Improved Computer-Aided Diagnosis System for Nonerosive Reflux Disease Using Contrastive Self-Supervised Learning with Transfer Learning

Article Improved Computer-Aided Diagnosis System for Nonerosive Reflux Disease Using Contrastive Self-Supervised Learning with Transfer Learning Junkai Liao 1, Hak-Keung Lam 1,*, Shraddha Gulati 2, and Bu Hayee 2 1 Department of Engineering, King’s College London, London, United Kingdom 2 King’s Institute of Therapeutic Endoscopy, King’s College Hospital NHS Foundation Trust, London, United Kingdom * Correspondence: hak-keung.lam@kcl.ac.uk Received: 5 February 2023 Accepted: 21 July 2023 Published: 26 September 2023 Abstract: The nonerosive reflux disease (NERD) is a common condition, the symptoms of which mainly include heartburn, regurgitation, dysphagia and odynophagia. The conventional diagnosis of NERD needs the endoscopic examination, biopsy of the lining of the esophagus (mucosa), and ambulatory pH testing over 24 to 96 hours, which is complex and time-consuming. To address this problem, a computer-aided diagnosis system for NERD (named NERD-CADS) has been proposed in our previous paper. The NERD-CADS offers a more convenient and efficient approach to diagnosing NERD, which only requires the input of endoscopic images into the computer to produce a nearly instant diagnostic result. The NERD-CADS uses a convolutional neural network (CNN) as a classifier and can classify the endoscopic images captured in the esophagus of both healthy people and NERD patients. This is, in fact, a classification problem of two classes: non-NERD and NERD. We conduct ten-fold cross-validation to verify the classification accuracy of the NERD-CADS. The experiment shows that the mean of ten-fold classification accuracy of the NERD-CADS test reaches 77.8%. In this paper, we aim to improve the classification accuracy of the NERD-CADS. We add the contrastive self-supervised learning as an additional component to the NERD-CADS (named NERD-CADS-CSSL), and investigate whether it can learn the capability of extracting representations to improve the classification accuracy. This paper combines the contrastive self-supervised learning with transfer learning, which first employs massive public image data to train the CNN by the contrastive self-supervised learning, and then uses the endoscopic images to fine-tune the CNN. In this way, the capability of extracting representations (learned by the contrastive self-supervised learning) can be transferred into the downstream task (NERD diagnosis). The experiment shows that the NERD-CADS-CSSL can obtain a higher mean (80.6%) in tests than the NERD-CADS (77.8%).