Application of semi-supervised learning models in the Los Angeles grading of reflux esophagitis
10.3969/j.issn.1005-202X.2025.09.016
- VernacularTitle:半监督学习模型在反流性食管炎Los Angeles分级中的应用
- Author:
Hang ZHAO
1
;
Xiaodan XU
;
Jinzhou ZHU
Author Information
1. 苏州大学附属第一医院消化内科,江苏 苏州 215506;苏州大学附属常熟医院消化内科,江苏 苏州 215005
- Publication Type:Journal Article
- Keywords:
reflux esophagitis;
self-supervised learning;
simple framework for contrastive learning of visual representation;
computer-aided diagnosis;
deep learning;
Grad-CAM
- From:
Chinese Journal of Medical Physics
2025;42(9):1236-1244
- CountryChina
- Language:Chinese
-
Abstract:
Objective To construct a classification model for the Los Angeles grading of endoscopic reflux esophagitis based on the SimCLR algorithm's semi-supervised learning framework.Methods The designed learning framework was pre-trained on a large unlabeled dataset through self-supervised learning,and further finely tuned on a small labeled dataset according to the Los Angeles grading criteria.The performance test on the model was conducted on an independent dataset,and the proposed model was compared with the models of supervised learning algorithms and endoscopists in terms of accuracy,Matthews correlation coefficient,and Cohen's kappa value.Finally,Grad-CAM and t-SNE were used for the visualization of the model's interpretation.Results The SimCLR model with ResNet as the backbone network showed superior performance in accuracy(0.840),Matthews correlation coefficient(0.800),and Cohen's kappa value(0.960)than the traditional supervised learning model with ResNet as the backbone(0.680,0.601,and 0.870)as well as junior endoscopists(0.770,0.713,and 0.940),but there was still a slight gap compared with senior endoscopists(0.850,0.813,and 0.960).In addition,the results of t-SNE showed that self-supervised learning in SimCLR was more effective in clustering multi-dimensional samples than traditional supervised transfer learning.Conclusion Compared with traditional supervised learning methods,semi-supervised learning demonstrates outstanding performance even with only a small number of labeled endoscopic images.
