BGT-Net: Boundary-Graph Transformer Network for Segmentation-Guided Seven-Class Dermoscopic Classification

Abstract

Accurate multi-class classification of dermoscopic skin lesions is still difficult because of irregular lesion morphology, fine differences between classes, and extreme class imbalance in clinical datasets. In this regard, this work presents a novel segmentation-guided hybrid approach called Boundary-Graph Transformer Network (BGT-Net). It improves the classification of seven classes of dermoscopic skin lesions (AKIEC, BCC, BKL, DF, MEL, NV, and VASC) by jointly modeling lesion geometry and internal texture patterns. The proposed approach begins with the use of high-quality manual segmentation to separate the lesion area and obtain accurate boundary information. A graph attention module is used to model geometric irregularity, boundary curvature, and asymmetry patterns based on boundary nodes. Meanwhile, a lightweight transformer branch is used to learn intra-lesion texture patterns based on segmented region patches. Cross-attention fusion is used to combine geometric and texture features for classification. Experimental results on a seven-class dermoscopic skin lesion dataset show that BGT-Net improves classification accuracy to 97.8%, macro F1 score of 0.96, and an AUC of 0.99, which is 12.4% higher than EfficientNet-B4 in accuracy and 0.15 higher in macro-F1. Moreover, the recall of minority classes like DF and VASC is substantially improved and more robust to class imbalance. Thus, the combination of boundary-aware graph modeling and transformer-based texture modeling improves dermoscopic skin lesion classification.