Wavelet Guided Visual State Space Model and Patch Resampling Enhanced U-Shaped Structure for Skin Lesion Segmentation

Automatic segmentation of skin lesions from dermoscopy images is crucial for the early diagnosis and treatment of skin cancer. However, this task presents significant challenges, including complex lesion shapes, indistinct boundaries, and variations in size and color. This study aims to enhance segmentation accuracy and robustness by integrating the Wavelet Guided Visual State Space (WGVSS) Block and Resampling Semantic Information Fusion (RSIF) into the segmentation process. Our proposed framework incorporates the WGVSS Block and RSIF within a U-shaped architecture specifically designed for skin lesion segmentation. The WGVSS Block addresses the challenge of effectively capturing multi-scale features, which are essential for accurately delineating irregular lesion boundaries. By employing discrete wavelet transform, this module decomposes features into low and high-frequency components. It utilizes selective scanning to learn frequency and spatial features that guide the extraction of relevant characteristics from the original image, thereby improving feature representation and enhancing the model’s ability to distinguish complex lesion shapes. On the other hand, RSIF tackles the issue of information loss that often occurs during the upsampling process, which can result in blurred or missing fine-grained details critical for accurate segmentation. By employing Resample Patch Resize, RSIF selectively preserves important features, ensuring that crucial details are maintained. This approach effectively mitigates the loss of resolution that typically occurs in conventional upsampling methods, thereby improving segmentation accuracy. Experimental validation on diverse medical datasets demonstrates that our approach significantly outperforms existing methods, such as VMUNet, H-vmunet and VMUNetv2, on the ISIC2017, ISIC2018, and PH2 datasets. The proposed architecture effectively addresses the challenges posed by irregular lesion shapes, varied sizes, and blurred boundaries, resulting in superior segmentation accuracy and robustness.