Towards deep-learning based detection and quantification of intestinal metaplasia on digitized gastric biopsies: a multi-expert comparative study

Current gastric cancer (GCa) risk systems are prone to errors since they evaluate a visual estimation of intestinal metaplasia percentages in histopathology images of gastric mucosa to assign a risk. This study presents an automated method to detect and quantify intestinal metaplasia using deep convolutional neural networks as well as a comparative analysis with visual estimations of three experienced pathologists. Gastric samples were collected from two different cohorts: 149 asymptomatic volunteers from a region with a high prevalence of GCa in Colombia and 56 patients from a tertiary hospital. Deep learning models were trained to classify intestinal metaplasia, and predictions were used to estimate the percentage of intestinal metaplasia and assign the OLGIM risk score. Results were compared with independent blinded metaplastic assessments performed by three experienced pathologists. The best-performing deep learning architecture classified intestinal metaplasia with F1-Score of 0.80 +- 0.01 and AUC of 0.91 +- 0.01. Among pathologists, inter-observer agreement by a Fleiss's Kappa score ranged from 0.20 to 0.48. In comparison, agreement between the pathologists and the best-performing model ranged from 0.12 to 0.35. Deep learning models show potential to reliably detect and quantify the percentage of intestinal metaplasia, achieving high classification performance. Visual estimation of intestinal metaplasia remains highly dependent on individual expertise, resulting in inter-observer variability. Deep learning models provide consistent estimates that could help reduce this subjectivity in risk stratification.