Similarity-Aware Evaluation, Probability Calibration, and Independent Replication for Reliable Gastrointestinal Image Classification

Abstract

High accuracy alone does not establish reliable medical-image classification because visually similar images may cross partitions and confidence may be miscalibrated. We evaluated a similarity-aware framework on a four-class gastrointestinal benchmark (4000 images) and independently replicated the methodology on GastroVision (8000 images; 27 classes). A 64-bit DCT perceptual hash defined groups at Hamming distance ≤ 4. Five-fold random and group-constrained protocols were matched by model and seed, with separate validation, calibration, and test subsets. We compared modern backbones, temperature scaling, a 2 × 2 Mixup/label-smoothing ablation, and deterministic maximum-softmax probability. On the primary benchmark, grouped EfficientNet-B0 achieved accuracy and macro-F1 of 0.9765 ± 0.0022. Grouping changed accuracy by +0.0025 relative to random splitting (Wilcoxon p = 0.625). Temperature scaling reduced expected calibration error from 0.1292 ± 0.0216 to 0.0123 ± 0.0030 without changing predictions, and deterministic MSP achieved error-detection AUROC of 0.9288 ± 0.0181. On GastroVision, random folds contained 78–101 detected cross-subset overlaps, whereas grouped folds contained none. Grouped EfficientNet-B0 achieved 0.8195 ± 0.0043 accuracy and 0.5792 ± 0.0148 macro-F1; ECE decreased from 0.1114 ± 0.0148 to 0.0323 ± 0.0048 after scaling. Similarity grouping removed detected resemblance without reducing accuracy, while calibration and uncertainty analyses exposed properties hidden by discrimination metrics. GastroVision supports methodological portability rather than direct external validation; pHash cannot establish patient or procedure independence, for which identifiers remain preferable.