Predicting Metaplasia in Upper Gastrointestinal Images from White Light Endoscopy

The risk of developing gastric cancer is associated with the presence of precursor lesions, like metaplasia, currently assessed by visually scoring histopathology images. Metaplasia appearance changes have been documented during endoscopy, however they need to be highlighted by expensive dyes or devices which require high levels of expertise to be properly applied. In contrast, White Light Endoscopy (WLE) is a more cost-effective option, but it reports inter-observer variability in metaplasia diagnosis. In this scenario, computational strategies may extract subvisual patterns and associate them with metaplasia during a conventional endoscopy. This paper introduces a method to extract features from WLE images to train a neural network for predicting the presence of metaplasia, in this case confirmed by histopathological analysis. With a set of 57 real cases (34 controls and 23 diagnosed with OLGIM 1 to 3). The method achieves an accuracy and F1-score of 76.5% and 73.7%, respectively for the binary problem.