INT Contour

INTContour is a software tool that utilizes machine learning, specifically Convolutional Neural Networks trained on expert-contoured medical images, to provide automatic segmentation of anatomical structures for radiation therapy treatment planning. Its purpose is to offer an efficient initial contouring solution for treatment targets and organs at risk across various body regions, with its performance validated as non-inferior to existing methods, while requiring subsequent review and potential editing by qualified medical professionals.

INTContour is a software-only medical device designed to assist in radiation therapy treatment planning. It employs a machine learning-based approach, specifically utilizing Convolutional Neural Networks (CNNs), to automatically segment anatomical structures in medical images such as CT and MRI scans. The system is trained on a library of expert-contoured cases to identify and delineate treatment targets and organs at risk. INTContour is intended for use by dosimetrists, medical physicists, and radiation oncologists as an initial segmentation tool, with its output provided in DICOM RTSTRUCT format. This output requires subsequent review and potential editing using appropriate external software, as INTContour itself does not include manual editing capabilities. The software operates as a service on a hospital's private network via a web-based interface and supports segmentation for the head and neck, thorax, abdomen, and male pelvis regions. The performance of INTContour was evaluated through non-clinical software verification and validation testing, focusing on model performance and software usability. Testing utilized data from multiple sources, distinct from the training dataset, covering the supported anatomical regions and diverse patient cases (ages 18-76, various cancers), with ground truth established by at least two trained personnel. Accuracy was assessed using the Dice metric for larger organs and the 95% Hausdorff Distance (HD95) for smaller structures. Statistical analysis demonstrated that INTContour's segmentation capabilities are non-inferior to those of a predicate device, meeting predefined acceptance criteria for all supported organs, thereby supporting its role as an initial step in the contouring process that necessitates professional oversight.

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 INTContour's CNN-based segmentation vary across different imaging modalities (CT vs. MRI) and specific imaging protocols for the supported anatomical regions?; What is the typical time-saving observed by clinicians (dosimetrists, medical physicists, radiation oncologists) when using INTContour as an initial segmentation tool compared to fully manual contouring or other semi-automated tools in a clinical setting?; Considering the convolutional neural network architecture, what methods are available or could be implemented to provide users with insights into the confidence levels or uncertainty of the generated contours for specific structures, particularly in atypical cases?; How robust is the INTContour algorithm to variations in image quality, patient anatomical diversity (e.g., post-surgical changes, presence of implants), and different cancer pathologies within the supported regions, beyond the dataset used for non-clinical testing?; What is the detailed performance (Dice metric and 95% Hausdorff Distance) on particularly challenging structures known for high inter-observer variability in manual contouring, and how does this specifically compare to multi-reader expert consensus?