Machine Learning-Based Detection of Intestinal Metaplasia Identifies Clinically Missed Cases.

CONTEXT.—: Gastric intestinal metaplasia is a recognized precursor and risk factor to gastric cancer, and correct diagnosis is required for clinical decision-making.

OBJECTIVE.—: To evaluate potential missed intestinal metaplasia diagnoses and develop automated approaches to quantification, we developed an artificial intelligence pipeline for review of both whole slide images and pathology reports.

DESIGN.—: A patch-based classifier was trained and applied to 1935 gastric biopsy specimens. Specimens with disagreement between the artificial intelligence pipeline and the clinical pathology report were reviewed by 3 pathologists who were blinded to the diagnoses.

RESULTS.—: Following review, 30 of 297 apparently "false-positive" artificial intelligence detections were determined to represent undetected intestinal metaplasia. The artificial intelligence quantification of intestinal metaplasia strongly agreed with human scoring (Spearman rank correlation coefficient, 0.90; P < .001). Finally, a large language model-based evaluation of pathology report text showed near perfect ability to categorize reports as positive or negative for intestinal metaplasia (99.3% of specimens) with a few reports evaluated as equivocal (0.7%).

CONCLUSIONS.—: An artificial intelligence-based quality assurance pipeline could detect missed intestinal metaplasia in as many as 1.5% of gastric biopsy specimens, highlighting a potentially addressable diagnostic gap. An automated pipeline could additionally provide quantitative information that could be informative for management.