Addressing Class Imbalance in Skin Lesion Segmentation: A U-NET Approach with Focal Loss and RESNET50V2

Abstract

This study proposes a robust skin lesion segmentation framework to improve early melanoma diagnosis. The approach integrates three key components: Th first one is a data augmentation through geometric transformations (rotation, flipping, zooming, and shearing) to improve generalization across diverse dermoscopic images; the second component is an hybrid U-Net architecture with a pre-trained ResNet50V2 encoder to enhance hierarchical feature extraction while preserving spatial resolution; and finally a focal Loss to address class imbalance by focusing training on hard-to-classify lesion pixels. Evaluated on the PH2 and ISIC 2016 datasets, the proposed model achieves significant improvements in Dice (96%) and Jaccard (97%) scores, outperforming baseline models. This work contributes a reliable and accurate computer-aided diagnosis (CAD) framework for early skin cancer detection.