Echocardiographic image segmentation using mutual boosting network

Abstract

The segmentation of cardiac boundary on echocardiographic images can provide important information for cardiac diagnostic and functional assessment. Due to the low image quality, traditional segmentation methods are subject to large performance variation and low segmentation accuracy. Manual segmentation, on the other hand, is slow, laborious, and observer-dependent, making it not suitable for real-time image-guided interventions. In this study, we developed a novel deep learning-based method to segment the myocardium, endocardium of the left ventricle and left atrium rapidly and accurately. The proposed method, named mutual boosting network, consists of three modules, i.e., localization module (L module), classification module (C module) and segmentation module (S module). The L module is used to detect the region-of-interests (ROIs) of cardiac substructures. The C module and S modules can then derive the classification and segmentation of each substructure within its respective detected ROIs. We conducted a five-fold cross-validation on 100 patients’ cases. The endocardium (LV_Endo) and epicardium (LV_Epi) of the left ventricle and left atrium (LA) were segmented using the proposed method. The segmentation accuracy was evaluated using the Dice similarity coefficient (DSC) and mean absolute distance (MAD). At the end diastole (ED) phase, the DSC and MAD are 0.94±0.02 and 0.05±0.05mm for the LV_Endo, 0.95±0.02 and 0.04±0.05mm for the LV_Epi, and 0.88±0.09 and 0.25±1.3mm for the LA. At the end systole (ES) phase, the DSC and MAD are 0.93±0.04 and 0.07±0.15mm for the LV_Endo, 0.95±0.03 and 0.05±0.1mm for the LV_Epi, and 0.90±0.06 and 0.14±0.64mm for the LA. The high DSC and sub-milimeter MAD values demonstrate the potential of the proposed method in myocardial functions assessment and real-time interventional image guidance.