Contour ProtégéAI

Contour ProtégéAI is a medical image processing software accessory that employs machine-learning neural networks to automatically segment anatomical structures in CT and MR images, aiding in radiation therapy planning and quantitative analysis. Its methodology involves training deep learning models on diverse clinical datasets and validating their performance against established segmentation techniques and user efficiency metrics to ensure accurate and time-saving contour generation for medical professionals.

Contour ProtégéAI, an accessory to MIM software, utilizes machine-learning algorithms to automatically generate anatomical contours on CT and MR medical images. Designed for use by trained medical professionals, its primary purpose is to assist in the automated processing of these images for applications such as quantitative analysis, aiding adaptive therapy, transferring contours to radiation therapy treatment planning systems, and archiving for patient management; the software supports DICOM 3.0 and can segment a variety of anatomical structures across different CT locations, as well as the prostate, seminal vesicles, and urethra in T2-weighted MR images, while it is crucial to note that Contour ProtégéAI is not intended for lesion detection, and all automatically generated results require review and potential editing using appropriate image visualization software. The system's machine-learning capabilities are based on neural network models, including the updated 4.1.0 models for Head and Neck CT, Thorax CT, and Whole Body CT, which were trained using 550 CT images from 41 global clinical sites with detailed re-segmentation by medical experts to ensure high-quality training data; performance evaluation was conducted on an independent set of 754 images, comparing Contour ProtégéAI to atlas-based segmentation from the MIM Maestro system using key metrics like Dice scores, Mean Distance to Agreement (MDA), cumulative Added Path Length (APL), and user-evaluated time savings, with results demonstrating that Contour ProtégéAI performed equivalently or superiorly to the reference atlas-based method across numerous anatomical structures, showing improvements in segmentation accuracy and efficiency, though known limitations exist related to image quality and specific patient demographics.

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 4.1.0 neural network models vary across different CT scanner manufacturers and imaging protocols not extensively represented in the training dataset (e.g., Toshiba which was only 2.7% of training data)?; Given the limited representation of patients under 40 in the training data, what specific validation studies have been or could be conducted to ensure robust performance and safety for pediatric or younger adult populations?; What are the specific anatomical structures or scenarios where the cumulative Added Path Length (APL) for Contour ProtégéAI, despite being non-inferior overall, still indicates a significant need for manual editing, and how does this compare to the predicate for those specific cases?; How does the system's performance degrade with increasing levels of image noise or motion artifacts, and are there built-in quality checks or warnings for suboptimal image inputs?; Beyond Dice and MDA, what is the impact of using Contour ProtégéAI on inter-observer variability in contouring compared to purely manual methods or atlas-based methods, particularly for complex structures like brachial plexus or lymph node regions?