MASH trial pathology scoring gains a regulator-qualified AI tool: AIM-MASH AI Assist leaves under 2% of cases needing secondary review

Liver biopsy scoring in MASH clinical trials now has a regulator-qualified AI-assisted workflow that works alongside pathologists to produce MASH CRN grades and stages for enrollment and endpoint assessment. Scoring previously relied on a single central reader, or in later-stage trials on dual readers with a third pathologist for adjudication or even three-pathologist consensus panels, while inter-reader and intra-reader disagreement persisted. Across 1,400 biopsies from three MASH clinical trials, AI-assisted pathologists' agreement with consensus ground truth was non-inferior to single unassisted pathologists on all four cardinal histologic features and greater for lobular inflammation and hepatocellular ballooning.

The tool outputs pixel-level prediction overlays for the four cardinal MASH features plus the corresponding CRN grades and stages directly from whole slide images of H&E- or Masson Trichrome-stained biopsies. Training used roughly 8,700 H&E-stained slides and roughly 7,600 Masson Trichrome-stained slides from six completed phase 2b and phase 3 MASH clinical trials, with about 60 expert MASH pathologists contributing over 100,000 annotations of tissue features, fibrosis and artifact regions. Ten sequential deployments on the same whole slide images were 100% repeatable for each of the four cardinal features; repeatability across rescans of the same glass slides on different days ranged from 92.6% to 96.3%, and reproducibility across scans at three different laboratories ranged from 84.7% to 91.2%.

In the AI-assisted workflow covered by the EMA qualification, pathologists review the model score, secondary review is triggered only by a disagreement of 2 or more points, and consensus calls follow if disagreement persists. The workflow keeps algorithm output under human oversight, preserving clinicians' ability to override results and check slides for additional safety findings. Under this workflow, fewer than 2% of all cases required secondary review and fewer than 1% required consensus calls.

The demonstrated scope is enrollment and endpoint scoring in MASH clinical trials: the tool is qualified in the EU and as a Drug Development Tool in the US for that setting, and is explicitly not for diagnostic procedures. Repeatability and reproducibility data come from rescans of the same glass slides on different days and at three laboratories, and the agreement comparison comes from 1,400 biopsies in three trials.

Scoring variability has been hypothesized to artificially lower observed effect sizes and make potentially effective drugs appear unsuccessful, and that remains a hypothesis in the source. A retrospective analysis of the WAYFIND trial showed that a result which missed its primary endpoint by manual scoring would have met it under the validated AIM-MASH AI Assist read workflow, but this is retrospective rather than prospective validation. The source notes that AI-based scoring is unprecedented in MASH clinical trials and that continued collaboration among PathAI, trial sponsors and regulatory agencies will be needed to incorporate the tool effectively into trial protocols.