HDF-SegNet: Dynamic Integration of Hierarchical Information for Cardiac Magnetic Resonance Image Segmentation

Zijuan Wang, Yucheng Chen, Nan Mu, Xiaoning Li

Abstract: Cardiac Magnetic Resonance Imaging (CMRI) presents complex morphological features in 3D space due to different cardiac cycles, respiratory motions and individual differences, with diverse variations in global morphology and local details between slices, which requires segmentation models to have accurate detail capturing and boundary recognition capabilities. In recent years, deep learning methods based on traditional Convolution Neural Network, such as U-Net architectures, use fixed-size convolution kernels, which are difficult to adapt to the complex deformation of the heart structure. In this paper, we propose a hierarchical dynamic fusion segmentation network (HDF-SegNet) with a two-branch architecture by combining deformable convolutions that can generate corresponding sensory field shapes according to the target deformation. In it, the upper branch reconstructs the original image and generates a priori information and local features to guide segmentation. The lower branch uses Dynamic Large Kernel(DLK) to model the global information. We use the Feature Interweaving Module(FIM) to refine the different layers of features and fuse all the features after selecting them dynamically by channel and spatially. We experimented on the publicly available ACDC cardiac dataset with an average Dice coefficient of 91.86%, the comparison experiments validate the superiority of HDF-SegNet on the task of CMRI segmentation.