Track: Clinical Studies

Abstract:

The development of effective, clinically relevant AI tools requires robust validation and close partnership with pathologists. This abstract details the 3-year co-development and validation of a novel AI-based diagnostic assistance tool in a dermatopathology laboratory designed for full workflow integration. The AI solution improves efficiency through intelligent case sorting, significant time savings, and seamless implementation within laboratory systems, including automated reporting. Its capabilities include the classification of skin lesions (melanoma, basal cell carcinoma, squamous cell carcinoma, naevi), automated measurement of lesion size and surgical margins, detection of perineural invasion, and the proposal of thumbnails for mitosis detection in melanoma. The project has included more than 9000 slides, 30 pathologists and three clinical studies. The methodology involved pathologists across all steps from data annotation to iterative testing. The iterative development model was proven critical for refining the tool's sensitivity (up to 91.6%) and ensuring its outputs were clinically intuitive and relevant. The studies demonstrated high diagnostic accuracy for lesion classification (between 96,1% to 98.7% depending on the type of lesion) and a strong concordance between AI-generated measurements and pathologist assessments. Feedback from participating pathologists affirmed the tool's utility in improving diagnostic support and workflow efficiency. Future work will, among other things, expand the tool's scope with finer-grained pathology classifications. In conclusion, this comprehensive validation confirms that a deeply collaborative, pathologist-centric development process is paramount for creating effective and trusted AI tools that can be successfully integrated into the clinical dermatopathology workflow.

Menu

Sophie MAZELLIER