Deep-learning-based Detection of Food Hypersensitivity from Confocal Laser Endomicroscopy Images of the Gastro-intestinal Tract

Abstract

Food hypersensitivity (FHS) is a relatively common pathological condition characterised by adverse reactions to specific foods, and that currently has limited diagnostic methods. Diagnostic tools using gastro-intestinal confocal laser endomicroscopy (CLE) images have recently been proposed for the assessment of FHS, but their interpretation can be challenging even for trained physicians. We propose to alleviate this problem by training machine learning models on CLE images of the gastro-intestinal tract for the binary classification problem of recognising images that show an adverse reaction to food intake. More specifically, the performances of four state-of-the-art image classification models (VGG16, Inception-v3, Xception, MobileNet-v2) are compared on one dataset acquired from 38 patients with proven FHS. Additionally, the models decisions are interpreted using the Grad-CAM technique. Our study shows that although all four models achieve satisfying classification performances, they learn very different features in terms of interpretability from the clinical perspective.