Development of machine learning tool for segmentation and parameter extraction in cardiac left ventricle ultrasound images of patients with cardiomyopathy

Abstract

This abstract presents development of an automatic diagnostic tool based on machine learning that analyses cardiac ultrasound images of patients with cardiomyopathy in several views (4 chamber apical,2 chamber apical and M mode view). This process is generally done manually,where a cardiologist manually extracts a region of interest (ROI),which is a time-consuming and error-prone task [1]. The main aim of the developed tool is to perform automatic left ventricle (LV) segmentation and to extract relevant parameters in order to estimate the severeness of cardiomyopathy in patients. Dataset included 1809 images with apical view and 53 images with M view from real patients collected at three Clinical Centres in UK and Serbia. Separate methodologies have been implemented for analysing apical and M mode view,including U-net for segmentation,after which parameters such as left ventricular length (LVL),internal dimension (LVID),posterior wall thickness (LVPW) and interventricular septum thickness (IVS) are calculated,both in systole and diastole (Fig. 1). The tool has also been implemented on the platform with a user-friendly interface,which allows these two modules to be used either separately or combined. In order to validate the model and compare the results between gold standard and developed methodology,two cardiology specialists have independently manually annotated LV and measured relevant parameters. The results show that the model achieves dice coefficient of 92.091% for segmentation and average root mean square error (RMSE) of 0.3052cm for parameter extraction in apical view images and average RMSE of 1.3548cm for parameter extraction in M mode view. Fully automatic detection of cardiomyopathy in cardiac LV ultrasound images can help clinicians in supporting diagnostic decision making and prescribing adequate therapy.