Mitral valve regurgitation is a heart valvular disease involving blood backflow from left ventricle to left atrium. It is the second most common valvular disease. Report reviewed that at least 1.7% of adults in the US had suffered from moderate MR, and approximately 45% of the 134874 people investigated in China were experiencing different levels of MR. However, not all MR patients require the same type of treatment. Thus, evaluating the severity of MR is one of the most significant aspects in treating MR. Until now, multiple methods and standards have been developed to obtain the severity of MR using Doppler echocardiography. Yet, results from these methods are not always accurate. For example, vena contracta width is one of the references to classify the severity. An overestimation of MR usually occurs when MR is not holosystolic. Therefore, techniques to avoid overestimation and underestimation of the severity of MR are extremely valuable to the treatment of MR. Deep learning is an expanding technique in medical imaging. Recent years, multiple studies have been conducted on medical images using neural networks, such as convolutional neural network (CNN) and deep belief neural network (DBN). In this project, we present a deep neural network to support a more accurate MR severity estimation than previous methods.