DUWS Net: Wavelet-based dual U-shaped spatial-frequency fusion transformer network for medical image segmentation

Abstract

Medical image segmentation is crucial for disease monitoring, diagnosis, and treatment planning. However, due to the complexity of medical images and their rich frequency information, networks face challenges in segmenting regions of interest using single-domain information. This study proposes a wavelet-transform-based dual U-Net fusion Transformer network for medical image segmentation, aiming to address the shortcomings of current methods. The network supplements spatial information through an external U-Net encoder-decoder structure, enabling deeper extraction of spatial features from the images. The internal U-shaped structure utilizes wavelet transform to capture low-frequency and high-frequency components of feature maps, performing linear self-attention interactions between these frequencies. This allows the network to learn global structures from low frequencies while capturing detailed features from high frequencies. Finally, spatial and frequency domain features are fused through alternating weighting based on spatial and channel dimensions. Experimental results show that the proposed method outperforms traditional single-domain segmentation models.