Abstract
Background & Aims:
Hepatocellular carcinoma (HCC) remains among the most lethal cancers worldwide, yet few tools exist to characterize HCC biology and stratify patient risk at the point of care. We developed TESERA (Tissue Encoded Subtype Elucidation and Risk Assessment), a computational pathology framework that derives both unsupervised subtypes and a prognostic index directly from hematoxylin and eosin (H&E) whole-slide images (WSIs).
Approach & Results:
Applied to The Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) cohort, unsupervised clustering of TESERA embeddings grouped tumors into Tissue Encoded Subtypes (TES1 and TES2) associated with different overall survival (OS; TES2 HR=4.20). TES1, the favorable-prognosis subtype, was enriched for metabolic pathways and
CTNNB1
mutations, while TES2 was enriched for proliferative pathways and
TP53
mutations. To predict patient-level survival outcomes, we trained a supervised model, the TESERA morphologic index (TESERA-MI), on TESERA embeddings, which predicted both OS (HR=3.17) and disease-free survival (HR=2.19) in TCGA-LIHC. TESERA-MI achieved the highest discrimination in a head-to-head comparison with clinical variables and molecular markers (C-index, 0.73; 1-, 3-, and 5-year AUROCs, 0.82, 0.80, and 0.79), remained independently prognostic after multivariable adjustment (HR=2.97), and stratified OS in an independent in-house cohort (HR=2.16).
Conclusions:
Released as a publicly accessible pipeline applicable to H&E WSIs, TESERA unifies morphologically derived metabolism-versus-proliferation subtypes with an externally validated, prognostic morphologic index competitive with existing clinical and molecular markers. These findings position routine histology as a low-cost, widely available substrate for biology-aware risk stratification in HCC.