AI Segmentation

AI Segmentation is a software medical device that employs deep learning and classical algorithms to automate the contouring of organs at risk from CT images, intended to support radiation therapy treatment planning by providing initial segmentations for review and editing by qualified professionals. Its efficacy and safety are supported by non-clinical performance data, including DICE similarity scores and expert qualitative assessments, demonstrating equivalence to its predicate version with enhancements like updated AI models and new editing tools.

AI Segmentation is a web-based software medical device designed to assist trained medical professionals in radiation therapy treatment planning. It utilizes a combination of deep learning and classical algorithms, operating in the cloud, to automatically segment organs at risk from CT images across predefined anatomical regions including the head and neck, thorax, pelvis, and abdomen. The software provides a user interface for visualization, review, and editing of these automated segmentations, which can then be exported to compatible treatment planning systems like the Eclipse Treatment Planning System. It is explicitly stated that AI Segmentation is not intended for making clinical decisions, offering medical advice, or evaluating radiation plans or treatment procedures; its outputs require review, potential editing, and final approval by a qualified physician. This version of AI Segmentation (v2.0) introduces updates and additions to its AI models for automated segmentation and contouring, which are described as static and non-adaptive, meaning they do not learn or change from user input over time. It also incorporates new simple editing tools for users. The performance of these updated and new AI models was evaluated through non-clinical testing, comparing auto-generated contours against ground truth contours using the DICE similarity index, and through qualitative assessment by clinical experts who rated the acceptability of the contours. The results indicated performance equivalent to the predicate device (version 1.0), meeting an 80% acceptability target by experts. No animal or clinical studies were conducted for this version.

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. What is the reported range and average DICE similarity coefficient for each specific anatomical structure across the head/neck, thorax, pelvis, and abdomen regions for the updated/new AI models?; How does the performance of AI Segmentation v2.0 vary with different CT image qualities or scanner protocols, and were these variables considered during non-clinical testing?; What are the specific functionalities of the 'simple editing tools' added in version 2.0, and how do they quantitatively impact the time required for contour review and adjustment by treatment planners?; Given the AI models are static and non-adaptive, what is Varian's plan for incorporating new data or evolving anatomical knowledge to maintain or improve model performance over time?; What were the specific criteria for the 'ground truth contours' used in the DICE similarity index calculation, and how was inter-observer variability among experts establishing these ground truths accounted for?