Multiparametric MR radiomics in brain glioma: models comparation to predict biomarker status

Jinlong He, Jialiang Ren, Guangming Niu, Aishi Liu, Qiong Wu, Shenghui Xie, Xueying Ma, Bo Li, Peng Wang, Jing Shen, Jianlin Wu, Yang GaoBMC Medical Imaging 2022 · DOI

TL;DR

This study explores using multi-sequence MRI radiomics to predict key genetic biomarkers (IDH, MGMT, TERT, 1p/19q) in brain gliomas. They found that combining radiomics with clinical data significantly improves prediction accuracy, especially for IDH and MGMT status, offering a promising non-invasive approach for personalized treatment.

What Problem Does This Paper Address?

Gliomas, the most common primary brain tumors, exhibit significant variability in their behavior, prognosis, and response to treatment, largely influenced by their underlying genetic makeup. Identifying these genetic biomarkers (like IDH mutations, MGMT promoter methylation, TERT promoter mutations, and 1p/19q co-deletion) is crucial for guiding treatment strategies and predicting patient outcomes. However, obtaining this information typically requires invasive tissue biopsies, which are not always feasible or desirable, especially pre-operatively. While imaging features can offer clues, a comprehensive and non-invasive method to predict multiple key biomarkers simultaneously has been lacking. This paper aims to bridge this gap by investigating the power of multiparametric MRI radiomics, combined with clinical data, to accurately predict these critical glioma biomarkers.

Methods

The researchers employed a retrospective analysis of 81 glioma patients. They extracted radiomic features from five different MRI sequences: T2WI, T1WI, DWI, ADC, and contrast-enhanced T1WI (T1+C). A total of 107 features were extracted per sequence, encompassing first-order (e.g., mean, standard deviation), shape, and various second-order texture metrics (GLCM, GLRLM, GLSZM, NGTDM, GLDM).

To build predictive models, they used a multi-step feature selection process:

  1. Univariate analysis: Identifying features with p < 0.05.
  2. Correlation analysis: Removing features with a correlation greater than 0.8.
  3. LASSO regularization: Selecting a parsimonious set of features.
  4. Multivariate stepwise regression: Retaining features with the minimal AIC.

This process was applied to create individual radiomic models for each MRI sequence. A "multi-sequence model" (labeled "All") was then built by combining meaningful features from all sequences. Finally, a "combined model" was developed by integrating the best radiomic features with clinical data (age, sex, WHO grade, etc.). The performance of these models was evaluated using Receiver Operating Characteristic (ROC) curve analysis and Decision Curve Analysis (DCA).

Datasets

The study utilized a dataset of 81 glioma patients diagnosed and treated at the Affiliated Hospital of Inner Mongolia Medical University between January 2018 and September 2021. The dataset included pre-operative MRI scans (T2W, T1W, DWI, ADC, T1W+C) and confirmed histopathological and molecular biomarker status (IDH, MGMT, TERT, 1p/19q). No new benchmark datasets were introduced.

Key Insights and Ideas

  • Radiomics is powerful, but context matters: Multiparametric radiomics from multiple MRI sequences ("All" model) generally outperformed single-sequence models in predicting IDH, MGMT, and TERT status. This highlights the complementary information provided by different imaging modalities.
  • Clinical data adds significant value: The "combined model," which integrated clinical features with radiomics, consistently showed superior predictive performance compared to radiomics-only models, particularly for IDH and MGMT. This underscores the importance of a holistic approach, blending imaging insights with patient demographics and tumor characteristics.
  • IDH and MGMT are more predictable: The models achieved the highest predictive accuracy for IDH and MGMT status, suggesting these biomarkers have more discernible imaging correlates compared to TERT and 1p/19q.
  • 1p/19q remains challenging: The study found no significant difference in predictive performance for 1p/19q status across the different models, indicating that this specific biomarker might be less amenable to prediction solely through current radiomic and clinical features.
  • DCA confirms clinical utility: Decision Curve Analysis demonstrated that the combined model offered a greater net clinical benefit across a range of threshold probabilities, reinforcing its potential for real-world application.

Key Results

  • The combined model achieved the highest Area Under the Curve (AUC) for predicting biomarker status:
  • IDH: AUC = 0.928
    • MGMT: AUC = 0.878
    • TERT: AUC = 0.757
  • The multi-sequence model ("All") also showed strong performance, outperforming single-sequence models for IDH, MGMT, and TERT.
  • For IDH prediction, the combined model (AUC 0.928) significantly outperformed single-sequence models (e.g., T2W AUC 0.752, T1W AUC 0.827, T1+C AUC 0.829) and the multi-sequence model (AUC 0.865).
  • For MGMT prediction, the combined model (AUC 0.878) was superior to single-sequence models and the multi-sequence model (AUC 0.833).
  • For TERT prediction, the combined model (AUC 0.757) showed improvement over DWI (AUC 0.635), ADC (AUC 0.654), and clinical models (AUC 0.657).
  • No significant differences were observed in predicting 1p/19q status among the models.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective design: This introduces the potential for unknown biases.
  • Risk of over-fitting: Given the number of variables and the dataset size, there's a possibility of over-fitting, especially with radiomic features.
  • Tumor segmentation: The segmentation included the entire tumor volume with high T2WI signal, without differentiating cystic, edematous, or parenchymal areas. Future studies could explore the impact of segmenting these specific sub-regions.
  • Scanner variability: While preprocessing steps were taken to mitigate scanner differences, external validation on multi-institutional datasets is crucial to improve the model's generalizability and clinical utility.

Future work should focus on multi-institutional validation, exploring more refined segmentation strategies, and potentially incorporating other imaging modalities or advanced AI techniques to further enhance prediction accuracy, especially for challenging biomarkers like 1p/19q.

Summary updated

BibTeX
@article{he2022multiparametric,
  author = {Jinlong He and Jialiang Ren and Guangming Niu and Aishi Liu and Qiong Wu and Shenghui Xie and Xueying Ma and Bo Li and Peng Wang and Jing Shen and Jianlin Wu and Yang Gao},
  title = {{Multiparametric MR radiomics in brain glioma: models comparation to predict biomarker status}},
  journal = {BMC Medical Imaging},
  year = {2022},
  doi = {10.1186/s12880-022-00865-8},
}
Figure