Personalized intrahepatic cholangiocarcinoma prognosis prediction using radiomics: Application and development trend

Pengyu Chen, Zhenwei Yang, Haofeng Zhang, Guan Huang, Qingshan Li, Peigang Ning, Haibo YuFrontiers in Oncology 2023 · DOI

TL;DR

This review explores how radiomics, the extraction of quantitative features from medical images, is being used to predict outcomes for intrahepatic cholangiocarcinoma (ICC), a challenging liver cancer. It highlights radiomics' potential to personalize treatment by predicting lymph node metastasis, microvascular invasion, and early recurrence, offering a more nuanced approach than traditional staging systems.

What Problem Does This Paper Address?

Intrahepatic cholangiocarcinoma (ICC) is a serious liver cancer with a poor prognosis, often diagnosed late. Current staging systems, like the AJCC TNM system, are not always accurate for predicting outcomes, especially after surgery. This makes it difficult for clinicians to tailor treatment strategies and predict how patients will fare. The paper addresses the need for more precise, personalized prognostic tools that can be used before treatment, leveraging the rich information hidden within medical images.

Methods

The paper focuses on radiomics, a technique that goes beyond visual interpretation of medical images (like CT and MRI) to extract hundreds or thousands of quantitative features. These features describe characteristics such as shape, texture, and intensity patterns within the tumor.

The core methodology involves:

  • Image Segmentation: Identifying and isolating the region of interest (ROI), typically the tumor, within the medical images.
  • Feature Extraction: Quantifying various characteristics of the segmented ROI using specialized software.
  • Feature Selection: Employing statistical methods (e.g., LASSO, mRMR) to identify the most relevant features from the large set extracted.
  • Model Construction: Using machine learning algorithms (e.g., Logistic Regression, SVM, Random Forest, CNNs) to build predictive models based on the selected features. These models can then predict specific clinical outcomes.

The paper also discusses the emerging role of deep learning (DL), particularly Convolutional Neural Networks (CNNs), which can automate feature extraction and even end-to-end modeling, potentially offering more powerful and less interpretable models.

Datasets

The paper reviews numerous studies that utilize various datasets. These datasets are primarily derived from clinical imaging of patients with intrahepatic cholangiocarcinoma. The studies often involve:

  • Single-center cohorts: Data collected from a single hospital.
  • Multi-center cohorts: Data pooled from multiple institutions, offering broader applicability.
  • Different imaging modalities: CT, MRI, Ultrasound (US), and PET/CT scans are all discussed.

The paper does not introduce a new benchmark dataset but rather synthesizes findings from existing research, highlighting the diversity of data sources and imaging techniques employed.

Key Insights and Ideas

  • Radiomics as a "Virtual Biopsy": Radiomics offers a non-invasive way to characterize tumor biology and heterogeneity, providing insights similar to a biopsy without the need for invasive procedures.
  • Personalized Prognosis: Radiomics models can predict specific outcomes like lymph node metastasis, microvascular invasion, and early recurrence, allowing for more tailored treatment plans than broad staging systems.
  • Deep Learning's Potential: DL, especially CNNs, shows promise in automating feature extraction and building more powerful predictive models, though interpretability remains a challenge.
  • Multimodal Approaches: Combining radiomics with clinical data (e.g., CA19-9 levels) or even other imaging modalities often leads to improved predictive performance.
  • Beyond the Tumor: Features extracted from the peritumoral region (area around the tumor) can also be valuable for prediction.

Key Results

The paper synthesizes results from numerous studies, demonstrating the utility of radiomics in predicting various aspects of ICC prognosis:

  • Lymph Node Metastasis (LNM): Radiomics models have shown strong performance in predicting LNM, with AUC values often exceeding 0.80, and in some cases reaching over 0.90.
  • Microvascular Invasion (MVI): Models have achieved AUCs ranging from 0.70 to over 0.95 for predicting MVI, with some studies suggesting PET/CT-based radiomics performs better than CT alone.
  • Early Recurrence (ER): Radiomics models have demonstrated high predictive accuracy for early recurrence, with AUCs frequently above 0.85 and some reaching 0.95 or higher.
  • Overall Survival (OS) / Progression-Free Survival (PFS): Radiomics models, often combined with clinical data, have shown improved prognostic prediction compared to traditional staging systems, with C-indices often in the range of 0.75-0.85.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Standardization Challenges: A lack of standardized imaging protocols and feature extraction methods across different centers and instruments can compromise model reproducibility and accuracy.
  • Small Sample Sizes and Single-Center Studies: Many studies rely on limited datasets from single institutions, increasing the risk of overfitting and limiting generalizability.
  • Interpretability of DL Models: While DL models show high performance, their "black box" nature makes it difficult for clinicians to understand the basis of their predictions, hindering trust and adoption.
  • Need for External Validation: More prospective, multi-center validation studies are needed to confirm the robustness and clinical utility of these radiomics models.
  • Plurality of Modalities: Expanding research to include more diverse imaging modalities (PET, US) and exploring their integration is crucial.
  • Radiogenomics and Immunotherapy: Future work should focus on integrating radiomics with genomic data (radiogenomics) and exploring its role in guiding immunotherapy decisions, offering a more comprehensive approach to personalized medicine.
  • Interpretability Research: Developing methods to improve the interpretability of radiomics and DL models is essential for clinical translation.

Summary updated

Abstract

Radiomics was proposed by Lambin et al. in 2012 and since then there has been an explosion of related research. There has been significant interest in developing high-throughput methods that can automatically extract a large number of quantitative image features from medical images for better diagnostic or predictive performance. There have also been numerous radiomics investigations on intrahepatic cholangiocarcinoma in recent years, but no pertinent review materials are readily available. This work discusses the modeling analysis of radiomics for the prediction of lymph node metastasis, microvascular invasion, and early recurrence of intrahepatic cholangiocarcinoma, as well as the use of deep learning. This paper briefly reviews the current status of radiomics research to provide a reference for future studies.
BibTeX
@inproceedings{chen2023personalized,
  author = {Pengyu Chen and Zhenwei Yang and Haofeng Zhang and Guan Huang and Qingshan Li and Peigang Ning and Haibo Yu},
  title = {{Personalized intrahepatic cholangiocarcinoma prognosis prediction using radiomics: Application and development trend}},
  booktitle = {Frontiers in Oncology},
  year = {2023},
  doi = {10.3389/fonc.2023.1133867},
}
Figure