MAEF-Net: Multi-Attention Efficient Feature Fusion Network for Deep Learning Segmentation

Segmentation of cardiac chambers and quantification of clinical functional parameters in dynamic echocardiography are the key to clinical diagnosis of heart disease. It is challenging and tedious for clinicians to identify end-diastolic frames (EDFs) and end-systolic frames (ESFs) in echocardiogram cine loops and to manually segment the left ventricle (LV). In this work, we presented a new multi-attention efficient feature fusion network (MAEF-Net) for automatic identification of EDFs and ESFs in echocardiogram videos and for automatic segmentation of the LV. Multi-attention mechanism was incorporated to guide the network to capture the characteristics of heart beat more efficiently, even in low-quality videos. At the same time, the deep supervision mechanism and spatial pyramid feature fusion were used to enhance the ability of feature extraction and gradient flow information was used to accurately segment the region of interest. MAEF-Net was trained and tested using the large public EchoNet-Dynamic data set. The Dice similarity coefficient of LV segmentation reached (93.10 ± 2.22)%. The key frame positioning error was (2.80 ± 0.22) frames. By using the information across multiple cardiac cycles, our model is more repeatable and robust with clinical interpretability. The proposed MAEF-Net may be used as a new method for automatic segmentation of the LV and identification of EDFs and ESFs.