Left Ventricle Segmentation using Bidirectional Convolution Dense Unet

Abstract

Cardiac magnetic resonance technique is most useful technique to diagnosis the early cardiovascular diseases as noninvasive method to reduce the mortality rate .There are so many methods developed to locate the heart chambers to derive cardiac indices, among them left ventricle play an important role. Since, there is overlapping of heart chamber during different cardiac phase, the segmentation of left ventricle (LV) is challenging task. To improve segmentation of Left ventricle, we propose a dense Unet based architecture with bidirectional convolution LSTM (long short term Memory). In this method, dense block and bidirectional convolution used to extract diverse features from Cardiac Magnetic Resonance (CMR) image to locate LV cavity. We use publically available Automated Cardiac Diagnosis Challenge (ACDC) dataset for training and testing of proposed model. The proposed method achieved high dice score 0.97(ED) & 0.92(ES) for ACDC test dataset. The proposed model also tested on Multi-Vendor and Multi-Disease Cardiac Image Segmentation Challenge (M & M) dataset and achieves 0.897 dice score that demonstrate the effectiveness and robustness of the network.

URN:NBN:sciencein.jist.2023.v11.417