Prediction of esophageal and gastric varices rebleeding for cirrhotic patients based on deep learning. (March 2023)

• Record Type: Journal Article
• Authors: Liu, Ziyi; Xu, Lulu; Qin, Na; Yang, Aisen; Chen, Yuan; Huang, Deqing; Shu, Jian
• Is Part Of: Biomedical signal processing and control. Volume 81(2023)
• Publication Date: 2023-03
• Subjects: Cirrhotic patient -- Esophageal and gastric varices rebleeding -- CECT image -- Deep learning; 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.2022.104420
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: When doctors use images to predict esophageal and gastric varices rebleeding (EGVR) for cirrhotic patients, the human can be affected by subjective factors. The purpose of this study is to use a deep learning method to automatically predict EGVR via contrast-enhanced computed tomography (CECT) sequences. A data set of 141 samples, which include 7050 CECT images retrospectively enrolled in five years, was used to train, validate and test the models. In order to sample and extract valuable stereo features from multiple slices, a novel module called Deformation Separable Convolutional Module is designed to replace the common pooling layer. In addition, multiple kinds of network structures are utilized to design a new deep learning model called Multi-slice Fusion Network (MSFN) for predicting EGVR. Various widely used models of deep learning for video processing as well as the proposed model were used to predict EGVR in the research process. The performance of these models was analyzed by comparing various evaluation metrics. In experiments, the most of metrics of best-performing MSFN (proposed) were higher than those of other methods. In terms of the calculation time of the algorithm, the average time cost of each image predicted by MSFN was only 0.4 s. The proposed MSFN not only has high precision and robustness but also has the condition of being applied to embedded medical devices. This technology has great potential for medical applications. The source code mentioned in this article will be released later. Highlights: In this paper, the deep learning algorithm is first used to predict EGVR. The DSCM is designed to filter and fuse multi-slice context information. The MSFN with information-fusion ability can achieve high prediction performance.