Abstract

Automated skin lesion detection is an essential first step for dermatological screening, yet most work targets lesion classification rather than localisation. Total body photography (TBP) sharpens this gap, producing hundreds of images per patient that each require reliable localisation at scale. We present, to our knowledge, the first systematic multi-architecture benchmark for lesion detection in TBP imagery. Rather than proposing a new detector, we examine design choices that shape detector performance, providing insights to guide their development. Using the iToBoS dataset, we evaluate seven architectures across fourteen configurations: four YOLO variants, two transformer setprediction detectors (MS-DETR, RT-DETRv2-S), and a distributionrefinement detector (D-FINE-S). Rather than an mAP leaderboard, we characterise reliability through error decomposition (TIDE), AP across IoU thresholds, a confidence × NMS-IoU sweep, and EigenCAM saliency. YOLOv12s with full augmentation gives the best F1 (0.640), while DFINE-S leads on mAP50 (0.674) and precision (0.707) through probabilistic coordinate refinement, with the smallest dependence on augmentation (+0.17 pp F1 versus +0.8–1.9 pp for the others), and EigenCAM shows its activations are lesion-focused where YOLOv12s exhibits signs of the Clever Hans effect; we therefore recommend D-FINE-S as the primary model. The sweep exposes four distinct confidence regimes (a YOLO plateau, MS-DETR threshold-invariance, and calibrated conf=0.50 operation for RT-DETRv2-S and D-FINE-S), and TIDE reveals complementary failure modes with distinct clinical implications.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ISIC_016.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=B3zd9oBiwC&nesting=2&sort=date-desc

BibTex

@InProceedings{YamMuh_Automated_MICCAISAT2026,
        author = { Yamin, Muhammad AND Murali, Praveen Kumar AND Das, Aritra AND Rajani, Hayat AND Saha, Anup AND Wahid, Sheikh Abdul AND Nwafor, Solomon Chibuzo AND Nazari, Sana AND Nuria, Ferrera AND Serra-García, Laura AND Lenoir, Clément AND Gracias, Nuno AND Malvehy, Josep AND Garcia, Rafael},
        title = { { Automated Skin Lesion Detection in Total Body Photography: A Multi-Architecture Benchmark } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17271},
        month = {pending},
        page = {pending}
}


back to top