Automated Aortic Valve Calcific Area Segmentation in Echocardiography Images Using Fully Convolutional Neural Networks

Aortic valve calcium scoring is extensively utilized for diagnosing, treating, monitoring, and assessing the risk of aortic stenosis and coronary artery disease. The gold standard method for determining aortic valve calcium score is computed tomography (CT). However, CT is expensive and exposes patients to radiation. Alternatively, echocardiography, which is a cheaper and radiation-free method, may not provide an objective result due to inter- and intra-observer variability among cardiologists. Therefore, automated measurement of aortic valve calcium score is essential. In this paper, a fully convolutional neural network approach is proposed for the segmentation of aortic valve calcified regions from echocardiography images, as the first step in predicting aortic valve calcium score. This aims to achieve a more objective, faster, easier, less costly, and radiation-free automatic calculation of aortic valve calcium score. The proposed method achieves a precision of 86.3%, recall of 80.27%, a dice coefficient of 76.27%, and a Jaccard index of 61.9% on the new dataset including 118 echocardiography images across 82 individual patients. These results show the feasibility and potential efficacy of fully automating aortic valve calcium scoring from echocardiography images.