Attention graph convolutional nets for esophageal contraction pattern recognition in high-resolution manometries. (July 2021)

• Record Type: Journal Article
• Authors: Wang, Zheng; Yan, Lu; Dai, Yuzhuo; Lu, Fanggen; Zhang, Jie; Hou, Muzhou; Liu, Xiaowei
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: High-resolution manometry (HRM) -- Contractile vigor propagation graph (CVP) -- Attention graph convolution network (GAT) -- Convolutional neural networks (ConvNet); Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102734
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We designed efficient vigor propagation graphs while capturing the contraction and pressure propagation between vigor. An interpretable mechanism explicitly models the relevant vigor propagation and to classify various vigor occurrences. An efficient inference scheme over vigor propagation graphs by utilizing a GAT with a sparse spatial sampling strategy. The introduced network has the potential to conduct contractile propagation reasoning over contractile propagation. The proposed approach achieves state-of-the-art results on HRM images. Abstract: Diagnosis of the esophageal motility disorders is ongoing in clinical evaluations, which is based on traditional method called High-resolution manometry (HRM). However, the huge raw swallow data sets from the HRM are not allowed the doctors to interpret and classify the patients with esophageal symptoms. To this end, modeling propagation between vigor is useful for recognizing esophageal contraction patterns in large-scale high-resolution manometries. In this paper, we learned a discriminative propagation between vigor using deep learning methods. Furthermore, we designed an efficient graph to incorporate contractile vigor propagation (CVP) that considers the contraction and pressure propagation between vigor in HRM images. Using an attention graph convolutional network (GAT), the edges in CVP can automatically learn the contraction patterns (i.e., local temporal information) trends in time series and propagation (i.e., global information) through the attention units. The attention mechanism layer leverages the short-term trend to improve the prediction accuracy. The quantitative experiments showed that the proposed method achieves high accuracy in esophageal contraction pattern recognition and demonstrates its effectiveness compared to existing traditional methods. We also visualized the learned vigor-specific propagation patterns in the contractile features, which show that the proposed CVP-GAT is able to develop interpretable propagation information for esophageal contraction pattern recognition.