Contour ProtégéAI

Contour ProtégéAI is a medical imaging software accessory that employs machine-learning based algorithms for the automated contouring of anatomical structures in CT and MR images to assist medical professionals in applications like quantitative analysis and radiation therapy planning. Its neural network models, trained on extensive, diverse datasets, demonstrate performance comparable or superior to traditional atlas-based segmentation methods, as measured by metrics like the Dice coefficient, though physician review of generated contours remains essential.

Contour ProtégéAI is an accessory software for MIM software, designed to be used by trained medical professionals for the automated processing of CT and MR digital medical images. It utilizes machine-learning algorithms, specifically neural network models, to automatically create contours of anatomical structures. These generated contours support applications such as quantitative analysis, aiding adaptive therapy, transferring contours to radiation therapy treatment planning systems, and archiving for patient follow-up. The system's 3.0.0 CT models were trained using 4061 CT images from 31 international clinical sites, while MR capabilities focus on segmenting normal structures of the prostate, seminal vesicles, and urethra within T2-weighted images. The software operates on various computer systems and can be deployed via MIMcloud or locally, but it is not intended for detecting or contouring lesions. The performance of Contour ProtégéAI was evaluated by comparing its new 3.0.0 CT neural network models against atlases created using the MIM Maestro reference device, with testing conducted on an independent set of 739 images. The Dice coefficient was used as the primary metric for comparison, with results indicating that Contour ProtégéAI demonstrated equivalent or superior performance to the MIM Maestro atlas segmentation across numerous anatomical structures. Equivalence was defined by the lower 95th percentile confidence bound of the Contour ProtégéAI segmentation being no more than 0.1 Dice lower than the mean MIM atlas performance. It is mandatory for users to employ appropriate image visualization software to review and, if necessary, edit the automatically generated contours to ensure accuracy for clinical use.

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 the Contour ProtégéAI machine learning models vary with different CT/MR image acquisition parameters (e.g., slice thickness, contrast, scanner vendor) not explicitly detailed in the training exclusion criteria?; What is the typical time required for clinicians to review and edit contours generated by Contour ProtégéAI for complex anatomical regions, and how does this compare to fully manual contouring?; Considering the training data excluded non-axial, thin-slice (<0.5mm), or non-Fan Beam/mV acquisitions, what is the expected performance degradation if the software is used on images with these characteristics?; How does the 'leave-one-out' analysis for the MIM Maestro atlas comparison account for potential institutional biases in segmentation practices if the test patient is from the same institution as the atlas data, even if the patient itself is excluded?; For structures with lower Dice coefficients (e.g., Bowel_Large, Colon_Sigmoid), what are the clinical implications of using these automated contours, and what specific guidance is provided for their review and correction?