CSS-UNet: Convolution-State Space Enhanced UNet for Cardiac MRI Segmentation

Abstract

Cardiac magnetic resonance imaging (CMR) is widely adopted in clinic for the assessment of cardiac anatomical structures and functions, and accurately segmenting left ventricular, myocardium and right ventricle from CMR plays an important role in clinical practice. Convolutional neural networks (CNNs), for example, U-Net, have been widely used in CMR segmentation. However, current CNN-based models focus on extracting local features via convolution modules, which cannot well understand the long-range dependencies within images, leading to the sub-optimal solution for CMR segmentation task. Inspired by Mamba that efficiently captures global context information using state space model, we propose a novel CMR segmentation model, called CSS-UNet, that can capture both local features and global contexts simultaneously by fusing features from convolution block and visual state space block. Our new model follows the design of U-Net architecture that contains encoder and decoder with skip connections, where a proposed feature fusion module, called spatial-state space, is seamlessly integrated into our model. By using the spatial-state space module, the low-level and high-level features can be extracted and fused for capturing both global and local information, enhancing the capability of feature extraction of CSS-UNet. We evaluate our proposed model on two public CMR datasets, and the experimental results reveal that our proposed model outperforms the most widely-used UNet, demonstrating the effectiveness of our model.