A Hybrid Deep Learning Approach for Skin Lesion Segmentation With Dual Encoders and Channel-Wise Attention

Abstract

Skin cancer poses a significant global health challenge due to its increasing incidence rates. Accurate segmentation of skin lesions is essential for early detection and successful treatment, yet many current techniques struggle to balance computational efficiency with the ability to capture complex lesion features. This paper aims to develop an advanced deep learning model that improves segmentation accuracy while maintaining computational efficiency, offering a solution to the limitations of existing methods. We propose a novel dual-encoder deep learning architecture incorporating Squeeze-and-Excitation (SE) attention blocks. The model integrates two encoders: a pre-trained ResNet-50 for extracting local features efficiently and a Vision Transformer (ViT) to capture high-level features and long-range dependencies. The fusion of these features, enhanced by SE attention blocks, is processed through a CNN decoder, ensuring the model captures both local and global contextual information. The proposed model was evaluated on three benchmark datasets, ISIC 2016, ISIC 2017, and ISIC 2018, achieving Intersection over Union (IoU) scores of 89.53%, 87.02%, and 84.56%, respectively. These results highlight the model’s ability to outperform current methods in balancing segmentation accuracy and computational efficiency. The findings demonstrate that the proposed model enhances medical image analysis in dermatology, providing a promising tool for improving the early detection of skin cancer.