AutoContour

AutoContour is a software system designed to streamline radiation therapy planning by providing automated, machine learning-driven contouring of anatomical structures on CT images, with subsequent review and editing capabilities through a web interface. Its methodology integrates local data monitoring, cloud-based algorithmic processing for initial contour generation, and a user-centric platform for refinement and DICOM RTSTRUCT export, aiming to enhance efficiency and consistency in the pre-treatment phase.

AutoContour, from Radformation, Inc., is a software tool aiding radiation treatment planners by automating and facilitating the contouring and review of anatomical structures in medical images for radiation therapy planning. Its architecture consists of three key parts: an 'agent' service on Windows OS that detects and uploads new CT datasets; a cloud-based service employing machine learning for automatic initial contour generation; and a web application for image registration, user review, contour editing, and exporting the final structure set as a DICOM RTSTRUCT file. Scientifically, AutoContour applies machine learning algorithms to CT image data to identify and delineate anatomical structures, a process central to its automated contouring. While primarily focused on CT for this automated task, it allows manual registration with MR and PET scans to incorporate multi-modal information for guiding contour refinement. The system's efficacy and safety were established through nonclinical performance data, including validation of its machine learning models on independent datasets and rigorous software verification, demonstrating its capability to perform as intended in preparing images for radiation therapy planning.

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. What specific machine learning architectures or models are employed by the cloud-based automatic contouring service, and what was the nature and diversity of the datasets used for training and validation to ensure its generalizability across different patient anatomies and image qualities?; How does AutoContour's performance in terms of contouring accuracy (e.g., Dice similarity coefficient, Hausdorff distance) for various anatomical structures compare quantitatively to manual contouring by experienced clinicians or to its predicate devices, particularly for complex or low-contrast regions?; What are the typical processing times for automatic contouring of a standard CT dataset, from upload by the agent service to the availability of initial contours in the web application, and how does this scale with dataset size or complexity?; What specific tools and functionalities are available within the web application for users to review and edit the automatically generated contours, and how does the system handle versioning or audit trails for these modifications?; Given that AutoContour primarily uses CT for automated contouring but allows MR/PET for registration, what are the limitations or best practices when using MR/PET data to guide the refinement of CT-based contours, especially concerning potential registration inaccuracies?