Boundary-aware Multi-stage with Mobile U-Net for Hair Segmentation in Dermoscopic Images

Skin lesion analysis is essential for early skin cancer identification.However, hair is a common artifact in skin lesion images that can obscure lesion regions and reduce the accuracy of automated analysis.Effective hair segmentation is essential for accurate hair recognition and removal as a preprocessing step to ensure these occlusions do not lead to inaccurate feature extraction, misclassification, or degraded performance of computer-based diagnostic tools.This study proposes a novel hair segmentation approach that integrates multi-stage fine-tuning with a boundary hybrid loss and incorporates the Convolutional Block Attention Module (CBAM) or the Boundary Guided Feature Fusion (BGFF) into the Mobile U-Net's skip connections.Experiments were conducted on the ISIC image dataset.The proposed method improves hair segmentation, increasing the base model's Dice score from 76.97% to 79.08%, thereby surpassing existing methods.By providing more accurate hair segmentation, this approach facilitates precise visualization of hair and ruler marks for artifact removal, thereby supporting dermatologists and computer-aided diagnostic tools in achieving more reliable identification of skin cancer.