Clinical validation of an AI-based pathology tool for scoring of metabolic dysfunction-associated steatohepatitis

Metabolic dysfunction-associated steatohepatitis (MASH) is a major cause of liver-related morbidity and mortality, yet treatment options are limited. Manual scoring of liver biopsies, currently the go

Review scope (28 September 2026): partial. The PMC full text was checked for the intended use, study timing and distinction between the consensus and median-panel comparisons. The review preserves the reported numerical results but is not a complete re-extraction of their confidence intervals, study populations, supplementary methods or conflicts of interest. The lecture-slide attribution to other tools remains unverified. See [[digital-pathology-evidence]] for the shared appraisal approach.

Summary

The analytical and clinical validation of AIM-MASH, an AI tool for scoring liver biopsies in metabolic dysfunction-associated steatohepatitis (MASH) drug trials. In these trials, histology decides who is enrolled and is the surrogate endpoint for accelerated approval. It is scored by hand on the NASH CRN system, with enough inter- and intra-reader variability to sink an underpowered phase 2b trial or force expensive consensus re-reads.

The study used archived specimens from completed global trials, with prospectively collected pathologist reads, to test the algorithm alone and as an assistant to expert hepatopathologists across 1,481 cases. Against the consensus reference, AI-assisted reads showed superior agreement to unassisted reads for lobular inflammation, hepatocellular ballooning, the enrolment criterion MAS ≥ 4 (with ≥ 1 in each component) and MASH resolution, and non-inferiority for steatosis and fibrosis. The algorithm's repeatability across scans and reproducibility across sites exceeded agreement between the study's own pathologists. These findings concern histology scoring in MASH clinical trials; they do not establish routine diagnostic effectiveness or patient-outcome benefit. Study design, Results and intended-use discussion.

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

Pulaski H, Harrison SA, Mehta SS, Sanyal AJ, Vitali MC, Manigat LC, et al. Clinical validation of an AI-based pathology tool for scoring of metabolic dysfunction-associated steatohepatitis. Nat Med. 2025;31(1):315–322. Epub 2024 Nov 4. doi: 10.1038/s41591-024-03301-2. PMID: 39496972. PMCID: PMC11750710.

• Model: AIM-MASH (AI-based measurement of MASH)

• Access: open access, full text in PubMed Central

Study at a glance

Item	Detail
Development (described in an earlier paper)	103,579 pathologist annotations on 6,235 H&E and 6,223 Masson's trichrome WSIs from six completed phase 2b/3 MASH trials
Architecture	Convolutional networks segment features into overlays and area proportions; graph neural networks predict the ordinal NASH CRN grade or stage per feature
Validation set	1,481 cases from completed trials of three drug candidates (semaglutide, pegbelfermin, resmetirom); about 13,000 independent reads
Timing and intended use	Archived trial material; prospectively collected assisted and unassisted reader assessments for validation of clinical-trial histology scoring
Ground truth	Two panels of hepatopathologists, each two primary readers plus a shared blinded tiebreaker, with a consensus call when all three disagreed
Comparators	Independent unassisted manual readers; AIM-MASH alone; AIM-MASH-assisted pathologists
Statistics	Non-inferiority margin of −0.1 in linearly weighted kappa against ground truth, then superiority testing
Scanner	Leica Aperio AT2 at ×40

Key findings

• Repeatability. The same slides scanned on three separate days gave mean agreement of 0.93–0.96 across steatosis, lobular inflammation, ballooning and fibrosis.

• Reproducibility across sites. With three external laboratories, operators and scanners, ballooning (0.91) met the 0.85 goal. Steatosis (0.86), lobular inflammation (0.85) and fibrosis (0.87) sat at the goal, with lower confidence bounds just below it. Every component still beat mean pairwise agreement between the study's pathologists.

• Algorithm alone vs manual readers. Superior for ballooning (+0.15 weighted kappa) and lobular inflammation (+0.12); non-inferior for steatosis and fibrosis.

• AI-assisted pathologists. Same pattern: superior for ballooning and inflammation, non-inferior for steatosis and fibrosis. For the trial composites, agreement with ground truth rose from 0.51 to 0.63 for MAS ≥ 4 with ≥ 1 in each category, and from 0.37 to 0.54 for MASH resolution.

• A second reference and comparator. In the median-panel analysis, AI-assisted agreement with the median reference from one reader panel was compared with agreement between that reference and the median scores of a second panel. Non-inferiority held for every component, but superiority was not observed. This differs from the primary comparison of assisted and individual unassisted reads against the consensus reference. Figure 4.

• Overlays. Three hepatopathologists judged the feature heatmaps on up to 160 frames per feature (from 222 WSIs). All overlays met the false-positive criterion. All except ballooning met the true-positive criterion; ballooning scored 0.87 and narrowly missed. The pathologists fully agreed that ballooning was present in only 55% of the frames any of them flagged.

• Assist workflow. The pathologist checks adequacy, staining and additional findings, but may override an AIM-MASH score only for a discrepancy of two points or more. The limit is deliberate, to stop inter-reader variability from creeping back in.

Limitations noted by the authors

• The specimens were archived trial material, but the validation reader assessments were collected prospectively. Re-staining or re-scanning was not available in this study; non-evaluable cases were under 4%. This differs from a live trial workflow in which quality-control problems could be addressed. Study design and limitations.

• New trial populations or drugs with novel mechanisms may challenge the locked algorithm, so ongoing performance monitoring is planned.

• The validated use is clinical-trial scoring. Routine diagnostic MASLD use may need further training and validation.

On the lecture slide

First row of a lecture slide titled "Top quantitative models in GI/liver path":

Disease	Model	Output	Training set	Reference
MASH	AIM-MASH (+qFIBS, Heinemann)	Steatosis, inflammation, ballooning, fibrosis scores	~103,579 annotations / 12,458 WSIs	Pulaski et al. Nat Med. 2025

The training figure matches the paper's Methods: 103,579 annotations on 6,235 + 6,223 = 12,458 WSIs. The slide's "+qFIBS, Heinemann" refers to other MASH quantification tools that are not part of this paper and are not captured here [unverified].