DeepContour (V1.0)

DeepContour (V1.0) is a medical image processing software that utilizes deep learning algorithms to automatically segment anatomical structures from CT scans for radiation therapy planning. Its methodology involves training models on diverse CT datasets, employing a two-step segmentation process, and validating performance through quantitative metrics like Dice coefficients and ASSD against established benchmarks.

DeepContour (V1.0) is a deep learning-based medical imaging software designed for trained healthcare professionals to automatically process CT images. Its primary function is to generate anatomical contours using deep learning algorithms, supporting tasks such as quantitative analysis, organ Hounsfield Unit (HU) distribution statistics, and the creation of patient management archives. The software processes CT input data and produces RTSTRUCT outputs, which can be transferred to Treatment Planning Systems (TPS). It also features an interactive contouring application for editing and reviewing automatically generated contours, supports 3D reconstruction, visual tools based on these contours, and performs rigid and elastic registration. DeepContour is intended for adult use only, specifically for patients previously selected for radiation therapy, covering anatomical regions such as the head & neck, thorax, abdomen, and pelvis. The software's deep learning models were trained on 800 CT images from 35 hospitals in China, covering various ages and body regions (head and neck, chest, abdomen, pelvic), with data augmentation used to enhance generalization. The training involved a two-step auto-segmentation process (region distinction then model-specific segmentation), with initial segmentations reviewed by radiation oncologists. Performance was verified using 203 CT images, including 100 clinical datasets from Peking Union Medical College Hospital and 103 images from American public datasets (LCTSC and Pancreas-CT). DeepContour demonstrated comparable or superior performance to predicate devices (AI-Rad CAI-Rad Companion Organs RT and Contour ProtégéAI) based on Dice coefficients and Average Symmetric Surface Distance (ASSD) for numerous organs at risk (OARs) across different anatomical regions.

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 DeepContour's performance vary across CT scanners from different manufacturers (Philips, GE, Siemens) beyond what was noted in the training data, particularly for less common or older models?; What are the specific failure modes or types of errors most commonly observed in the automated contouring, and what mechanisms are in place for users to efficiently identify and correct them?; The training data was primarily from Chinese hospitals; how was the model's generalizability to diverse global populations (beyond the American public datasets) ensured or tested, especially concerning variations in anatomy or image acquisition protocols?; What is the typical processing time for a standard CT dataset, and how does it scale with the complexity of the case or the number of organs being contoured?; Are there specific anatomical structures or pathological conditions where DeepContour's automated segmentation accuracy is known to be significantly lower, requiring more intensive manual review and correction?