endoExplain: A reproducible protocol for auditing score-localisation discordance in colonoscopy image classifiers

Abstract

Background and objective: A high classifier score and a plausible class-activation map (CAM) are often presented together, although neither establishes that the other is reliable. We introduce endoExplain as a reproducible protocol for auditing score-localisation discordance rather than as a new detector or explanation algorithm. Methods: Content hashing separated HyperKvasir development images from 1,000 masked images before training. EfficientNet-B0, ResNet-34 and ConvNeXt-Tiny were trained with three seeds each. Scores were temperature scaled using validation data only. Grad-CAM, Grad-CAM++, XGrad-CAM, HiResCAM and Eigen-CAM were evaluated on identical image-mask pairs, alongside random and centre baselines. Outcomes combined peak localisation, overlap, a top-20% deletion response, score-threshold sensitivity and adjustment for lesion size and centrality. The selected checkpoint was transferred without retraining to three external mask cohorts. Results: Temperature scaling reduced test expected calibration error from 0.0167 to 0.0115. Among 172 reserved images with scaled score at least 0.90, peak-outside-lesion rates ranged from 4.1% to 62.2% across CAMs. Method dependence remained evident across architectures and seeds, although method rankings were not universal. Spatial alignment and deletion response were not interchangeable. A random-deletion control also produced positive logit drops, limiting specificity claims based on deletion alone. External positive-mask results were dataset dependent. A source-category audit also exposed that 149/155 test positives were dyed-lifted polyps, materially bounding classifier claims. Conclusions: endoExplain makes calibration, spatial agreement, perturbation response and transfer separately inspectable. The results caution against using a score or visually persuasive CAM as evidence of lesion localisation or model reasoning.