A Digital Pathology Resource for Liver Cancer Quantification with Datasets, Benchmarks, and Tools

Ying Xiao, Shimiao Tang, Xitong Ling, Weiming Chen, Jun Wang, Jiawen Li, Huaitian Yuan, Jianghui Yang, Bowen Li, Huan Li, Yiting Meng, Tian Guan, Yonghong He, Hongfang YinarXiv 2026 · arXiv

TL;DR

This paper introduces HepatoBench and HepatoQuant, an open-source dataset and end-to-end pipeline designed to automate the fine-grained quantification of liver cancer tissue components. By combining patch-level classification with whole-slide image (WSI) segmentation using state-of-the-art pathology foundation models, the framework shifts liver cancer pathology from subjective, qualitative analysis to objective, reproducible regional statistics.

What Problem Does This Paper Address?

Hepatocellular carcinoma (HCC) is a massive global health burden, and examining tissue slides under a microscope remains the gold standard for diagnosis. Pathologists don't just look for tumor cells; they evaluate the surrounding microenvironment—looking at fibrosis, inflammation, and necrosis—to estimate a patient's risk of cancer recurrence.

However, traditional pathological assessment relies heavily on manual, qualitative, or semi-quantitative judgments. This introduces high subjectivity and low reproducibility. While deep learning could help, most publicly available digital pathology datasets only provide labels at the whole-slide image (WSI) level. Fine-grained, well-annotated patch-level datasets for liver cancer are incredibly scarce, hindering the development of automated, reproducible quantitative tools.

Methods

The authors developed an end-to-end pipeline called HepatoQuant that elegantly splits the problem into two parallel tasks before recombining them for final analysis:

  1. Patch-Level Tissue Classification: WSIs are tiled into small patches. The authors leverage frozen, pre-trained pathology foundation models (PFMs)—specifically UNI, Virchow-V2, and Gigapath—as feature extractors. A lightweight linear-probe classification head is trained on top of these features to categorize each patch into one of seven distinct tissue types.
  2. WSI-Level Region Segmentation: In parallel, a segmentation model is fine-tuned using Low-Rank Adaptation (LoRA) on top of a pathology foundation model to delineate the broader tumor vs. non-tumor (normal) regions across the entire slide.

Finally, the pipeline integrates these two outputs. By mapping the fine-grained patch classifications back onto the broader segmentation mask, HepatoQuant calculates precise regional tissue composition statistics (e.g., the percentage of fibrosis specifically within the tumor region versus the normal region).

Figure 1. End-to-end quantitative analysis pipeline for hepatocellular carcinoma whole slide images.

Datasets

The authors introduce and open-source HepatoBench, a curated, seven-class patch-level histopathology dataset.

  • Size & Resolution: 140,000 annotated patches extracted at 20× magnification (150×150 pixels).
  • Categories: Tumor (TUM: 13,200), Fibrosis (FIB: 26,364), Inflammation (INF: 4,776), Necrosis (NEC: 22,023), Normal tissue (NOR: 13,200), Bile duct reaction (REA: 1,540), and Steatosis (STE: 6,198).
  • Quality Control: Rigorous filtering was applied to remove duplicates (via MD5 hashing), ambiguous morphologies, and artifacts like blur, wrinkles, or severe staining variations.
  • WSI Dataset: 111 fully annotated whole-slide images used to train and evaluate the regional tumor/non-tumor segmentation model.

Key Insights and Ideas

  • The Power of Linear Probing on PFMs: Instead of computationally expensive full-model fine-tuning, simply freezing massive pathology foundation models and training a single linear layer yields exceptional classification accuracy. This underscores the robust, generalized representations these foundation models inherently possess.
  • Regional Context is Crucial: Identifying a tissue component (like inflammation) is only half the battle. Its clinical significance changes depending on whether it is located inside the tumor microenvironment or within healthy tissue. HepatoQuant’s ability to provide region-aware quantification bridges the gap between raw deep learning outputs and actionable clinical insights.

Key Results

The authors benchmarked three prominent pathology foundation models on HepatoBench: UNI, Gigapath, and Virchow-V2.

  • Top Performers: Gigapath and Virchow-V2 noticeably outperformed UNI across all primary metrics. Both achieved a Balanced Accuracy (BACC) of 0.967 and an F1-Score of 0.969, compared to UNI’s BACC of 0.913 and F1-Score of 0.920.
  • Classification Reliability: Area Under the ROC Curve (AUC) scores reached near-perfection, with Gigapath hitting 0.999 and Virchow-V2 achieving a flawless 1.000.
  • Confusion matrices revealed that Gigapath and Virchow-V2 maintained exceptionally clean diagonal performance, showing minimal misclassification even among morphologically challenging categories like inflammation (INF) and normal tissue (NOR).
Figure 2. Benchmark results of pathology foundation models on HepatoBench.

Limitations and Future Work

While HepatoBench and HepatoQuant offer a robust foundation for automated liver pathology, the authors explicitly note that this dataset is intended for research and benchmarking, not as a standalone tool for direct clinical diagnosis.

Future work will likely focus on validating the pipeline across multi-center datasets to ensure robustness against diverse staining protocols and scanning hardware. Additionally, integrating these automated spatial tissue statistics with downstream clinical outcomes and multimodal genomic data represents a promising frontier for personalized cancer treatment.

Summary updated

Abstract

Liver cancer, especially hepatocellular carcinoma (HCC), imposes a substantial global disease burden. Accurate diagnosis and prognostic assessment directly influence treatment selection and patient survival, and pathological examination remains the gold standard for liver cancer diagnosis. Identifying diverse tissue components and pathological subtypes on histopathology slides is crucial for estimating postoperative recurrence risk and overall prognosis. However, most publicly available resources are still provided at the whole-slide image (WSI) level, and well-annotated datasets for fine-grained tissue component identification in liver cancer are scarce, which hinders reproducible model development and the deployment of quantitative analysis tools. To address this gap, we release HepatoBench, a patch-level image database for liver cancer with annotations for seven key tissue categories. Based on HepatoBench, we train and open-source a deep learning classification model as a tissue recognition tool. Furthermore, we train a WSI-level tumor/non-tumor segmentation model to automatically localize lesion regions across entire slides. By integrating the patch-level tissue classifier with the WSI-level segmentation model, we build HepatoQuant, an end-to-end, disease-specific regional quantification tool for liver cancer, enabling a unified workflow from WSIs to tissue composition parsing and quantitative statistics. We also open-source HepatoBench, the benchmarking protocol, and supporting tools, providing a solid foundation for automated regional quantification and fair method comparison in liver cancer pathology.
BibTeX
@misc{xiao2026digital,
  author = {Ying Xiao and Shimiao Tang and Xitong Ling and Weiming Chen and Jun Wang and Jiawen Li and Huaitian Yuan and Jianghui Yang and Bowen Li and Huan Li and Yiting Meng and Tian Guan and Yonghong He and Hongfang Yin},
  title = {{A Digital Pathology Resource for Liver Cancer Quantification with Datasets, Benchmarks, and Tools}},
  year = {2026},
  eprint = {2604.22858},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.22858},
}
Figure