GRP-090 A Novel Superpixel–RAG–Transformer Approach for Three-Class Melanoma Segmentation

Melanoma is one of the deadliest skin cancers, with early detection relying on accurately identifying both the lesion core and its often ambiguous border. Traditional CNN and U-Net models struggle with fuzzy transitions and irregular boundaries. We propose a three-class segmentation framework that labels regions as background, border, or lesion core. Our method over-segments images into superpixels, builds a Region Adjacency Graph (RAG) to capture spatial context, and generates embeddings using transformer-based autoencoders. This approach combines local image statistics with global semantic structure. Experiments on the HAM10000 dataset show improved precision and recall, especially in challenging border regions, outperforming CNN/U-Net baselines. Our results highlight the value of explicitly modeling boundaries for accurate and interpretable melanoma segmentation.