AutoContour (Model RADAC V4)

AutoContour Model RADAC V4 is software that assists radiation therapy treatment planners by automatically contouring anatomical structures within CT and MR medical images using machine learning. It allows users to review and modify the automatically generated contours and export them as DICOM-compliant structure sets for use in radiation therapy treatment planning systems.

AutoContour V4 adds 77 new CT structure models and 18 new MR structure models compared to the predicate V3, expands DICOM export capabilities to include deformable registration and dose files, and supports image registration (rigid and deformable). The underlying CNN architecture, software platform, and core functionality remain unchanged from the predicate device.

The device was validated using Dice Similarity Coefficient (DSC) testing on independent datasets not used for training, with pass criteria of 0.8 for large structures, 0.65 for medium structures, and 0.5 for small structures. Ground truth contours were generated using consensus NRG/RTOG guidelines by three clinically experienced experts (two radiation therapy physicists and one dosimetrist).

AutoContour V4 is substantially equivalent because it uses the same CNN architecture, software components, and operating platform as the predicate V3, with only additive changes. All new structure models passed identical DSC validation criteria as predicate models, the same risk mitigations apply to new models, and external clinical testing demonstrated mean clinical ratings of 4.57 for CT and 4.6 for MR structures (requiring only minor edits for clinical use), consistent with predicate performance.