Modified MultiResUNet for Left Ventricle Segmentation from Echocardiographic Images

An accurate assessment of heart function is crucial in diagnosing the cardiovascular disease. One way to evaluate or detect the disease can use echocardiography, by detecting systolic and diastolic volumes. However, manual human assessments can be time-consuming and error-prone due to the low resolution of the image. One way to detect heart failure on echocardiogram is by segmenting the left ventricle on the echocardiogram using deep learning. In this study, we modified the MultiResUNet model for left ventricle segmentation in echocardiography images by adding Atrous Spatial Pyramid Pooling block and Attention block. The use of multires blocks from MultiResUnet is able to overcome the problem of multi-resolution segmentation objects, where the segmentation objects have different sizes. This problem has similar characteristics to echocardiographic images, where the systole and diastole segmentation objects have different sizes from each other. Performance measure were evaluated using Echonet-Dynamic dataset. The proposed model achieves dice coefficient of 92%, giving an additional 2% performance result compared to the MultiResUNet.