Predicting Endoscopic Grading of Gastric Intestinal Metaplasia using Small Patches

Gastric intestinal metaplasia (GIM) characterization is challenging for humans and AI models. Deep learning solutions for this task are sensitive to training data, which is particularly concerning given the wide range of acquisition conditions, sampling biases, and overall scarcity of high-quality datasets.In this paper, we set forth the GIM self-similarity hypothesis where we assume that an underlying stationary self-similar process governs the structural changes observed in the mucosa. To validate this hypothesis we show that a deep learning model can map an adequately placed patch to the endoscopic grading of GIM (EGGIM) of the entire still frame.To evaluate our approach, we collected and annotated both retrospective and prospective datasets with EGGIM scores. Our results are promising: using leave-one-patient-out cross-validation, the predictions from a ResNet-50 model can be used to correctly stratify the risk for 57 out of 65 patients with perfect sensitivity on an extremely biased dataset.