YOLOv10-based detection of melanocytic nevi: reverse exclusion optimization for melanoma screening

Abstract

Malignant melanoma is the deadliest skin cancer, yet its early dermoscopic presentation closely mimics benign melanocytic nevi. Conventional visual or dermoscopic screening therefore suffers from high miss rates and generates excessive biopsies. In this study we focus on Chinese East-Asian patients and introduce a reversed-exclusion strategy—classifying “benign first, exclude malignancy”: lesions that fully meet benign nevus criteria are deemed low-risk; all others are flagged as high-risk. Building on the real-time detector YOLOv10, we incorporate three medical-oriented upgrades: (i) a PP-LCNet backbone to preserve sub-3 mm textures; (ii) a Multiscale Contextual Attention (MCA) neck to enhance cross-scale aggregation; and (iii) a Shape-IoU loss that jointly optimises position, scale, and curvature. The model was trained on a multi-centre dermoscopic dataset from three tertiary hospitals in mainland China (2,040 benign nevi) and independently tested on 365 biopsy-proven melanomas collected at the same medical institution but drawn from a demographically distinct patient cohort, achieving a detection mAP@0.5 of 97.69% for benign lesions and a melanoma false-negative rate (FNR) of only 0.27%. By delivering high-confidence benign identification followed by malignant exclusion, the proposed model offers a high-precision, low-risk pathway for early melanoma screening in Chinese clinical settings. It can markedly reduce unnecessary biopsies while keeping the miss rate below the clinical safety ceiling of 0.5%, thus preserving the life-saving window afforded by early detection