An automatic classification of metaplasia in gastric histopathology images

Gastric metaplasia (GM) has been classically related to the risk of progressing from gastritis to gastric cancer. Therefore, quantification of such progression is crucial to establish the type of intervention and to determine prognosis. Currently, the Operative Link for Gastritis Assessment (OLGA) and the Operative Link on Gastritis Assessment based on Intestinal Metaplasia (OLGIM) are the acknowledged protocols to assess and stage the risk of GM progression, from the lowest stage (stage 0, no metaplasia) to the highest (stage IV, severe metaplasia). However, these systems are qualitative, prone to error by the dependence on the expert and restricted by the number of biopsies required per patient. Hence, this paper presents an exploration of state-of-the-art convolutional neural networks (CNN) for the automatic classification of metaplasia in histopathology images of gastric tissue. The experimental results show that the best model was VGG16, under a binary cross entropy training, achieving an average accuracy of 0.76 +/- 0.022 and an F1-Score of 0.76 +/- 0.024 in test. Additionally, predictions were compared with the real annotations made by the expert, where the ResNet50 obtained the best performance with a Dice Score of 0.93 +/- 0.074 and its corresponding Jaccard Index of 0.87 +/- 0.129.