Multiparametric MRI‐based model for prediction of local progression of hepatocellular carcinoma after thermal ablation

Chao Chen, Qiuying Han, He Ren, Siyi Wu, Yangyang Li, Jiandong Guo, Xinghai Li, Xiang Liu, Chengzhi Li, Yunfei TianCancer 2023 · DOI

TL;DR

This paper introduces a novel deep learning radiomics model (DLRMM) that integrates multiparametric MRI data from both the tumor and ablation zone to predict local tumor progression (LTP) in hepatocellular carcinoma (HCC) patients after thermal ablation. The DLRMM significantly outperforms traditional clinical models, offering a more accurate and personalized approach to risk stratification and treatment planning.

What Problem Does This Paper Address?

Hepatocellular carcinoma (HCC) is a major global health concern, and thermal ablation (TA) is a common treatment for early-stage HCC. However, a significant challenge is the high rate of local tumor progression (LTP) after TA, which can lead to poor patient outcomes. Predicting LTP accurately is crucial for tailoring treatment strategies and improving patient survival. Current methods for assessing risk factors like ablative margin are often difficult to measure accurately due to factors like liver deformation. This paper aims to bridge this gap by developing a more robust and automated predictive model using advanced imaging techniques.

Methods

The researchers developed a Deep Learning Radiomics (DLR) model, termed DLRMM, that leverages multiparametric MRI data. The core of their approach involves:

  • Data Acquisition: Multiparametric MRI data (T1WI+C, T2WI, and DWI) were collected from patients before and after thermal ablation.
  • Feature Extraction:
    • Radiomics: Standard radiomics features were extracted from regions of interest (ROIs) encompassing the tumor and ablation zone.
  • Deep Learning (DL): A ResNet18 model, pre-trained on ImageNet and fine-tuned on MRI data, was used to extract DL-based radiomics features. This approach allows the model to learn complex, hierarchical features directly from the images, potentially capturing subtle patterns missed by traditional radiomics.
  • Model Development:
    • DLR Model: A DLR model was constructed using the extracted DL radiomics features.
  • Clinical Model: A model was built using established clinical risk factors (e.g., tumor size, number, ablative margin, and location adjacent to major vessels).
  • Integrative Model (DLR-Clinical): The DLR model was further combined with significant clinical variables to create a comprehensive predictive tool.
  • Model Evaluation: The performance of these models was assessed using Area Under the Receiver Operating Characteristic Curve (AUC) and compared using the DeLong test. Other metrics like accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were also reported.

Datasets

The study utilized a multi-institutional dataset comprising:

  • Primary Cohort (PC): 189 patients.
  • External Test Cohort 1 (ETC1): 135 patients.
  • External Test Cohort 2 (ETC2): 93 patients.

This provided a robust framework for both training and external validation of the developed models. No new benchmark datasets were introduced; rather, existing patient data from multiple centers were pooled and split.

Key Insights and Ideas

  • DL Radiomics Captures More Than Traditional Radiomics: The deep learning component of the DLRMM was able to extract features that significantly improved predictive performance, suggesting it can capture more nuanced information about tumor characteristics and the post-ablation environment than traditional radiomics alone.
  • Tumor Margin and Ablation Zone are Crucial: Including features from both the tumor and the ablation zone was key. This highlights that not only the initial tumor characteristics but also the success and extent of the ablation itself are critical for predicting future progression.
  • Integration is Key for Optimal Performance: While the DLR model showed strong performance, combining it with established clinical risk factors (like tumor number, safe margin, and proximity to major vessels) yielded the best predictive accuracy. This suggests that a hybrid approach, leveraging both advanced imaging and clinical knowledge, is most effective.
  • Personalized Risk Stratification: The DLR-clinical model effectively stratified patients into low- and high-risk groups for LTP, which directly correlated with overall survival. This allows for personalized risk assessment and potentially tailored follow-up or adjuvant treatment strategies.

Key Results

  • The DLR model achieved AUCs of 0.864 (PC), 0.843 (ETC1), and 0.858 (ETC2), significantly outperforming the clinical model (AUCs around 0.801-0.804).
  • The integrated DLR-clinical model further improved performance, with AUCs of 0.871 (PC), 0.869 (ETC1), and 0.869 (ETC2).
  • The DLR-clinical model demonstrated a significant difference in overall survival between low-risk and high-risk groups across all cohorts, indicating its prognostic value.

Limitations and Future Work

The authors acknowledge several limitations:

  • Sample Size: The relatively small sample size, particularly in the primary cohort, might lead to mild overfitting. Future work should involve collecting more data to confirm the model's robustness.
  • Variability in MRI Machines and IPTA Techniques: Differences in MRI scanner parameters and TA techniques (MWA vs. RFA) across participating hospitals could influence model performance.
  • Exclusion of Molecular Biomarkers: The study did not incorporate molecular biomarkers (e.g., NLR, MMP7, MUC1) or genomic information, which could potentially enhance the model's accuracy.

Future directions include:

  • Collecting larger, multi-center datasets to validate and refine the DLR model.
  • Investigating the inclusion of molecular and genomic data for a more comprehensive predictive framework.
  • Designing prospective clinical trials to further establish the clinical utility and reliability of this DLR-based tool for guiding treatment decisions and surveillance strategies in HCC patients undergoing thermal ablation.

Summary updated

Abstract

The DLRMM model is essential to identify LTP risk of HCC patients who underwent TA and may potentially benefit personalized decision-making.
BibTeX
@inproceedings{chen2023multiparametric,
  author = {Chao Chen and Qiuying Han and He Ren and Siyi Wu and Yangyang Li and Jiandong Guo and Xinghai Li and Xiang Liu and Chengzhi Li and Yunfei Tian},
  title = {{Multiparametric MRI‐based model for prediction of local progression of hepatocellular carcinoma after thermal ablation}},
  booktitle = {Cancer},
  year = {2023},
  doi = {10.1002/cam4.6277},
}
Figure