Automated left ventricle localization from cardiac MR images using deep learning

Ischaemic heart disease is caused by the blockage of blood flow to the heart and led to an increase in the number of deaths in Malaysia with a total of 18,267 deaths reported in 2018. The use of late gadolinium enhancement (LGE) contrast agents in cardiac magnetic resonance imaging (CMR) for the evaluation of post-MI patients have demonstrated an incremental prognostic value. The LGE allows for direct visualisation of scarred myocardial tissue as a region interest (ROI) that is being enhanced. By assessing the enhanced ROI in the left ventricular (LV), the myocardial scar can be detected and diagnosed. However, the current approach for detecting myocardial scar in LV from CMR images is done visually by the radiologist and is time-consuming and subject to variability. Therefore, it is proposed to incorporate computer-aided diagnosis using a deep learning approach based on the YOLO algorithm for automatically locating the LV region that will improve the diagnostic accuracy of the myocardial scar tissue via LGE data acquired in post-MI patients. In this work, a total of 159 images from 10 subjects are selected and split into three datasets i.e. training, validating and testing datasets with the ratio of 80%, 10% and 10% respectively. The deep learning techniques based on YOLOv2 and YOLOv3 with three different solvers (ADAM, RMSProp and SGDM) are used to evaluate the performance of the automated LV localization from the CMR images. The highest localization accuracy is obtained from YOLOv2 using an ADAM solver with an average precision (AP) of 100% and mean intersection over union (IoU) of 89%.