Development of a deep learning algorithm for Paneth cell density quantification for inflammatory bowel disease

Background Alterations in ileal Paneth cell (PC) density have been described in gut inflammatory diseases such as Crohn's disease (CD) and could be used as a biomarker for disease prognosis. However, quantifying PCs is time-intensive, a barrier for clinical workflow. Deep learning (DL) has transformed the development of robust and accurate tools for complex image evaluation. Our aim was to use DL to quantify PCs for use as a quantitative biomarker. Methods A retrospective cohort of whole slide images (WSI) of ileal tissue samples from patients with/without inflammatory bowel disease (IBD) was used for the study. A pathologist-annotated training set of WSI were used to train a U-net two-stage DL model to quantify PC number, crypt number, and PC density. For validation, a cohort of 48 WSIs were manually quantified by study pathologists and compared to the DL algorithm, using root mean square error (RMSE) and the coefficient of determination ( r 2 ) as metrics. To test the value of PC quantification as a biomarker, resection specimens from patients with CD (n = 142) and without IBD (n = 48) patients were analysed with the DL model. Finally, we compared time to disease recurrence in patients with CD with low versus high DL-quantified PC density using Log-rank test. Findings Initial one-stage DL model showed moderate accuracy in predicting PC density in cross-validation tests (RMSE = 1.880, r 2 = 0.641), but adding a second stage significantly improved accuracy (RMSE = 0.802, r 2 = 0.748). In the validation of the two-stage model compared to expert pathologists, the algorithm showed good performance up to RMSE = 1.148, r 2 = 0.708. The retrospective cross-sectional cohort had mean ages of 62.1 years in the patients without IBD and 38.6 years for the patients with CD. In the non-IBD cohort, 43.75% of the patients were male, compared to 49.3% of the patients with CD. Analysis by the DL model showed significantly higher PC density in non-IBD controls compared to the patients with CD (4.04 versus 2.99 PC/crypt). Finally, the algorithm quantification of PCs density in patients with CD showed patients with the lowest 25% PC density (Quartile 1) have significantly shorter recurrence-free interval (p = 0.0399). Interpretation The current model performance demonstrates the feasibility of developing a DL-based tool to measure PC density as a predictive biomarker for future clinical practice. Funding This study was funded by the National Institutes of Health (NIH).