Triplet Attention-Enhanced UNet Architectures for Advanced Skin Lesion Segmentation

ccurate segmentation of skin lesions is critical the early detection and successful treatment of skin cancer, particularly melanoma, which has a significant impact on patient survival rates. This paper investigates a novel approach to skin lesion segmentation by integrating the Triplet Attention mechanism into UNet-based architectures, aiming to enhance the precision and reliability of automated segmentation in der-moscopic images. The proposed model leverages the strengths of both UNet’s encoder-decoder structure and the attention mechanism’s ability to focus on critical features while suppressing irrelevant background noise. Additionally, an Adaptive Weighted Loss function is introduced to tackle the class imbalance often present in skin lesion datasets, thereby enhancing segmentation performance. Extensive experiments were conducted using publicly available datasets, demonstrating significant improvements in segmentation accuracy and robustness, particularly in challenging cases with unclear lesion boundaries and varying illumination conditions. Our results indicate that the Triplet Attention-Enhanced UNet model, combined with the Adaptive Weighted Loss function, outperforms existing methods, providing a more reliable tool for dermatologists in the early diagnosis and treatment planning of skin cancer.