Multiscale Kernel Atrous Convolution Network for Segmentation of Myocardium

Cardiovascular diseases are a global leading cause of death. The automatic segmentation of the left ventricle (LV) from magnetic resonance (MR) images is essential for quantitative analysis and assessment of cardiac functioning and for monitoring and diagnosing several heart diseases. In this paper, we introduce a multi-scale pre-trained framework enhanced by a kernel-based atrous spatial pyramid pooling block which facilitates information propagation by concatenating distinct features from the encoder blocks at various stages. Additionally, we incorporate convolutional block attention modules (CBAM) into the encoder to attain adaptive refinement features and employed different types of residual blocks for multiscale upsampling features at each decoder stage, and we have introduced efficient residual blocks based on dilated convolutional layers in the decoder. We use variety of pre-trained networks, including X-Net, UNet, DSNet, Xception, and Efficient models, serving as the base networks which seamlessly integrated with kernel-based atrous spatial pyramid pooling module to aggregate information from the encoder side. Experiments on the Myocardium benchmark dataset demonstrate better performance in basal and middle slices compared to state-of-the-art methods, although performance in apical slices is relatively less impressive. We have also made our code publicly accessible ¹ . ¹ https://github.com/RespectKnowledge/LGEMRI MyoSeg DL