Artificial intelligence in liver diseases: Improving diagnostics, prognostics and response prediction

David Nam, Julius Chapiro, Valérie Paradis, Tobias Paul Seraphin, Jakob Nikolas KatherJHEP 2022 · DOI

TL;DR

Artificial intelligence (AI), particularly deep learning, shows immense promise for revolutionizing liver disease diagnosis, prognosis, and treatment response prediction by extracting valuable insights from complex histopathology and radiology data. However, significant challenges remain in standardizing data, validating models, and integrating these AI tools into routine clinical practice.

What Problem Does This Paper Address?

Hepatology, the study of liver diseases, is a complex field where clinicians rely on a vast array of data modalities—from patient history and lab tests to imaging and biopsies—to make critical decisions. Interpreting this data is challenging, leading to diagnostic ambiguities. While AI, especially deep learning (DL), has advanced significantly in extracting clinically relevant information from complex datasets, its application in hepatology lags behind other medical fields. This paper addresses this gap by summarizing the current state of AI in hepatology, focusing on histopathology and radiology, and outlining a roadmap for its future development and clinical implementation.

Methods

The paper discusses two main categories of AI techniques used in hepatology:

  • Classical Machine Learning (ML) / Radiomics: This approach relies on "handcrafted features" defined by human investigators. For example, in radiomics, quantitative visual features from images (like tumor size or intensity) are extracted and fed into ML algorithms (e.g., random forests) for classification.
  • Deep Learning (DL): DL methods, primarily Convolutional Neural Networks (CNNs), automatically learn features directly from complex data like images. Unlike classical ML, DL models don't require pre-defined features, making them more flexible and powerful for tasks like image processing. The paper notes that DL is increasingly dominating AI research in hepatology due to its superior performance.

The paper also touches upon multimodal learning, where AI integrates data from various sources (e.g., images, clinical data, genetic data) to improve decision-making, a crucial aspect for the complex nature of hepatology.

Datasets

The paper does not explicitly list specific datasets used for training or evaluation in a dedicated section. However, it mentions:

  • MEDLINE database: Used for a quantitative survey of ML/DL research in hepatology.
  • Whole slide images (WSIs): Commonly used for histopathology studies, particularly from patients with non-alcoholic fatty liver disease (NAFLD), non-alcoholic steatohepatitis (NASH), or hepatocellular carcinoma (HCC).
  • CT scans, MRI scans, and H&E-stained histopathology slides: These are the primary imaging modalities analyzed.
  • The Cancer Genome Atlas (TCGA) database: Mentioned as a resource for early DL studies in HCC, though potential biases within it are noted.
  • Liver Tumor Segmentation Benchmark (LiTS): A public dataset for liver CTs used for evaluating segmentation algorithms.

No newly introduced benchmarks are highlighted.

Key Insights and Ideas

  • AI's potential to democratize expertise: AI can automate the interpretation of complex data, potentially making expert-level diagnostic capabilities more accessible.
  • Histopathology and radiology are rich data sources: Both image types contain a wealth of information that AI can leverage for diagnosis, prognosis, and predicting treatment response.
  • DL is the current frontier: Deep learning, with its ability to automatically learn features, is rapidly becoming the dominant approach in AI for hepatology, outperforming traditional ML methods.
  • The "black box" problem persists: While AI can achieve high accuracy, understanding why a model makes a certain prediction (explainability) is crucial for clinical trust and adoption.
  • Data standardization is a major hurdle: The lack of universal standards for data formatting, image compression, and metadata storage in histopathology hinders the development and deployment of AI tools.
  • Clinical implementation lags research: There's a significant gap between the growing number of AI research studies and the limited number of clinically approved AI applications in hepatology.

Key Results

The paper summarizes the state of the art through various studies, highlighting:

  • High accuracy in specific tasks: Studies report high AUCs (Area Under the Curve) for tasks like classifying liver masses (median accuracy of 0.84), differentiating HCC from benign lesions (AUC of 0.92), and staging liver fibrosis (AUCs of at least 0.95).
  • Improved diagnostic capabilities: AI models have shown the ability to segment organs and lesions with high accuracy, grade HCC nuclei, and identify tumor tissue.
  • Prognostic power: DL models have demonstrated the ability to predict disease progression, recurrence after resection, and survival in patients with HCC, sometimes outperforming traditional clinical and pathological features.
  • Potential to replace biopsy: CNN classification tools show promise in grading NAFLD and liver fibrosis, potentially reducing the need for invasive liver biopsies.
  • Growing research trend: The number of ML/DL studies in hepatology has grown exponentially in recent years, with a notable increase in DL studies compared to handcrafted feature-based studies.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • Lack of large-scale clinical validation: Many studies are proof-of-concept and require further large-scale clinical validation and regulatory approval.
  • Data standardization: A critical need exists for universal standards in data formatting, image compression, and metadata storage, especially in histopathology.
  • Bias in datasets: Ensuring epidemiological diversity in training datasets is crucial to avoid biases and ensure generalizability across different populations.
  • Explainability and interpretability: Developing methods to understand AI model predictions is essential for clinical trust and adoption.
  • Infrastructure and training: Building healthcare infrastructure and training medical personnel in "digital literacy" are vital for AI deployment.
  • Multimodal integration: Further research is needed to effectively integrate diverse data types (clinical, imaging, genetic) for more comprehensive decision-making.
  • Standardized software suites: The authors envision a future with standardized, potentially open-source, software suites that incorporate various AI plug-ins to facilitate adoption and reduce costs.
  • Proactive bias mitigation: AI algorithms should incorporate continuous learning to monitor for and adapt to biases in real-time.
  • Prospective testing: AI algorithms need to be tested on prospectively collected data to assess their robustness.

Summary updated

Abstract

Clinical routine in hepatology involves the diagnosis and treatment of a wide spectrum of metabolic, infectious, autoimmune and neoplastic diseases. Clinicians integrate qualitative and quantitative information from multiple data sources to make a diagnosis, prognosticate the disease course, and recommend a treatment. In the last 5 years, advances in artificial intelligence (AI), particularly in deep learning, have made it possible to extract clinically relevant information from complex and diverse clinical datasets. In particular, histopathology and radiology image data contain diagnostic, prognostic and predictive information which AI can extract. Ultimately, such AI systems could be implemented in clinical routine as decision support tools. However, in the context of hepatology, this requires further large-scale clinical validation and regulatory approval. Herein, we summarise the state of the art in AI in hepatology with a particular focus on histopathology and radiology data. We present a roadmap for the further development of novel biomarkers in hepatology and outline critical obstacles which need to be overcome.
BibTeX
@inproceedings{nam2022artificial,
  author = {David Nam and Julius Chapiro and Valérie Paradis and Tobias Paul Seraphin and Jakob Nikolas Kather},
  title = {{Artificial intelligence in liver diseases: Improving diagnostics, prognostics and response prediction}},
  booktitle = {JHEP},
  year = {2022},
  doi = {10.1016/j.jhepr.2022.100443},
}
Figure