Defining Pathomics Pathomics is an emerging field that applies AI algorithms to extract high-throughput quantitative biological features from whole-slide histopathological images (WSIs) - capturing patterns invisible to the human eye that reflect tumor biology, heterogeneity, and clinical outcomes.
The HCC Problem HCC is the sixth most common cancer and the third leading cause of cancer mortality globally, with over 860,000 new cases and 750,000 deaths in 2022. The 70% five-year recurrence rate after surgery and the 67% of patients presenting at intermediate or advanced stage underscore the urgent need for better prognostic tools.
What Pathomics Offers Unlike radiomics (imaging features) or genomics (DNA mutations), pathomics operates directly on H&E-stained tissue - the universally available material in HCC surgery and biopsy. By quantifying features like cellular morphology, nuclear architecture, tissue architecture, and immune infiltration from WSIs, pathomics can predict outcomes that current staging systems miss.
Review Scope This narrative review from Nanjing University Medical School synthesizes current pathomics applications in HCC across: (1) basic pathological diagnosis tasks (tumor identification, histological classification, grading, MVI detection), (2) advanced prognostic prediction (genetic markers, immune signatures, survival, treatment response), and (3) the emerging field of pathological foundation models.
Standard Pipeline Steps The pathomics workflow involves: WSI acquisition and preprocessing (digitization, color normalization), patch segmentation and expert annotation of regions of interest, feature extraction using pretrained CNN or ViT models, attention-based multiple instance learning (MIL) for slide-level aggregation, and model evaluation with AUROC, F1, and decision curve analysis.
CNN vs. Vision Transformer CNN backbones (ResNet50, EfficientNet-B0) remain dominant for patch-level feature extraction at 224-512 pixel tiles. Vision transformers (ViT-base, 12 layers) are increasingly used to capture long-range tissue structure relationships across distant image regions - an advantage for tumors with heterogeneous spatial patterns like HCC.
Attention-Based MIL Because WSIs contain millions of pixels, direct image analysis is computationally infeasible. Multiple instance learning aggregates patch-level features into slide-level predictions using a 128-256 dimensional attention head that scores each patch's diagnostic or prognostic relevance. The top 10-20 scoring patches drive the final prediction, downweighting uninformative areas.
Multi-Modal Integration The frontier of pathomics integrates histological features with radiomics (from CT or MRI), genomics, transcriptomics, and clinical records into multi-modal AI frameworks. Bi-modal pathomics-radiomics models and tri-modal models incorporating clinical variables consistently outperform single-modality approaches in HCC prognosis prediction.
Tumor Identification AI pathomics reliably distinguishes HCC from non-tumor liver tissue with AUROC values approaching 1.000. One study using a global-label DL framework achieved precision of 1.0 in test sets, while another demonstrated AUROC 0.998-1.000 for tumor identification. These results establish AI as a reliable quality control tool for ensuring only tumor regions are analyzed in downstream tasks.
Histological Classification Multiple studies distinguish HCC from cholangiocarcinoma (CCA) and other liver cancer subtypes. Cheng et al. developed HnAIM for seven hepatocellular nodular lesion categories (AUROC 0.935), while Jang et al. distinguished HCC, CCA, and metastatic colorectal cancer with AUROC above 0.95 internally (though dropping to 0.745 on external validation). Calderaro et al. reclassified the rare combined HCC-CCA subtype with AUROC 0.99.
Pathological Grading HCC differentiation grade (Edmondson-Steiner classification) predicts prognosis but requires expert pathology. SENet, the best-performing AI model in a comparative study of five architectures, achieved 95.27% accuracy for three-tier grade classification. Another study using multiphoton microscopy-based pathomics achieved over 90% accuracy for label-free automated grading.
MVI Detection Breakthrough Multiple pathomics models now detect MVI from WSIs. Chen et al.'s weakly supervised model achieved AUROC 0.904/0.871 internally and externally. Zhang et al.'s MVI-AIDM (the study covered in LIV133) achieved 97.49%/94.25% accuracy and detected significantly more MVI-positive cases than pathologists - establishing pathomics as the new benchmark for MVI diagnosis.
Genetic Marker Prediction Pathomics models predict high-risk gene expression from histological images without sequencing. Models for EZH2 (AUROC 0.742 testing), TNFRSF4 (AUROC 0.723), and ANGPT2 (AUROC 0.719 external validation) demonstrate that tissue morphology encodes molecular information. High pathomics scores correlated with worse overall survival and distinct immune cell infiltration patterns.
Survival Prediction Saillard et al.'s pioneering study achieved concordance indices of 0.78 and 0.75 for overall survival prediction after HCC resection - surpassing clinical staging systems. The identified high-risk histological features (macrotrabecular-massive subtype, vascular space, cellular atypia, nuclear pleomorphism) align with known pathological risk markers, validating the biological plausibility of the AI predictions.
Recurrence Prediction Multiple models predict recurrence-free survival, with Qu et al. achieving AUROC 0.857/0.852 for 3-year RFS and Schmauch et al.'s combined clinico-pathomics model achieving C-index 0.77 for disease-specific survival. A consistent finding is that high-recurrence-risk tumors show stroma, high cellular atypia, deeply stained nuclei, and lack of immune infiltration.
Treatment Response Prediction ABRS-P predicted atezolizumab-bevacizumab response with significant PFS differences in an external immunotherapy cohort (median PFS 12 vs. 7 months, p=0.014). Li et al.'s CNN-SASM predicted sorafenib benefit with AUROC 0.813/0.707 internally and externally. These results position pathomics as a potential companion diagnostic for HCC systemic therapy selection.
What Are Foundation Models? Pathological foundation models (PFMs) are large neural networks pre-trained on millions of WSI tiles using self-supervised learning - learning general tissue representations without task-specific labels. Models like UNI, Prov GigaPath, CONCH, Virchow, and CHIEF have been published in Nature-tier journals and represent a paradigm shift toward general-purpose pathology AI.
How PFMs Work for HCC PFMs encode slides into rich feature representations that can be fine-tuned for specific HCC tasks (grading, MVI, survival) with small labeled datasets - dramatically reducing the annotation burden that has limited pathomics research. A PFM can be adapted to a new HCC task with 50-200 labeled cases rather than the hundreds or thousands required when training from scratch.
AI Agents: The Future Vision PFMs are evolving toward 'AI agents' - systems that understand clinical queries, select appropriate sub-models, and generate analytical insights in natural language. Interactive models like PathChat and PathAsst can already respond to pathologist questions about tissue morphology. The goal is a generalist AI that can autonomously perform the full pathology diagnostic pipeline.
HCC-Specific Foundation Models Needed While general pathology foundation models are powerful, HCC-specific vertical models that incorporate hepatocyte morphology, fibrosis patterns, and immune microenvironment features unique to liver cancer are needed. Models like DINOPath for gastrointestinal cancers and ArteraAI for prostate cancer demonstrate the value of disease-specific foundation model development.
Data Heterogeneity Variability in scanner hardware, staining protocols, color normalization standards, and annotation criteria across institutions creates domain shift that causes models trained at one center to underperform at another. Solutions include international WSI annotation consortia, standardized staining/scanning guidelines, and federated learning frameworks that train collaboratively without sharing raw patient data.
Model Interpretability Deep learning's 'black-box' nature limits clinician trust. While Grad-CAM, SHAP, and concept activation vectors can visualize attention regions, these tools do not quantitatively link image features to molecular pathways or outcome mechanisms. A multi-scale interpretability framework connecting morphological findings to biological mechanisms is urgently needed.
Ethical and Regulatory Barriers WSI data sharing raises privacy concerns, and AI diagnostic tools require regulatory clearance as software as medical devices. Regulatory frameworks differ substantially between FDA (US), CE marking (Europe), and Asian regulatory bodies - requiring jurisdiction-specific validation studies that are expensive and time-consuming.
The Path Forward The review recommends: prospective multi-center studies (to overcome retrospective single-center bias), multi-modal data integration (combining pathomics with radiomics, genomics, and clinical records), establishment of industry-standard guidelines and compliant platform infrastructures, and coordination between researchers, hospitals, and industry to build and share external validation cohorts.