MCA-Net: Multi-scale Context-aware Aggregation Network for Medical Image Segmentation

Abstract

Segmentation is still a challenging task in medical image analysis, especially for multi-organ segmentation, in which the volume difference of each organ is large and the shape is irregular. In this paper, we propose a Multi-scale Context-aware Aggregation Network (MCA-Net) for multi-organ segmentation. MCA-Net mainly contains three major components: an encoder, a Multi-scale Context-aware Aggregation module, and a decoder. A hybrid encoder combining CNN and Transformer is adopted to compensate for the shortcomings of convolution operation. The multi-scale context-aware aggregation module is composed of multi-scale residual atrous convolutional (MRAC) block and residual pyramid pooling (RPP) block. Also, Convolutional Block Attention Module (CBAM) block is introduced to enhance important features and suppress unnecessary ones. Additionally, to reduce the loss of low-level information, cascaded up-sampling is employed to recover the resolution of original image, and skip connections are introduced in the middle to further integrate the feature information. Our method achieves competitive results on multi-organ segmentation and cardiac segmentation datasets compared with the state-of-the-art approaches.