Boundary Attentive Spatial Multi-scale Network for Cardiac MRI Image Segmentation

Accurate automatic segmentation of cardiac MRI images can be used for clinical parameter calculation and provide visual guidance for surgery, which is important for both diagnosis and treatment of cardiac diseases. Existing automatic segmentation methods for cardiac MRI are based on U-shaped network structure introducing global pooling, attention and etc. operations to extract more effective features. These approaches, however, suffer from a mismatch between sensing receptive field and resolution and neglect to pay attention to object boundaries. In this paper, we propose a new boundary attentive multi-scale network based on U-shaped network for automatic segmentation of cardiac MRI images. Effective features are extracted based on channel attention for shallow features. With the goal of increasing segmentation accuracy, multi-scale features are extracted using densely coupled multi-scale dilated convolutions. In order to improve the ability to learn the precise boundary of the objects, a gated boundary-aware branch is introduced and utilized to concentrate on the object border region. The effectiveness and robustness of the network are confirmed by evaluating this method on the ACDC cardiac MRI dataset to produce segmentation predictions for the left ventricle, right ventricle, and myocardial. Comparative studies demonstrate that our suggested method produces superior segmentation outcomes when compared to other cardiac MRI segmentation methods.