HiCASCADE: Hierarchical Multi-scale Transformer with CASCADE for Medical Image Segmentation

Various hierarchical CNN networks have proven their effectiveness in the field of medical image segmentation, however, due to the inherent shortcomings of CNNs, it is difficult for a single CNN network to fully utilize the long-range dependent information. Transformer has excellent ability to capture long-range features, which can make up for the shortcomings of CNN in this respect. However, previous work ignored the intrinsic structure unique to the medical image segmentation task, and could not adequately capture the local and long-range feature information of medical images, and there are limitations in the use of fusion of multi-scale features and it can't identify organ border areas very well. In this paper, we propose a new hierarchical network architecture, HiCASCADE, which combines the advantages of CNN and Trasformer to extract multi-stage features and enhance them by cascading attention modules. Specifically, we use hybrid CNN-Transformer modules to extract multiscale features, enhance the feature representation by efficiently fusing features at different scales using attention gate (AG), and then enhance the fused long-range and local contextual features in both the channel and spatial dimensions by a convolutional attention module (CAM) and suppress irrelevant background information. We train the model on the synapse multi-organ segmentation dataset and achieve an excellent score of dice 81.2, which outperforms medical image segmentation models such as Attention UNet, TransUnet, SwinUnet, and PVT-CASCADE.