Deep learning‐based accurate diagnosis and quantitative evaluation of microvascular invasion in hepatocellular carcinoma on whole‐slide histopathology images

Xiuming Zhang, Xiaotian Yu, Wenjie Liang, Zhong-Liang Zhang, Shengxuming Zhang, Linjie Xu, Han Zhang, Zunlei Feng, Mingli Song, Jing Zhang, Shi FengCancer 2024 · DOI

TL;DR

This paper introduces MVI-AIDM, a deep learning model that automates the diagnosis of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) from whole-slide histopathology images. MVI-AIDM not only achieves high accuracy but also provides quantitative spatial information, significantly improving upon traditional manual methods.

What Problem Does This Paper Address?

Microvascular invasion (MVI) is a critical prognostic factor for hepatocellular carcinoma (HCC), strongly correlating with early recurrence and poorer survival after surgical resection. However, diagnosing MVI through traditional histopathology is a challenging task. It's subjective, time-consuming, and prone to inter-observer variability, especially for subtle cases of micro-MVI. Furthermore, traditional methods lack the ability to provide precise quantitative information about MVI, such as the number of invaded vessels or the spatial relationship to the tumor, which are crucial for personalized treatment and prognosis. This gap highlights the need for an automated, accurate, and quantitative diagnostic tool.

Methods

The authors developed a three-step deep learning model, MVI-AIDM, designed to mimic the manual diagnostic process of pathologists:

  1. Tumor Region Detection: This initial step uses a pathological classification framework with a noise-rectifying loss function to identify cancerous regions within whole-slide images (WSIs).
  2. Microvascular Segmentation: Once tumor regions are identified, the model employs a ResNet18 classification model and DeepLabv3 for semantic segmentation to locate and segment microvessels outside the tumor.
  3. MVI Cell Classification: This final step uses a weakly supervised method with a cell localization and classification branch to identify cancer cells within the segmented microvessels, thereby diagnosing MVI. This approach leverages correlation filtering to achieve cell segmentation without requiring explicit cell-level annotations.

The model's performance was evaluated using standard metrics like accuracy, precision, recall, F1 score, and AUC.

Datasets

  • Internal Dataset: 5517 WSIs from 753 HCC patients from the First Affiliated Hospital of Zhejiang University (FAHZJU). This dataset was split into a training set (70%) and a test set (30%).
  • External Validation Dataset: 358 WSIs of HCC from The Cancer Genome Atlas (TCGA) database. These slides were manually annotated for MVI presence.

Key Insights and Ideas

  • Mimicking Pathologists: The model's strength lies in its ability to simulate the step-by-step diagnostic process of pathologists, making its outputs more interpretable and potentially easier for clinicians to adopt.
  • Weakly Supervised Learning for Cell Classification: The innovative use of weakly supervised learning for MVI cell classification, relying on patch-level labels rather than precise cell annotations, is a significant advancement, reducing the annotation burden.
  • Quantitative Spatial Information: Beyond just diagnosis, MVI-AIDM provides crucial quantitative spatial information (e.g., distance to tumor, MVI area, cancer cell count within MVI), which is currently lacking in traditional pathology and offers deeper insights into prognosis.
  • Improved Detection Rate: The AI model significantly outperforms pathologists in detecting MVI, particularly in challenging cases, suggesting it can serve as a powerful auxiliary tool to reduce missed diagnoses.

Key Results

  • High Accuracy: MVI-AIDM achieved an accuracy of 94.25% on the independent external validation set (TCGA).
  • Improved Detection Rate: The MVI positive detection rate of MVI-AIDM (70.85%) was significantly higher than that of pathologists (64.13%, p<0.001) in the internal test set.
  • Efficiency: The model reduced diagnosis time from an average of 28.7 minutes per case for a senior pathologist to just 9.1 minutes from slide scanning to result output.
  • Tumor Region Detection: Achieved an AUC of 0.88 and 0.82 on the FAHZJU test set and TCGA cohort, respectively.
  • MVI Cell Classification: Achieved an AUC of 0.95 and 0.88 for MVI cell classification on the FAHZJU test set and TCGA cohort, respectively.

Limitations and Future Work

  • MVI Grading Debate: The authors acknowledge that there is ongoing international debate regarding MVI grading. They plan to further investigate MVI grading standards and factors associated with prognosis using the extracted MVI features.
  • Types of MVI: The study focused on general MVI detection. Future work will aim to differentiate and recognize the three types of MVIs in HCC: portal invasion, hepatic vein invasion, and hepatic artery invasion.
  • False Positives: The model may exhibit minor false positives due to artifacts like displaced cells, requiring further verification by pathologists.
  • Sample Size Impact: The authors plan to investigate the impact of varying sample sizes on the diagnostic performance of the model, as insufficient samples can lead to overfitting.
  • Data Availability: Patient data is not publicly available due to privacy concerns but can be accessed by researchers meeting specific criteria.

Summary updated

Abstract

We developed a deep learning diagnostic model, which performed well and improved the efficiency and accuracy of MVI diagnosis. The model provided spatial information of MVI that was essential to accurately predict HCC recurrence after surgery.
BibTeX
@inproceedings{zhang2024deep,
  author = {Xiuming Zhang and Xiaotian Yu and Wenjie Liang and Zhong-Liang Zhang and Shengxuming Zhang and Linjie Xu and Han Zhang and Zunlei Feng and Mingli Song and Jing Zhang and Shi Feng},
  title = {{Deep learning‐based accurate diagnosis and quantitative evaluation of microvascular invasion in hepatocellular carcinoma on whole‐slide histopathology images}},
  booktitle = {Cancer},
  year = {2024},
  doi = {10.1002/cam4.7104},
}
Figure