EndoExplain framework audits AI colonoscopy with spatial, temporal, and quality checks

Current AI-assisted colonoscopy systems typically output frame-level confidence scores, but confidence alone fails to reveal whether a prediction is spatially plausible, temporally consistent, or reliable under degraded image quality. To address this, researchers introduce EndoExplain, a lightweight and reproducible audit framework that layers multiple diagnostic signals: classification confidence, lesion segmentation, CAM-style visual attribution, attribution-mask alignment, frame-quality indicators, and temporal event summarization. This approach separates conflated signals in computer-aided detection pipelines, enabling clinicians to review AI outputs in a structured, cockpit-style interface. The framework was tested on the HyperKvasir dataset, with an EfficientNet-B0 classifier reaching 0.9280 test accuracy over ten endoscopic classes and 0.9969 ROC-AUC for polyp family detection. The segmentation model (U-Net++ with EfficientNet-B1 backbone) achieved Dice 0.9318 and IoU 0.8826.

A strict top-20% multi-method attribution audit revealed that the choice of attribution method significantly changes explanation-mask agreement: Eigen-CAM produced the strongest overlap, while Grad-CAM++ remained weakly aligned. This ranking was preserved in a frozen-model external sanity check on ETIS-LaribPolypDB and CVC-ClinicDB, though overlap was dataset-dependent. The temporal component was validated via a human-reviewed clip-level benchmark over 60 HyperKvasir videos, achieving an event F1 score of 0.8081 at a pre-specified threshold of 0.85 using one-to-one event matching with at least 1 second overlap. A clinician-informed external plausibility review confirmed the clinical readability of separating confidence, localisation, attribution, quality metadata, and temporal context. The authors emphasize this is a retrospective research prototype, not a clinically validated medical device.