Calibrated selective referral for skin lesion AI on HAM10000

Skin-lesion artificial intelligence is often evaluated using aggregate accuracy or AUROC, but triage-support use also requires calibrated probabilities, explicit referral behaviour, subgroup reliability, and robustness checks. We evaluated a seven-class HAM10000 dermoscopy workflow using lesion-level splitting, validation-only model selection, temperature scaling, selective referral, subgroup analysis, bootstrap uncertainty, stress testing, and ISIC2018 Task 3 related-domain testing. The validation-selected model was an image-only ResNet50 trained with inverse-frequency weighted cross-entropy. On the internal test set it achieved accuracy 0.745, balanced accuracy 0.673, macro-F1 0.625, and macro one-vs-rest AUROC 0.941. Temperature scaling improved probability quality without changing top-1 predictions, reducing Brier score from 0.362 to 0.337, negative log-likelihood from 0.731 to 0.636, and 15-bin expected calibration error from 0.111 to 0.034. At a validation-selected entropy threshold targeting selective risk no higher than 0.150, held-out test coverage was 0.774, selective risk was 0.171, and 77 of 219 melanoma cases were referred. ISIC2018 Task 3 related-domain testing on 1,512 matched images showed accuracy 0.731, macro-F1 0.639, and macro AUROC 0.933. This is a retrospective calibrated triage-support benchmark, not autonomous diagnosis or prospective external validation.