Contour ProtégéAI

Contour ProtégéAI is a medical image processing software that employs machine-learning based neural network models to automate the delineation of anatomical structures in CT and MR images for applications like radiation therapy planning. The system's 4.0.0 CT models, trained on 326 diverse CT images and tested on 819 independent images, demonstrated equivalent or superior segmentation accuracy, measured by Dice scores, compared to a reference atlas-based method.

Contour ProtégéAI, developed by MIM Software Inc., is an advanced medical imaging software accessory designed to assist trained medical professionals by automatically processing CT and MR images. It employs machine-learning algorithms, specifically neural networks, to create anatomical contours. These contours are critical for various clinical applications, including quantitative analysis, aiding adaptive therapy, transferring data to radiation therapy treatment planning systems, and archiving for patient follow-up. The system supports DICOM 3.0 standard images and is intended for segmenting anatomical structures across diverse CT locations and specific structures like the prostate, seminal vesicles, and urethra in T2-weighted MR images, with the explicit note that users must review and, if necessary, edit the automatically generated results using appropriate visualization software; it is not intended for lesion detection. The latest iteration features new 4.0.0 CT neural network models, which were trained using a dataset of 326 CT images sourced from 37 international clinical sites, ensuring data used for testing was from entirely separate institutions. These images underwent a rigorous re-segmentation process by medical professionals based on established clinical guidelines to ensure high-quality training data. The performance of Contour ProtégéAI was evaluated against a reference device (MIM Maestro atlases built from the same training data) using a test set of 819 independent images, with the Dice similarity coefficient as the primary metric. Results demonstrated that Contour ProtégéAI's automated segmentations were equivalent or superior across numerous anatomical structures in Head and Neck, Thorax, Abdomen, Pelvis, and SurePlan MRT CT models, and preliminary user feedback indicated comparable time-savings to other automated segmentation tools.

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 4.0.0 CT neural network models, as measured by Dice scores, vary when applied to CT images from scanners or acquisition protocols significantly different from those in the 326-image training set?; What is the quantifiable impact on clinical workflow, such as average time saved per case, when using Contour ProtégéAI for segmentation compared to manual contouring or the predicate atlas-based segmentation?; Given that the training data excluded patients from test institutions, what strategies were employed to ensure model generalization across diverse patient demographics and anatomical variations not explicitly represented in the 37 training sites?; The document lists several structures for which Dice scores are provided. Does the system segment additional anatomical structures in CT and MR images beyond those specifically detailed in the performance tables, and if so, what is their expected accuracy?; How does Contour ProtégéAI handle images with significant artifacts or poor image quality, and what is the recommended protocol for user review and correction in such cases to ensure clinical accuracy?