Predicting Endoscopic Grading of Gastric Intestinal Metaplasia using Small Patches

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Characterizing Gastric Intestinal Metaplasia (GIM) is challenging for both human endoscopists and artificial intelligence (AI) models.
• Deep learning (DL) models for GIM analysis are highly sensitive to variations in training data, including acquisition conditions, sampling biases, and data scarcity.

Purpose of the Study:

• To propose and validate the Gastric Intestinal Metaplasia self-similarity hypothesis, suggesting a stationary self-similar process underlies mucosal structural changes.
• To demonstrate that a DL model can predict the endoscopic grading of GIM (EGGIM) for an entire endoscopic image from a single, well-placed patch.

Main Methods:

• Development of a DL model, specifically ResNet-50, to analyze endoscopic images.
• Collection and annotation of retrospective and prospective datasets containing EGGIM scores.
• Validation using leave-one-patient-out cross-validation to assess model performance on diverse patient data.

Main Results:

• The ResNet-50 model successfully predicted EGGIM scores from image patches.
• Leave-one-patient-out cross-validation demonstrated that the model could correctly stratify risk for 57 out of 65 patients.
• The model achieved perfect sensitivity in risk stratification, even on a highly biased dataset.

Conclusions:

• The GIM self-similarity hypothesis provides a novel framework for understanding GIM.
• The proposed DL approach shows significant potential for accurate and efficient GIM risk stratification.
• This method offers a promising solution for improving GIM characterization, especially in data-limited scenarios.