U-SwinTrans: Automated Skin Lesion Segmentation Using Swin Transformer

Skin lesions are difficult to segment because of their different shapes, sizes, colors, and textures. While Convolutional Neural Networks (CNNs) have proven effective at medical image segmentation, typical U-shaped networks that employ convolutional operations may struggle to capture global features successfully. To overcome these challenges, we incorporated Swin Transformer Blocks into a U-shaped architecture, enhancing the network’s capability to capture global features more effectively. We also used traditional image enhancement techniques such as DCT, DWT, and FFT to preprocess input images for better model performance. Our method entailed tokenizing preprocessed pictures into patches and putting them into the U-SwinTransformer encoder-decoder architecture, which learned both local and global semantic characteristics required for successful lesion segmentation. A comprehensive analysis of the ISIC 2017 dataset proves the effectiveness of our approach in skin lesion segmentation.