Answers are grounded in this paper and cite their sources.
Abstract Background & aims Advances in digital pathology and artificial intelligence (AI) are driving progress toward personalized clinical management. In hepatocellular carcinoma (HCC), AI-based models using digitized hematoxylin and eosin (H&E) slides can be a robust tool to predict outcome-related molecular profiles and presence of microvascular invasion (mVI), which can be useful in the clinical setting. Methods A transformer-based deep-learning (DL) model was deployed using digitized H&E slides from 431 resected HCC patients (training cohort). 5-fold cross-validation was applied, and the model was tested on two external cohorts: TCGA-LIHC (n=363) and advanced-stage HCC cohort (n=64). Results The DL model effectively predicted outcome-related molecular profiles, distinguishing poor-prognosis (S1/S2, proliferation) from good-prognosis (S3, non-proliferation) subclasses. In internal cross-validation, mean area under the curves (AUC) were 0.75 for proliferation and 0.79 for non-proliferation subclasses. This performance was reproduced in the TCGA test set with AUCs ranging from 0.72-0.80 and in advanced HCC from 0.76-0.81. In the test sets, AI-predicted non-proliferation subclass was associated with a longer median OS compared to proliferation subclass (5.8 vs 3.5 years in TCGA; p=0.02). For mVI prediction, the DL model achieved a mean AUC of 0.70 in the internal cross-validation and 0.62 in the TCGA. AI-predicted mVI was associated with shorter OS (4.9 vs 7.6 years for non-mVI; p=0.003) and an immunosuppressive microenvironment (p=0.002). Conclusions Our H&E-based AI model enables accurate prediction of outcome-related molecular subtypes of poor prognosis and presence of mVI, offering a scalable and accessible tool to extract clinically relevant features from routine histology. Impact and implications Outcome-related molecular profiles and the presence of microvascular invasion (mVI) are critical determinants of prognosis and treatment decisions in hepatocellular carcinoma (HCC). This study presents an artificial intelligence (AI)-based method that analyzes routine H&E-stained slides and accurately predicts: (a) biologically relevant HCC molecular subtypes associated with patient outcomes, and (b) the presence of mVI, a well-established predictor of poor outcomes and risk of recurrence, that currently requires meticulous pathological assessment of multiple H&E slides. These AI tools can offer a scalable method to support personalized treatment decisions, such as transplant eligibility, trial enrollment, or neo/adjuvant therapy planning, and may improve clinical management of HCC. Our findings lay the groundwork for incorporating AI-assisted pathology into future prospective studies aimed at improving HCC clinical management.
No takes yet. Share an insight, caveat, or question.
Seraphin et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: