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

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Accurate delineation of epidermis and tumor boundaries is crucial for melanoma staging.
• Pixel-level annotation on whole-slide images (WSIs) is time-consuming and prone to inter-observer variability.
• Standardized, publicly available benchmarks with expert-validated labels are needed to advance automated melanoma staging research.

Purpose of the Study:

• Introduce Mel-DEPTHS, a novel benchmark dataset for epidermis and tumor segmentation in melanoma WSIs.
• Facilitate and standardize research in automated melanoma staging.
• Provide expert-validated labels and fixed data partitions for reproducible research.

Main Methods:

• Developed the Mel-DEPTHS dataset with 50 anonymized melanoma WSIs and pixel-level masks for epidermis and tumor regions.
• Implemented an Expert-Supervised Iterative Self-Training (ESIST) protocol to reduce annotation burden.
• Benchmarked six state-of-the-art segmentation models (UNet, UNet++, UNet3+, UPerNet, TransUNet, ConvUNeXt) using WSI-level metrics (precision, recall, IoU, Dice).

Main Results:

• TransUNet demonstrated the highest segmentation performance, followed closely by ConvUNeXt and UperNet.
• Three-fold cross-validation confirmed consistent model rankings and robustness of the labels.
• The Mel-DEPTHS dataset offers the necessary fidelity and diversity for clinically relevant segmentation.

Conclusions:

• Mel-DEPTHS establishes a standardized benchmark for computational pathology in melanoma research.
• The ESIST protocol offers an efficient method for generating expert-validated annotations.
• This work promotes reproducibility and accelerates the development of automated melanoma staging tools.