Simultaneous Right Ventricle End-diastolic and End-systolic Frame Identification and Landmark Detection on Echocardiography

End-diastolic (ED) and end-systolic (ES) frame identification and landmark detection are crucial steps of estimating right ventricle function in clinic practice. However, the complex morphology of the right ventricle and low-quality echocardiography pose challenges to these tasks. This study proposes a multi-task learning (MTL) framework to simultaneously identify the right ventricle ED and ES frames and detect anatomical landmarks for echocardiography. The framework contains an encoder and two branches: frame-branch and landmark-branch. The convolution neural network (CNN) encoder is employed for extracting the shared features of two branches. The frame-branch is built with a recurrent neural network (RNN) to select ED and ES frames. A heatmap-based model is used as the landmark-branch to detect the landmarks. Furthermore, instead of directly regressing the indexes of ED/ES frames, we form the frame identification as a curve regression problem, which achieves considerable performance. Experiments performed on the echocardiography dataset of 105 patients validate the effectiveness of the proposed approach, which leads to the average frame difference of 1.59 (±1.34) frames (ED) and 1.56 (±1.35) frames (ES) on the frame identification task, and the percentage of correctly predicted landmarks is 83.3%. These results demonstrated that our method outperforms most existing methods.