Estimation of the Left Ventricle Volume Using Semantic Segmentation

The left ventricle (LV) volume is a major indicator of heart disease. Current clinical procedure for volume estimation from magnetic resonance (MR) images include manual inspection by the experienced cardiologists, which is often a time-consuming process. This paper presents an automated method for LV volume estimation at two points in the time. The method consists of training a convolutional neural network (CNN) with the MR images from the Sunnybrook dataset. After that, the trained model is used to predict left ventricular segments in MR images from the Kaggle dataset. From predicted segments, the ventricular volumes at end-systole (ESV) and end-diastole (EDV) are calculated using numerical integration. The obtained results estimate the ESV and EDV with a mean absolute error of 14.688 and 19.222 mL, respectively.