Lesion-aware prototype contrastive learning for port-wine stain classification

Image-based classification of pink-red and purple-red port-wine stain (PWS) subtypes is a fine-grained visual recognition task. However, subtle chromatic differences, blurred lesion boundaries, and heterogeneous acquisition conditions make this task challenging. To address these issues, we developed LAPC, a lesion-aware prototype contrastive framework that combines dual-statistic channel-spatial attention with prototype-discriminative representation learning. The attention branch combines average and maximum feature statistics with learned directional spatial responses through residual channel and spatial recalibration. The contrastive branch mines lesion representations through supervised contrastive learning, class-prototype aggregation, and inter-class margin separation. On an expanded internal cohort of 887 PWS images, five-fold evaluation showed higher mean Accuracy, Recall, AUC, and BACC for LAPC than the compared methods under both backbones; these numerical comparisons do not establish statistical significance. Ablation experiments evaluated the individual and combined effects of lesion-aware attention and prototype contrastive learning across two backbone architectures, with supplementary experiments conducted on PH2 in a separate dermoscopic setting.