A Multi-Fusion Residual Attention U-Net Using Temporal Information for Segmentation of Left Ventricular Structures in 2D Echocardiographic Videos

The interpretation of cardiac function using echocardiography requires a high level of diagnostic proficiency and years of experience. This study proposes a multi-fusion residual attention U-Net, MURAU-Net, to construct automatic segmentation for evaluating cardiac function from echocardiographic video. MURAU-Net has two benefits: (1) Multi-fusion network to strengthen the links between spatial features. (2) Inter-frame links can be established to augment the temporal coherence of sequential image data, thereby enhancing its continuity. To evaluate the effectiveness of the proposed method, we performed nine-fold cross-validation using CAMUS dataset. Among state-of-the-art methods, MURAU-Net achieves highly competitive score, for example, Dice similarity of 0.952 (ED phase) and 0.931 (ES phase) in ${\mathrm{LV}}_{\text{Endo}}$, 0.966 (ED phase) and 0.957 (ES phase) in ${\mathrm{LV}}_{\mathrm{Epi}}$, and 0.901 (ED phase) and 0.917 (ES phase) in $\mathrm{LA}$, respectively. It also achieved the Dice similarity of 0.9313 in the EchoNet-Dynamic dataset for the overall left ventricle segmentation. In addition, we show MURAU-Net can accurately segment multiclass cardiac ultrasound videos and output the animation of segmentation results using the original two-chamber cardiac ultrasound dataset MUCO.