CCNET: Cascading Convolutions for Cardiac Segmentation

Abstract

Myocardial segmentation plays a pivotal role in the clinical diagnosis of cardiac diseases. The difference in size and shape of the heart poses an extensive challenge to the clinical diagnosis. Being specific, the large amount of noise generated by the cardiac magnetic resonance (CMR) images also gives rise to substantial interference in the clinical diagnosis. Inspired by associated tasks, we put forward a network for the myocardium segmentation. In the proposed methodology, at first, we establish numerous sub-sampling layers in a bid to attain the high-level features, together with fusing the feature information of different visual fields by assuming different convolution kernel sizes. Thereafter, high-level features coupled with initial input features are merged by means of a plurality of cascaded convolution layers. It is capable of directly improving the performance of myocardium segmentation. We perform an assessment of our approach on 165 CMR T1 mapping images with lower PSNR, and the results demonstrate that our architecture outperforms previous approaches.