Background Crohn’s disease (CD) shows marked histopathologic heterogeneity, and conventional histologic assessment remains dependent on subjective visual evaluation, limiting reproducibility and objective quantification. We aimed to develop and clinically validate a deep learning-based framework for quantitative assessment of endoscopic biopsies in CD. Methods We retrospectively analyzed 3,641 endoscopic biopsy slides from 687 patients with CD using a hierarchical deep learning workflow implemented on the Palgo platform. This workflow integrated tissue-compartment segmentation, crypt-level segmentation and classification, and inflammatory cell detection and classification. From the resulting whole-slide outputs, we derived quantitative histologic metrics, including cryptitis and crypt-abscess ratios, mucosal and submucosal inflammatory cell densities, hotspot-based density measures, and mucosal stromal proportion. Concordance with the pathologist assessment was evaluated using concordance correlation coefficients, and associations with clinical indicators were examined using Spearman’s rank correlation coefficients. Results The models showed robust internal performance across the main segmentation and classification tasks, with Dice coefficients and AUC values generally exceeding 0.95. AI-derived crypt abscess and cryptitis ratios showed good concordance with pathologist assessment, whereas agreement was weaker for certain cell-level outputs, particularly submucosal plasma cell density (CCC = 0.376). Grade-stratified calibration indicated a tendency toward AI over-detection in submucosal regions with negative or low manual scores, underscoring the need for cautious interpretation of this metric. Several AI-derived features were associated with clinical or endoscopic activity; however, these associations were generally weak to moderate and should be regarded as exploratory. In particular, crypt abscess ratio was associated with endoscopic activity; submucosal eosinophil density and the mucosa-to-submucosa plasma cell ratio were associated with CDAI; and mucosal stromal proportion was associated with CRP. Exploratory group comparisons further suggested greater submucosal inflammatory infiltration in CD biopsies than in the heterogeneous non-CD inflammatory colitis group. Conclusions This AI-based framework provides objective and reproducible tissue-level quantification of CD histopathology and may complement routine pathologic assessment. By integrating structured quantitative outputs with heatmap-based visual review, it offers a practical pathologist-supervised approach for standardizing histologic evaluation in CD.
Jiang et al. (Fri,) studied this question.