A feasibility study using quantitative and interpretable histological analyses of celiac disease for automated cell type and tissue area classification

Histological assessment is essential for the diagnosis and management of celiac disease. Current scoring systems, including modified Marsh (Marsh-Oberhuber) score, lack inter-pathologist agreement. To

Summary

A feasibility study of a fully automated, interpretable model of celiac disease histology from H&E-stained duodenal biopsies. The motivation is that the modified Marsh (Marsh–Oberhuber) classification, the standard way of grading celiac mucosa, has poor agreement between pathologists.

Convolutional networks trained on pathologist annotations identify cells (separating out intraepithelial lymphocytes), tissue regions (villous versus crypt epithelium) and artifacts. Human-interpretable features built from those predictions track the Marsh score. Area measures of villous atrophy fall as Marsh rises (r = −0.79), measures of crypt hyperplasia rise with it (r = 0.71), and further features separate celiac disease from normal duodenum.

The abstract in the description: frontmatter is the paper's own, verbatim. The sections below are my own-words digest of the abstract; the full text is open access in PubMed Central but was not captured here.

Citation

Griffin M, Gruver AM, Shah C, Wani Q, Fahy D, Khosla A, et al. A feasibility study using quantitative and interpretable histological analyses of celiac disease for automated cell type and tissue area classification. Sci Rep. 2024;14(1):29883. Published 2024 Dec 2. doi: 10.1038/s41598-024-79570-1. PMID: 39622903. PMCID: PMC11612272.

• Access: open access, full text in PubMed Central

• Overlap: four authors (Najdawi, Fahy, Khosla, Jayson) also wrote the ulcerative colitis model captured alongside this note, and both build human-interpretable features from cell and tissue models.

Study at a glance

Item	Detail
Source material	Duodenal biopsies of varying celiac disease severity plus normal duodenum, from a large central laboratory
Celiac disease slides	N = 318: training 230, validation 60, test 28
Normal duodenum slides	N = 58: training 40, validation 12, test 6
Models	Convolutional networks for cells, tissue regions and artifacts, trained on pathologist annotations
Readout	Human-interpretable features, correlated with modified Marsh scores (Spearman)

Key findings

• The models detected intraepithelial lymphocytes, separated villous from crypt epithelium, and flagged artifacts.

• Villous-atrophy area measures correlated negatively with Marsh score (r = −0.79); crypt-hyperplasia measures correlated positively (r = 0.71).

• Features distinguishing celiac disease from normal duodenum were identified.

• The authors frame the model as explainable and fully automated, with possible uses in diagnosis, prognosis, clinical trials and treatment-response monitoring.

Reading notes

My own notes, not content taken from the paper.

• The held-out test set is small (28 celiac and 6 normal slides), which fits the "feasibility" framing.

• The model is benchmarked against the Marsh score, the very system the paper says pathologists do not agree on. The correlations are only as meaningful as that reference, the same trap the gold standard paradox describes.

On the lecture slide

From a lecture slide titled "Top quantitative models in GI/liver path":

Disease	Model	Output	Training set	Reference
Celiac disease	Griffin CNN	IEL density, villous:crypt ratio	270 slides	Griffin et al. Sci Rep. 2024

"270 slides" matches the training split: 230 celiac + 40 normal. The full dataset was 376 slides.