Medical image segmentation network based on multi-scale cross-attention and wavelet transform

Medical image segmentation is a key technology for computer-aided diagnosis and treatment planning, essential for early disease detection, precise diagnosis, and prognosis evaluation. With advancements in clinical imaging technologies, deep learning segmentation models represented by U-Net have been widely applied in medical image processing. However, challenges persist when processing complex medical images, including insufficient multi-scale feature capture, limited long-distance dependency modeling, and inadequate integration of global contextual information. To address these challenges, this paper proposes MSCWNet, a medical image segmentation network based on multi-scale cross-attention and wavelet transform. The network introduces a Multi-Scale Cross-Attention Module (MSCAM) to enhance feature selection capability and long-distance dependency modeling, employs Wavelet Transform Convolution (WTConv) for effective multi-scale feature extraction, and implements a Dynamic Upsampling module (DySample) to improve global context integration and detail reconstruction. Experiments on four medical image segmentation datasets, GlaS, MoNuSeg, CVC-ClinicDB, and Kvasir-Seg, demonstrate that MSCWNet achieves superior performance compared to eight mainstream segmentation models, with significant improvements in Dice and IoU metrics. The code of MSCWNet is available at https://github.com/qf1518/MSCW .