MSGM-Net: multi-scale general-modal network for disease and organ segmentation

General and effective medical image segmentation methods achieve pixel-level accurate segmentation of lesions or organs. Existing 2-D networks have demonstrated their effectiveness on single-modal segmentation tasks, but difficult to exploit complementary information among modalities and detect the complete target location on multi-modal segmentation tasks. Moreover, most of the existing convolutional neural network (CNN) variant networks fail to effectively extract long-range dependency and global context information because of the limitation of the receptive field. To address the above issues, this study proposes a multi-scale general medical image segmentation 2-D network (MSGM-Net) based on CNN and transformer, which has a wider encoder for extracting more dense local details. Specifically, a channel-space attention module is designed to independently group feature maps and calculate channel attention weights within each group, which successfully fuses channel information among different scales and establishes an effective connection between channel and spatial features. We also introduce a multi-scale convolution module to acquire more abundant multi-receptive field features, which is beneficial for extracting multi-scale features and obtaining fine-grained local information. Extensive experiments on typical single-modal and multi-modal image segmentation tasks demonstrate the accuracy and effectiveness of MSGM-Net and show that our approach is significantly superior to current state-of-the-art methods. The code is available at https://github.com/Victory1120/MSGM-Net .