Multi-scale wavelet network algorithm for pediatric echocardiographic segmentation via hierarchical feature guided fusion

Abstract

The automatic segmentation of critical anatomical structures in pediatric echocardiography is the essential steps for early diagnosis and treatment of congenital heart disease. However, current segmentation algorithms rarely extract the information based on effective feature enhancement algorithms. Simultaneously, the algorithms are susceptible to image quality and the lack of detail information. To solve this, we propose a multi-scale wavelet network (MS-Net) combined with a bidirectional feature fusion (BFF-Net) and a wavelet-Unet (W-Unet) for end-to-end pediatric echocardiographic segmentation. In MS-Net, the entire network uses the discrete wavelet transform (DWT) instead of the sampling operation to reduce the impact of image noise while avoiding information loss. The algorithm enhances the edge information by designing the edge attention module (EAM) in the BFF-Net branch and fuses the context and detail information via the bidirectional feature fusion. Secondly, this algorithm uses W-Unet to obtain the detail features of high-resolution images by network depth reduction and propagation method, which supplements the features extracted by the BFF-Net branch. Finally, the hierarchical features of BFF-Net and W-Unet are fused and updated by guided filtering (GF) to obtain the final segmentation prediction. Using 127 pediatric echocardiographic cases of self-selection as the experimental dataset, the left atrium and left ventricle of the echocardiogram were segmented. The Dice coefficient values of 0.9532 and 0.9155, the pixel accuracy of 0.9914, and the specificity values of 0.9975 and 0.9984 were obtained. It was thus verifying its potential and effectiveness as a clinical auxiliary tool.