Abstract / Summary
Abstract We adapted two explainability methods commonly used in image classification, KernelSHAP and Grad-CAM, to colonoscopy video classification.We evaluated their plausibility by measuring how well their approximation of EndoFMLV’s visual focus for polyp class predictions aligned with the visible polyp region. Both methods showed low plausibility, and plausibility alone could not reveal model faults reliably.
Topics
Colorectal Cancer Screening and DetectionOpen access
Primary Source
Current Directions in Biomedical Engineering