AutoContour Model RADAC V3

AutoContour Model RADAC V3 is a software device utilizing machine learning algorithms to automatically segment anatomical structures on CT and MR medical images for radiation therapy treatment planning. Its methodology involves deep learning models trained on diverse datasets, with performance validated through Dice Similarity Coefficients and qualitative clinical reviews on independent external datasets, demonstrating its utility in assisting treatment planners.

AutoContour Model RADAC V3 is a software system designed to assist radiation treatment planners by automatically contouring anatomical structures within CT and MR medical images, thereby streamlining preparation for radiation therapy. The system operates by processing DICOM-compliant image data through machine learning algorithms, specifically deep-learning based models. These models are trained on extensive imaging datasets covering various anatomical regions including the head and neck, thorax, abdomen, and pelvis for adult patients. The software comprises a .NET client application for image loading, review, and editing of contours; a local agent service for monitoring new datasets; and a cloud-based service for performing the automatic contouring. This version, RADAC V3, expands upon its predecessor by incorporating 90 new CT models and 11 new MR models for structure segmentation, aiming to improve accuracy and utility in clinical settings. The performance of AutoContour Model RADAC V3 was evaluated through non-clinical testing using independent datasets, distinct from training data, to ensure model generalizability. Validation focused on Dice Similarity Coefficient (DSC) scores and qualitative clinical appropriateness assessed by independent experts. For CT models, training on an average of 373 sets yielded mean DSCs of 0.88 (large structures), 0.88 (medium), and 0.75 (small); external validation on The Cancer Imaging Archive (TCIA) datasets resulted in mean DSCs of 0.90, 0.83, and 0.79 respectively, with an average qualitative clinical rating of 4.5 (out of 5), indicating minor edits needed. MR models, trained on datasets like TCIA GLIS-RT and Prostate-MRI-US-Biopsy, showed training DSCs of 0.87 (medium) and 0.74 (small); external validation on unique datasets produced similar DSCs and an average qualitative rating of 4.4, confirming their clinical utility in assisting with the contouring process.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the performance of AutoContour RADAC V3, particularly the Dice Similarity Coefficient (DSC) and qualitative review scores, vary across different imaging scanner manufacturers and acquisition protocols not explicitly detailed in the training/testing sets?; What is the observed impact on clinical workflow efficiency (e.g., time saved in contouring) when using AutoContour RADAC V3 compared to manual contouring or previous versions, across different anatomical regions and user experience levels?; Given the demographic data of the training sets (e.g., 84% White for CT), what steps are being taken or planned to ensure the robustness and accuracy of the AI models across more diverse patient populations and rarer anatomical variations?; How do the CT structures generated via post-processing/boolean operations from previously released models (marked 'N/A*' for CNN training) perform in scenarios where the primary models (e.g., Eye, Rib) exhibit lower confidence or accuracy for a specific patient?; What are the specific failure modes or types of inaccuracies most commonly observed in the AI-generated contours for both CT and MR models, and what is the typical user intervention required to correct them according to the qualitative reviews?