S3040 Deep Learning Algorithm for Duodenal Biopsy Classification

Introduction: Duodenal biopsy combined with tissue transglutaminase (TTG-IgA) antibodies is the gold standard for celiac disease diagnosis. Issues with biopsy include artifacts, poorly oriented specimens, and inter-observer variability. Artificial intelligence has the potential to address these challenges. We aimed to determine if deep learning, image analysis using artificial intelligence, can be used on duodenal biopsies to accurately diagnose celiac disease. Methods: The target study population were individuals with a duodenal biopsy and TTG-IgA antibody test. Inclusion criteria were adults ≥18 years. Exclusion criteria were those following a gluten-free diet. The study period was 2008-2021. Patients were grouped into 5 categories: controls, celiac disease, increased intraepithelial lymphocytes (IELs) only, IELs with a positive TTG-IgA antibody, and seronegative villous atrophy. Biopsy were stained with hematoxylin and eosin. Whole slide images were obtained and placed on a High-Performance Computing node with a graphics processing unit. Clustering-constrained attention multiple-instance learning was the deep learning algorithm used. A patch size of 256 × 256 pixels was used for the segmentation step. The dataset was split into training, validation, and testing splits with 10-fold cross validation. Accuracy and area under the receiver operating characteristics curve (AUC) were reported for each fold. Results: A total of 2,392 patients’ whole slide images were included: 1290 controls, 414 celiac disease, 441 IELs only, 84 IELs with a positive TTG-IgA, and 163 with seronegative villous atrophy. For all categories, the average AUC was 0.849 with an accuracy of 71%. The best performing fold had an AUC of 0.884 with 75% accuracy. The confusion matrix for this fold demonstrates difficulty differentiating between IELs with a positive TTG-IgA antibody and other classifications. The model reliably differentiated celiac disease from controls average with AUC 0.992 and accuracy 97%. To distinguish celiac disease from seronegative villous atrophy the average AUC was 0.810 with accuracy 76%. Conclusion: This deep learning model can discriminate between celiac disease and normal histology. It can also separate patients into 5 classifications, representing a promising tool that would benefit from further optimization to improve diagnostic accuracy, reduce the manual pathology review burden, and enhance resource management.