Trichrome-based deep learning for quantitative assessment of steatosis, hepatocellular ballooning, and fibrosis in metabolic dysfunction-associated steatohepatitis
Virchows Archiv · 07.11.2026
DOI: 10.1007/s00428-026-04771-0 ↗Kurz gefasst
Fachartikel in der Zeitschrift Virchows Archiv, online erschienen am 07.11.2026. Verfasst von Byoung Uk Park und 19 weiteren Autor:innen. Beteiligte Einrichtung: Department of Laboratory Medicine and Pathology, University of Minnesota, Minneapolis, Minnesota, USA. park3145@umn.edu. Die Arbeit nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Aims
This study aimed to develop and evaluate Tri-MASH, a trichrome-based deep learning model for quantitative assessment of steatosis, classic hepatocellular ballooning, and fibrosis, and to explore their spatial relationships. Inflammatory-cell density was evaluated as an exploratory measure. Methods and results: Tri-MASH was trained using 161 training specimens and internally evaluated on 30 randomly selected, nonoverlapping specimens. External evaluation comprised 39 biopsies independently scored by four pathologists and 524 baseline and week-72 biopsies from 262 clinical-trial participants. Internal-validation F1 scores ranged from 78.9% for inflammatory-cell detection to 99.2% for sinusoidal fibrosis. In cohort 1, associations with median pathologist assessments were very strong for steatosis (Rs = 0.85) and fibrosis (Rs = 0.87–0.88) and moderate for ballooning (Rs = 0.55). In cohort 2, exploratory pooled analyses showed moderate associations for steatosis (Rs = 0.52), ballooning (Rs = 0.54), and total and portal-septal fibrosis (Rs = 0.50); confidence intervals and p values were unadjusted for repeated observations within participants. Inflammatory-cell density correlated poorly with lobular inflammation grade (Rs=-0.01 to 0.14). Continuous measurements characterized within-arm longitudinal variation, including variation despite unchanged fibrosis stage. In a selected 50-biopsy subset enriched for NAS ballooning grade 2, case-specific median distances to sinusoidal fibrosis were shorter for ballooned hepatocytes than for lipid droplets ( p < 0.0001).
Conclusions
Tri-MASH enables integrated quantitative and spatial analysis of steatosis, classic hepatocellular ballooning, and fibrosis on a single trichrome section. The current data support its use as an adjunctive quantitative research platform rather than a replacement for conventional H&E-based assessment or standalone clinical grading; inflammatory-cell output did not reproduce conventional lobular inflammation grading.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Department of Laboratory Medicine and Pathology, University of Minnesota, Minneapolis, Minnesota, USA. park3145@umn.edu.
- Department of Preventive, Occupational, and Aerospace Medicine, Mayo Clinic, Rochester, Minnesota, USA.
- Department of Laboratory Medicine and Pathology, Mayo Clinic, Rochester, Minnesota, USA.
- MySME, Birmingham, Alabama, USA.
- Department of Pathology, University of Virginia, Charlottesville, Virginia, USA.
- Aiforia Technologies, Rochester, Minnesota, USA.
Zitieren
APA
Park, B. U., Hassan, B., Nazli, S., Smith, L., Shabaan, A., Patil, A., Stetzik, L., Elhalaby, R., DaCosta, A., Sivasubramaniam, P., Jiezhen, T. L., Puri, G., Malhi, H., Gores, G., Goodman, Z. D., Salomao, M., Pai, R., Hartley, C. P., Graham, R. P., & Moreira, R. K. (2026). Trichrome-based deep learning for quantitative assessment of steatosis, hepatocellular ballooning, and fibrosis in metabolic dysfunction-associated steatohepatitis. Virchows Archiv. https://doi.org/10.1007/s00428-026-04771-0
Vancouver
Park BU, Hassan B, Nazli S, Smith L, Shabaan A, Patil A, et al. Trichrome-based deep learning for quantitative assessment of steatosis, hepatocellular ballooning, and fibrosis in metabolic dysfunction-associated steatohepatitis. Virchows Archiv. 2026. doi:10.1007/s00428-026-04771-0
BibTeX
@article{Park2026,
author = {Park, Byoung Uk and Hassan, Bashar and Nazli, Saadiya and Smith, Lindsey and Shabaan, Abdelrahman and Patil, Ameya and Stetzik, Lucas and Elhalaby, Rofyda and DaCosta, Adilson and Sivasubramaniam, Priya and Jiezhen, Tracy Loh and Puri, Gyanendra and Malhi, Harmeet and Gores, Gregory and Goodman, Zachary D and Salomao, Marcela and Pai, Rish and Hartley, Christopher P and Graham, Rondell P and Moreira, Roger K},
title = {Trichrome-based deep learning for quantitative assessment of steatosis, hepatocellular ballooning, and fibrosis in metabolic dysfunction-associated steatohepatitis},
journal = {Virchows Archiv},
year = {2026},
doi = {10.1007/s00428-026-04771-0}
}
EUVIMED