A clinician-written, training-free rule for skin-lesion borders in smartphone photographs: a pre-registered sealed test against a blind tracer, a trained network, SAM and an informed null

v3 (28 Oct 2026): corrects the network’s training set from ISIC 2018 to the ISIC 2017 training set (2,000 images) in the abstract, Methods/Comparators and reference 1. No change to results or numbers.

v2 (21 Oct 2026): adds files missing from v1, with no change to the manuscript, code, data or results: the pre-registration of the sealed confirmatory test (frozen 14 Oct 2026, 20:33 BRT), Addendum D naming the reference tracer (10:40 BRT) and the predictions (12:21–12:22 BRT; annex 20:30–20:32 BRT, added to the same document). Thresholds of the rule are withheld (patent applications). SHA-256 of each file in HASHES_v2.txt.

Author’s page and other studies: https://fabiohumaniza-cyber.github.io/ · Project: https://deolhonapele.com.br

Purpose. Primary-care teams usually first image a skin lesion with a smartphone, yet segmentation methods are validated on dermoscopy. We asked whether a written, training-free rule, adapted from a dermoscopy rule in one day of pre-specified development, delineates lesions in smartphone photographs better than an informed null.

Methods. The engine was frozen by hash after five one-rule rounds on 243 previously exposed PAD-UFES-20 photographs. A sealed set of 300 unused photographs (60 each of actinic keratosis, basal cell carcinoma, nevus, squamous cell carcinoma, seborrhoeic keratosis; one per patient) was traced by an independent blind nurse and run once, sealed, through the engine and four comparators: the previous rule, a U-Net trained on ISIC 2017, SAM ViT-B prompted at the centre, and tuned Otsu. A centred circle of the tracer’s median lesion area was the informed null. Refusals scored zero; hypotheses were tested in a fixed sequence.

Results. The tracer delineated 188 photographs. Median Dice was 0.745 (95% CI 0.663–0.785) for the engine, 0.784 for the U-Net, 0.693 for SAM and 0.671 for the null. The primary hypothesis, engine better than null, failed (paired mean difference −0.052; Wilcoxon p = 0.84). Post hoc, the engine exceeded the null on accepted photographs (median +0.059), but its 27 refusals cost more. On nevi and seborrhoeic keratoses it was close to the U-Net; on pink lesions it did not exceed the null.

Conclusion. A first-cycle written rule is competitive on pigmented lesions in smartphone photographs but not yet on pink ones; centred clinical datasets need an informed null.

Deposit: per-photograph results, test-set draw, sealing and analysis scripts, and all hashes. The frozen engine is available for audit by hash under a confidentiality agreement.