AutoContour Model RADAC V2

AutoContour RADAC V2 is a software device that employs deep learning algorithms to automatically generate anatomical contours on CT and MR medical images, thereby assisting radiation treatment planners. Its methodology involves processing DICOM images via a client application connected to local or cloud-based services, supporting user review and modification, with its efficacy validated through quantitative performance metrics like Dice Similarity Coefficient, sensitivity, and specificity on diverse datasets.

AutoContour RADAC V2 is a software tool intended to assist radiation treatment planners by automatically contouring and enabling review of anatomical structures within medical images, in preparation for radiation therapy treatment planning. It processes DICOM-compliant CT or MR image data using machine learning-based contouring, with its deep-learning models trained on extensive imaging datasets covering anatomical organs of the head and neck, thorax, abdomen, and pelvis for adult male and female patients. The system architecture includes a .NET client application for users to load images, perform registration, review and edit contours, and export structure sets; a local agent service to monitor for new datasets for automatic contouring; and a cloud-based or local automatic contouring service that generates the initial contours based on the uploaded image sets. This version, AutoContour RADAC V2, builds upon its predecessor by expanding machine-learning based contouring and manual ROI manipulation to Magnetic Resonance (MR) image types, increasing the number of supported CT structure models (including 58 new CT models), and enhancing image registration capabilities to include automatic deformable image registration for transferring structure contours between image sets. The device's performance was validated non-clinically using metrics such as the Dice Similarity Coefficient (DSC). For CT structures, mean DSC results were 0.94 for large volume, 0.82 for medium volume, and 0.61 for small volume structures. For the six MR models, a mean DSC of 0.67 was achieved. Sensitivity and specificity analyses were also conducted on independent test datasets, with performance found to be consistent with state-of-the-art literature and a reference device. The system mandates a user review and approval process for all generated contours prior to their export to a treatment planning system.

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 V2's machine learning models vary across different patient demographics (e.g., age, ethnicity, body habitus) not extensively detailed in the provided summary statistics?; What specific deep learning architectures are employed for the CT and MR contouring models, and what were the key considerations during their training and optimization?; Are there plans to expand the MR model library beyond the current six structures, particularly for other common radiotherapy sites, and what is the anticipated validation process for new MR models?; How does the automatic deformable image registration feature account for potential anatomical changes or inconsistencies between image sets, and what are its limitations?; Given the reported specificity for CT models (ranging from 0 to 100%), what are the most common types of false positive contours generated, and how effectively does the user-specified structure template system mitigate these in diverse clinical scenarios?