as time pressures, anatomic variants in the spine, and similarities in adjacent vertebrae are associated with incorrect level treatments of the spine. The purpose of this work was to mitigate such challenges by fully automating the treatment planning process for diagnostic and simulation computed tomography (CT) scans. METHODS AND MATERIALS: Vertebral bodies are labeled on CT scans of any length using 2 intendent deep-learning models-mirroring 2 different experts labeling the spine. Then, a U-Net++ architecture was trained, validated, and tested to contour each vertebra (n = 220 CT scans). Features from the CT and auto-contours were input into a random forest classifier to predict whether vertebrae were correctly labeled. This classifier was trained using auto-contours from cone beam computed tomography, positron emission tomography/CT, simulation CT, and diagnostic CT images (n = 56 CT scans, 751 contours). Auto-plans were generated via scripting. Each model was combined into a framework to make a fully automated clinical tool. A retrospective planning study was conducted in which 3 radiation oncologists scored auto-plan quality on an unseen patient cohort (n = 60) on a 5-point scale. CT scans varied in scan length, presence of surgical implants, imaging protocol, and metastatic burden. RESULTS: The results showed that the uniquely designed convolutional neural networks accurately labeled and segmented vertebral bodies C1-L5 regardless of imaging protocol or metastatic burden. Mean dice-similarity coefficient was 85.0% (cervical), 90.3% (thoracic), and 93.7% (lumbar). The random forest classifier predicted mislabeling across various CT scan types with an area under the curve of 0.82. All contouring and labeling errors within treatment regions (11 of 11), including errors from patient plans with atypical anatomy (eg, T13, L6) were detected. Radiation oncologists scored 98% of simulation CT-based plans and 92% of diagnostic CT-based plans as clinically acceptable or needing minor edits for patients with typical anatomy. On average, end-to-end treatment planning time of the clinical tool was less than 8 minutes. CONCLUSIONS: This novel method to automatically verify, contour, and plan palliative spine treatments is efficient and effective across various CT scan types. Furthermore, it is the first to create a clinical tool that can automatically verify vertebral level in CT images.

Clinical acceptability of fully automated external beam radiotherapy for cervical cancer with three different beam delivery techniques.

PURPOSE: To fully automate CT-based cervical cancer radiotherapy by automating contouring and planning for three different treatment techniques. METHODS: We automated three different radiotherapy planning techniques for locally advanced cervical cancer: 2D 4-field-box (4-field-box), 3D conformal radiotherapy (3D-CRT), and volumetric modulated arc therapy (VMAT). These auto-planning algorithms were combined with a previously developed auto-contouring system. To improve the quality of the 4-field-box and 3D-CRT plans, we used an in-house, field-in-field (FIF) automation program. Thirty-five plans were generated for each technique on CT scans from multiple institutions and evaluated by five experienced radiation oncologists from three different countries. Every plan was reviewed by two of the five radiation oncologists and scored using a 5-point Likert scale. RESULTS: Overall, 87%, 99%, and 94% of the automatically generated plans were found to be clinically acceptable without modification for the 4-field-box, 3D-CRT, and VMAT plans, respectively. Some customizations of the FIF configuration were necessary on the basis of radiation oncologist preference. Additionally, in some cases, it was necessary to renormalize the plan after it was generated to satisfy radiation oncologist preference. CONCLUSION: Approximately, 90% of the automatically generated plans were clinically acceptable for all three planning techniques. This fully automated planning system has been implemented into the radiation planning assistant for further testing in resource-constrained radiotherapy departments in low- and middle-income countries.

VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images.

Vertebral labelling and segmentation are two fundamental tasks in an automated spine processing pipeline. Reliable and accurate processing of spine images is expected to benefit clinical decision support systems for diagnosis, surgery planning, and population-based analysis of spine and bone health. However, designing automated algorithms for spine processing is challenging predominantly due to considerable variations in anatomy and acquisition protocols and due to a severe shortage of publicly available data. Addressing these limitations, the Large Scale Vertebrae Segmentation Challenge (VerSe) was organised in conjunction with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) in 2019 and 2020, with a call for algorithms tackling the labelling and segmentation of vertebrae. Two datasets containing a total of 374 multi-detector CT scans from 355 patients were prepared and 4505 vertebrae have individually been annotated at voxel level by a human-machine hybrid algorithm (https://osf.io/nqjyw/, https://osf.io/t98fz/). A total of 25 algorithms were benchmarked on these datasets. In this work, we present the results of this evaluation and further investigate the performance variation at the vertebra level, scan level, and different fields of view. We also evaluate the generalisability of the approaches to an implicit domain shift in data by evaluating the top-performing algorithms of one challenge iteration on data from the other iteration. The principal takeaway from VerSe: the performance of an algorithm in labelling and segmenting a spine scan hinges on its ability to correctly identify vertebrae in cases of rare anatomical variations. The VerSe content and code can be accessed at: https://github.com/anjany/verse.

Clinical implementation of automated treatment planning for whole-brain radiotherapy.

The purpose of the study was to develop and clinically deploy an automated, deep learning-based approach to treatment planning for whole-brain radiotherapy (WBRT). We collected CT images and radiotherapy treatment plans to automate a beam aperture definition from 520 patients who received WBRT. These patients were split into training (n = 312), cross-validation (n = 104), and test (n = 104) sets which were used to train and evaluate a deep learning model. The DeepLabV3+ architecture was trained to automatically define the beam apertures on lateral-opposed fields using digitally reconstructed radiographs (DRRs). For the beam aperture evaluation, 1st quantitative analysis was completed using a test set before clinical deployment and 2nd quantitative analysis was conducted 90 days after clinical deployment. The mean surface distance and the Hausdorff distances were compared in the anterior-inferior edge between the clinically used and the predicted fields. Clinically used plans and deep-learning generated plans were evaluated by various dose-volume histogram metrics of brain, cribriform plate, and lens. The 1st quantitative analysis showed that the average mean surface distance and Hausdorff distance were 7.1 mm (±3.8 mm) and 11.2 mm (±5.2 mm), respectively, in the anterior-inferior edge of the field. The retrospective dosimetric comparison showed that brain dose coverage (D99%, D95%, D1%) of the automatically generated plans was 29.7, 30.3, and 32.5 Gy, respectively, and the average dose of both lenses was up to 19.0% lower when compared to the clinically used plans. Following the clinical deployment, the 2nd quantitative analysis showed that the average mean surface distance and Hausdorff distance between the predicted and clinically used fields were 2.6 mm (±3.2 mm) and 4.5 mm (±5.6 mm), respectively. In conclusion, the automated patient-specific treatment planning solution for WBRT was implemented in our clinic. The predicted fields appeared consistent with clinically used fields and the predicted plans were dosimetrically comparable.

Technical Note: Dose prediction for head and neck radiotherapy using a three-dimensional dense dilated U-net architecture.

PURPOSE: Radiation therapy treatment planning is a time-consuming and iterative manual process. Consequently, plan quality varies greatly between and within institutions. Artificial intelligence shows great promise in improving plan quality and reducing planning times. This technical note describes our participation in the American Association of Physicists in Medicine Open Knowledge-Based Planning Challenge (OpenKBP), a competition to accurately predict radiation therapy dose distributions. METHODS: A three-dimensional (3D) densely connected U-Net with dilated convolutions was developed to predict 3D dose distributions given contoured CT images of head and neck patients as input. While traditional augmentation techniques such as rotations and translations were explored, it was found that training on random patches alone resulted in the greatest model performance. A custom-weighted mean squared error loss function was employed. Finally, an ensemble of best-performing networks was used to generate the final challenge predictions. RESULTS: Our team (SuperPod) placed second in the dose stream of the OpenKBP challenge. The average mean absolute difference between the predicted and clinical dose distributions of the testing dataset was 2.56 Gy. On average, the predicted normalized target DVH metrics were within 3% of the clinical plans, and the predicted organ at risk DVH metrics were within 2 Gy of the clinical plans. CONCLUSIONS: The developed 3D dense dilated U-Net architecture can accurately predict 3D radiotherapy dose distributions and can be used as part of a fully automated radiation therapy planning pipeline.

Generating High-Quality Lymph Node Clinical Target Volumes for Head and Neck Cancer Radiation Therapy Using a Fully Automated Deep Learning-Based Approach.

PURPOSE: To develop a deep learning model that generates consistent, high-quality lymph node clinical target volumes (CTV) contours for head and neck cancer (HNC) patients, as an integral part of a fully automated radiation treatment planning workflow. METHODS AND MATERIALS: Computed tomography (CT) scans from 71 HNC patients were retrospectively collected and split into training (n = 51), cross-validation (n = 10), and test (n = 10) data sets. All had target volume delineations covering lymph node levels Ia through V (Ia-V), Ib through V (Ib-V), II through IV (II-IV), and retropharyngeal (RP) nodes, which were previously approved by a radiation oncologist specializing in HNC. Volumes of interest (VOIs) about nodal levels were automatically identified using computer vision techniques. The VOI (cropped CT image) and approved contours were used to train a U-Net autosegmentation model. Each lymph node level was trained independently, with model parameters optimized by assessing performance on the cross-validation data set. Once optimal model parameters were identified, overlap and distance metrics were calculated between ground truth and autosegmentations on the test set. Lastly, this final model was used on 32 additional patient scans (not included in original 71 cases) and autosegmentations visually rated by 3 radiation oncologists as being "clinically acceptable without requiring edits," "requiring minor edits," or "requiring major edits." RESULTS: When comparing ground truths to autosegmentations on the test data set, median Dice Similarity Coefficients were 0.90, 0.90, 0.89, and 0.81, and median mean surface distance values were 1.0 mm, 1.0 mm, 1.1 mm, and 1.3 mm for node levels Ia-V, Ib-V, II-IV, and RP nodes, respectively. Qualitative scoring varied among physicians. Overall, 99% of autosegmented target volumes were either scored as being clinically acceptable or requiring minor edits (ie, stylistic recommendations, <2 minutes). CONCLUSIONS: We developed a fully automated artificial intelligence approach to autodelineate nodal CTVs for patients with intact HNC. Most autosegmentations were found to be clinically acceptable after qualitative review when considering recommended stylistic edits. This promising work automatically delineates nodal CTVs in a robust and consistent manner; this approach can be implemented in ongoing efforts for fully automated radiation treatment planning.

The Emergence of Artificial Intelligence within Radiation Oncology Treatment Planning.

BACKGROUND: The future of artificial intelligence (AI) heralds unprecedented change for the field of radiation oncology. Commercial vendors and academic institutions have created AI tools for radiation oncology, but such tools have not yet been widely adopted into clinical practice. In addition, numerous discussions have prompted careful thoughts about AI's impact upon the future landscape of radiation oncology: How can we preserve innovation, creativity, and patient safety? When will AI-based tools be widely adopted into the clinic? Will the need for clinical staff be reduced? How will these devices and tools be developed and regulated? SUMMARY: In this work, we examine how deep learning, a rapidly emerging subset of AI, fits into the broader historical context of advancements made in radiation oncology and medical physics. In addition, we examine a representative set of deep learning-based tools that are being made available for use in external beam radiotherapy treatment planning and how these deep learning-based tools and other AI-based tools will impact members of the radiation treatment planning team. Key Messages: Compared to past transformative innovations explored in this article, such as the Monte Carlo method or intensity-modulated radiotherapy, the development and adoption of deep learning-based tools is occurring at faster rates and promises to transform practices of the radiation treatment planning team. However, accessibility to these tools will be determined by each clinic's access to the internet, web-based solutions, or high-performance computing hardware. As seen by the trends exhibited by many technologies, high dependence on new technology can result in harm should the product fail in an unexpected manner, be misused by the operator, or if the mitigation to an expected failure is not adequate. Thus, the need for developers and researchers to rigorously validate deep learning-based tools, for users to understand how to operate tools appropriately, and for professional bodies to develop guidelines for their use and maintenance is essential. Given that members of the radiation treatment planning team perform many tasks that are automatable, the use of deep learning-based tools, in combination with other automated treatment planning tools, may refocus tasks performed by the treatment planning team and may potentially reduce resource-related burdens for clinics with limited resources.

Evaluation of a multiview architecture for automatic vertebral labeling of palliative radiotherapy simulation CT images.

PURPOSE: The purpose of this work was to evaluate the performance of X-Net, a multiview deep learning architecture, to automatically label vertebral levels (S2-C1) in palliative radiotherapy simulation CT scans. METHODS: For each patient CT scan, our automated approach 1) segmented spinal canal using a convolutional-neural network (CNN), 2) formed sagittal and coronal intensity projection pairs, 3) labeled vertebral levels with X-Net, and 4) detected irregular intervertebral spacing using an analytic methodology. The spinal canal CNN was trained via fivefold cross validation using 1,966 simulation CT scans and evaluated on 330 CT scans. After labeling vertebral levels (S2-C1) in 897 palliative radiotherapy simulation CT scans, a volume of interest surrounding the spinal canal in each patient's CT scan was converted into sagittal and coronal intensity projection image pairs. Then, intensity projection image pairs were augmented and used to train X-Net to automatically label vertebral levels using fivefold cross validation (n = 803). Prior to testing upon the final test set (n = 94), CT scans of patients with anatomical abnormalities, surgical implants, or other atypical features from the final test set were placed in an outlier group (n = 20), whereas those without these features were placed in a normative group (n = 74). The performance of X-Net, X-Net Ensemble, and another leading vertebral labeling architecture (Btrfly Net) was evaluated on both groups using identification rate, localization error, and other metrics. The performance of our approach was also evaluated on the MICCAI 2014 test dataset (n = 60). Finally, a method to detect irregular intervertebral spacing was created based on the rate of change in spacing between predicted vertebral body locations and was also evaluated using the final test set. Receiver operating characteristic analysis was used to investigate the performance of the method to detect irregular intervertebral spacing. RESULTS: The spinal canal architecture yielded centroid coordinates spanning S2-C1 with submillimeter accuracy (mean ± standard deviation, 0.399 ± 0.299 mm; n = 330 patients) and was robust in the localization of spinal canal centroid to surgical implants and widespread metastases. Cross-validation testing of X-Net for vertebral labeling revealed that the deep learning model performance (F1 score, precision, and sensitivity) improved with CT scan length. The X-Net, X-Net Ensemble, and Btrfly Net mean identification rates and localization errors were 92.4% and 2.3 mm, 94.2% and 2.2 mm, and 90.5% and 3.4 mm, respectively, in the final test set and 96.7% and 2.2 mm, 96.9% and 2.0 mm, and 94.8% and 3.3 mm, respectively, within the normative group of the final test set. The X-Net Ensemble yielded the highest percentage of patients (94%) having all vertebral bodies identified correctly in the final test set when the three most inferior and superior vertebral bodies were excluded from the CT scan. The method used to detect labeling failures had 67% sensitivity and 95% specificity when combined with the X-Net Ensemble and flagged five of six patients with atypical vertebral counts (additional thoracic (T13), additional lumbar (L6) or only four lumbar vertebrae). Mean identification rate on the MICCAI 2014 dataset using an X-Net Ensemble was increased from 86.8% to 91.3% through the use of transfer learning and obtained state-of-the-art results for various regions of the spine. CONCLUSIONS: We trained X-Net, our unique convolutional neural network, to automatically label vertebral levels from S2 to C1 on palliative radiotherapy CT images and found that an ensemble of X-Net models had high vertebral body identification rate (94.2%) and small localization errors (2.2 ± 1.8 mm). In addition, our transfer learning approach achieved state-of-the-art results on a well-known benchmark dataset with high identification rate (91.3%) and low localization error (3.3 mm ± 2.7 mm). When we pre-screened radiotherapy CT images for the presence of hardware, surgical implants, or other anatomic abnormalities prior to the use of X-Net, it labeled the spine correctly in more than 97% of patients and 94% of patients when scans were not prescreened. Automatically generated labels are robust to widespread vertebral metastases and surgical implants and our method to detect labeling failures based on neighborhood intervertebral spacing can reliably identify patients with an additional lumbar or thoracic vertebral body.