A deep-learning-based model for assessment of autoimmune hepatitis from histology: AI(H)

Histological assessment of autoimmune hepatitis (AIH) is challenging. As one of the possible results of these challenges, nonclassical features such as bile-duct injury stays understudied in AIH. We a

Summary

Autoimmune hepatitis (AIH) has no pathognomonic lesion. Its biopsies show the elementary lesions of any chronic hepatitis, graded semi-quantitatively with poor inter-observer agreement, and non-classical features such as bile duct injury are understudied.

AI(H) is a stack of convolutional networks trained on the Aiforia platform. Each model targets one part of AIH histology:

• on H&E: tissue, liver microanatomy, necroinflammation, portal inflammation, immune-cell types and bile duct damage;

• on Sirius red: fibrosis.

The models were trained on pre-treatment biopsies from University Hospital Basel. Against pathologist annotations, test accuracy ranged from 72% (immune-cell classification) to 99% (tissue detection). The tool found bile duct injury in about two-thirds of AIH biopsies.

The abstract in the description: frontmatter is the paper's own, verbatim. The sections below are my own-words digest of the abstract and the open-access full text in PubMed Central.

Citation

Ercan C, Kordy K, Knuuttila A, Zhou X, Kumar D, Koponen V, et al. A deep-learning-based model for assessment of autoimmune hepatitis from histology: AI(H). Virchows Arch. 2024;485(6):1095–1105. Epub 2024 Jun 15. doi: 10.1007/s00428-024-03841-5. PMID: 38879691. PMCID: PMC11666607.

• Platform: Aiforia (cloud-based, supervised deep learning)

• Access: open access, full text in PubMed Central

Study at a glance

Item	Detail
Cohort	116 adult AIH patients (94 women, 22 men; mean age 59), Basel archive 1996–2020; 123 paired H&E and Sirius red pre-treatment slides
Split	Random 80/20: 99 training slides, 24 test slides
Annotations	Immune cells by two pathologists (7,868 immune-cell annotations in total); all other classes by one pathologist
Clinical grading	Ishak grading and staging by two pathologists in consensus, plus International AIH Pathology Group criteria
Scanner	3DHISTECH Pannoramic SCAN II, 40× (0.24 µm/px)

Key findings

Model	Stain	Test accuracy
Tissue detection	H&E / Sirius red	99.4% / 99.4%
Liver microanatomy (portal, parenchyma, central vein)	H&E / Sirius red	88.0% / 94.0%
Necroinflammation (interface hepatitis, focal and confluent necrosis)	H&E	83.9%
Bile duct damage	H&E	81.7%
Portal inflammation (mild / moderate / severe)	H&E	79.2%
Immune cells (lymphocyte, plasma cell, macrophage, eosinophil, neutrophil)	H&E	72.4%
Fibrosis	Sirius red	87.6% (88.0% in the full-text Results)

• Bile duct injury appeared in 68.6% of biopsies (81 cases). That is closer to one earlier report (72%) than to others (24%, 83%).

• Tracking the pathologists. AI(H) counts of focal necrosis, the extent of interface hepatitis and the moderate portal-inflammation ratio all rose in step with the pathologists' Ishak scores, and so did the densities of all five immune-cell types.

• Toward the consensus criteria. Fed into the International AIH Pathology Group criteria, AI(H) outputs sorted biopsies into "likely" and "possible" AIH with 88.2% agreement with the pathologist. Most errors came from overcalling interface hepatitis and from faded archival slides.

• Other hepatitides. Exploratory runs on a handful of drug-induced, HBV and HCV biopsies recognised the same elementary lesions, with some errors.

Limitations noted by the authors

• Single institution, with most annotations by one pathologist. Consensus grading meant inter-observer variability could not be measured.

• Accuracy fell in heavily necroinflamed or architecturally disrupted biopsies, and immune-cell detection struggled in dense infiltrates.

• The tool is image-only: no clinical or laboratory data, and other liver diseases were excluded, so it is not a diagnostic tool.

• Fibrosis used Sirius red only. There was no elastic stain to separate recent collapse from established fibrosis.

On the lecture slide

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

Disease	Model	Output	Training set	Reference
Autoimmune hepatitis	AI(H) (Ercan)	Immune-cell density, necrosis	99 WSIs	Ercan et al. Virchows Arch. 2024

"99 WSIs" matches the training split (99 training, 24 test, of 123 slides).