MS UX-Net: A Multi-scale Depth-Wise Convolution Network for Medical Image Segmentation

Semantic segmentation of 3D medical images plays an important role in assisting physicians in diagnosing and successively studying the progression of the disease. In recent years, transformer-based models have achieved state-of-the-art performances on several 3D medical image segmentation tasks. However, these methods still suffer from huge model sizes and high complexity. On the other hand, large-kernel depth-wise convolution networks have shown great potential to encode contextual information more efficiently and effectively than transformer-based networks in natural image segmentation tasks. Inspired by the success of large-kernel depth-wise convolution networks and their variants, we propose a multi-scale lightweight depth-wise convolution network termed MS UX-Net, a U-shaped network mainly composed of convolution operations with different kernel sizes. Specifically, we design a multi-scale feature extraction module for feature encoding, which extracts features into four different scales and learns special features of different scales effectively. Furthermore, we adopt multi-scale depth-wise convolution rather than single-scale standard convolution during the decoding stages, which yields a notable reduction in both the number of model parameters and the computational complexity. The competitive results on two public FLARE2021 and Synapse datasets and a private Pancreatic tumor dataset demonstrate the effectiveness of our method.