Introduction: Histological assessment of celiac disease active autoimmune insult and severity relies on villus height to crypt depth ratio (Vh:Cd) and intraepithelial lymphocyte (IEL) count, measures that are time-consuming and subjective. Furthermore, manual measurement of Vh:Cd depends on biopsy orientation. An opportunity exists to streamline the Vh:Cd and IEL measurement processes. Here, we examined an AI-powered tool for this purpose.
Methods: CeliacExplore,* a digital pathology tool that extracts quantitative histology features (N=360) from digitized, H&E-stained duodenal biopsies, was deployed on specimens (N=135) from the CeDLara Phase 3 trial (NCT03644069). Extracted features included villus epithelium and crypt epithelium area proportions and IEL density. Spearman correlations were computed between digital features and manual Vh:Cd and IEL count across all slides and separately within ‘well-oriented’ and 'maloriented' subsets (i.e., slides with suboptimal tissue orientation on the original WSI, requiring tissue recut for manual Vh:Cd scoring). A multivariable linear regression model incorporating three Vh:Cd-related digital features was evaluated.
Results: CeliacExplore-derived IEL density strongly correlated with manual IEL count (⍴=0.739; p<0.001) and MARSH score (⍴=0.676; p<0.001), while digital villous and crypt epithelium area proportions correlated with manual Vh:Cd (⍴=0.444; p<0.001) and MARSH score (⍴=-0.441; p<0.001). A multivariable regression model incorporating CeliacExplore epithelial cell and tissue features accurately predicted Vh:Cd regardless of orientation [RMSE=0.577 (‘well-oriented’); =0.562 (‘maloriented’)].
Conclusion: CeliacExplore features demonstrate robust concordance with histological measures of celiac disease severity. The persistence of feature correlation on maloriented slides highlights the potential of digital pathology to provide reliable information beyond the limitations of conventional histological review, with implications for both clinical trials and routine practice.
*For Research Use Only. Not for use in diagnostic procedures.