DS-UNeXt: depthwise separable convolution network with large convolutional kernel for medical image segmentation

Accurate automatic segmentation of medical images is required in computer-aided diagnosis systems in clinical medicine. Convolutional neural networks (CNNs) based on U-shaped structures are widely used in medical image segmentation tasks. However, due to the intrinsic locality of the convolution operation, it is difficult for CNN-based approaches to learn the global information and long-range semantic information interactions using Swin-Unet. However, we find that UNet and Swin-Unet have the worst segmentation performance on small masses. To remedy this problem, this paper presents an end-to-end depthwise separable U-shaped convolution network with a large convolution kernel (DS-UNeXt) for the medical image segmentation of computed tomography (CT) images and magnetic resonance images (MRIs). Our network has a larger receptive field to extract features, which is useful for boosting the performance of multiscale medical segmentations. In DS-UNeXt, parallel depthwise separable spatial pooling (PDSP) is proposed to aggregate the global information. PDSP consists of multiple parallel depthwise separable convolutions to enhance the high-level semantic features. The proposed DS-UNeXt achieves Dice indices of 80.65% and 90.88% on the synapse for the multiorgan segmentation dataset and the automatic cardiac diagnosis challenge (ACDC) dataset, respectively. Moreover, extensive experiments show that DS-UNeXt transcends several state-of-the-art segmentation networks.