Mel-DEPTHS: a benchmark dataset for epidermis and tumor segmentation for melanoma staging.

Accurate delineation of epidermis and tumor boundaries is central to melanoma staging, yet pixel-level annotation on whole-slide images (WSIs) is labor-intensive and inconsistent across observers. Advancing this field requires standardized, publicly available benchmarks with expert-validated labels. We introduce Mel-DEPTHS, a new benchmark dataset for epidermis and tumor segmentation, designed to accelerate and standardize research for automated melanoma staging. Mel-DEPTHS comprises 50 anonymized melanoma WSIs (40x, 0.25[Formula: see text]m/pixel) with pixel-level masks for epidermis and tumor regions. Clinical variables such as invasion depth, ulceration, and pT stage are provided alongside fixed train/test partitions to ensure reproducibility. To mitigate annotation burden, we developed an Expert-Supervised Iterative Self-Training (ESIST) protocol: a pretrained model generates pseudo-labels, which dermatopathologists iteratively refine for retraining. We benchmarked six state-of-the-art segmentation models (UNet, UNet++, UNet3+, UPerNet, TransUNet, ConvUNeXt) using WSI-level precision, recall, IoU, and Dice. TransUNet achieved the best performance, closely followed by ConvUNeXt and UperNet. Three-fold cross-validation also confirmed consistent model rankings and label robustness. Mel-DEPTHS provides the fidelity and diversity necessary for clinically meaningful segmentation. It establishes a standardized benchmark and fosters reproducibility in computational pathology.