WSA-UDNet: A Novel Framework for Medical Image Segmentation

Medical image segmentation is crucial for clinical diagnosis and treatment. Accurate segmentation remains challenging due to noise, artifacts, limited receptive fields, and insufficient multi-scale context modeling in existing methods. To address these limitations, we propose WSA-UDNet, a novel segmentation framework built upon an enhanced U-Net architecture. The core of WSA-UDNet is the WTConv module, which integrates convolution with wavelet transform to decompose features into low- and high-frequency components. This design extends the receptive field, captures multi-frequency responses, and effectively suppresses noise without increasing parameters. To further improve information aggregation, we design the MSCAT dual-stage collaborative attention mechanism, where a CECA channel attention module strengthens local–global feature interactions, and a lightweight single-head spatial attention, SHSA, efficiently captures global context while preserving fine details. Additionally, we designed the MSFA module to align multi-level semantic features, bridging the encoder–decoder gap and enhancing learning effectiveness. Experimental results on the BraTS2018 dataset show a Dice similarity coefficient of 89.1%, while the Synapse and HOCM datasets achieve 84.26% and 79.55%, respectively. These results demonstrate that WSA-UDNet achieves superior segmentation accuracy, strong generalization, and robust sensitivity to class imbalance, providing an effective and efficient solution for medical image segmentation that balances semantic richness with computational efficiency.