TL;DR
This paper introduces a novel radiomics-based algorithm using advanced MRI techniques (APTw and DWI) to non-invasively predict 1p/19q co-deletion status in low-grade gliomas (LGG). The developed model demonstrates excellent predictive performance, comparable to experienced neuroradiologists, offering a promising tool for clinical decision-making in LGG management.
What Problem Does This Paper Address?
Low-grade gliomas (LGGs) are a group of brain tumors that, while generally slower-growing than high-grade gliomas, can still significantly impact patient prognosis. A critical factor influencing LGG treatment and outcomes is the presence or absence of 1p/19q co-deletion. This genetic alteration is particularly important for oligodendrogliomas, as it is associated with a better prognosis and a higher response rate to chemotherapy and radiotherapy. Currently, determining 1p/19q co-deletion status typically requires invasive methods like fluorescence in situ hybridization (FISH) after surgical biopsy. This creates a need for non-invasive methods to predict this crucial genetic marker, allowing for better surgical planning and treatment stratification before or during initial diagnosis. While conventional MRI features have been explored, advanced MRI techniques like Amide Proton Transfer Weighted (APTw) and Diffusion Weighted Imaging (DWI) offer richer information about tissue microenvironment and composition, which could potentially improve prediction accuracy.
Methods
The researchers developed a radiomics classification algorithm by extracting a large number of features from multiparametric MRI scans. Here's a breakdown of their approach:
- Image Acquisition: They utilized three-dimensional APTw, DWI, and conventional MRI sequences (T1, T2, FLAIR) from patients diagnosed with LGG.
- Radiomics Feature Extraction: A total of 8454 radiomic features were extracted from these images using the Feature Explorer software. These features encompassed various categories, including shape, first-order, texture (GLCM, GLRLM, GLSZM, NGTDM, GLDM), and wavelet-transformed features.
- Feature Selection: To identify the most relevant features and avoid overfitting, they employed a Least Absolute Shrinkage and Selection Operator (LASSO) regression model with fivefold cross-validation. This process selected 8 highly relevant radiomic features.
- Model Development: Three distinct models were built:
- Clinical Model: Based on significant demographic and clinical factors identified through multivariable logistic regression.
- Radiomics Model: Utilized the selected radiomic features to generate a radiomics score (Rad-score) through a linear combination.
- Combined Model: Integrated both the radiomics score and significant clinical factors.
- Model Evaluation: The performance of these models was assessed using receiver operating characteristic (ROC) curve analysis, DeLong's test for comparing AUCs, and decision curve analysis.
Datasets
The study retrospectively analyzed data from 90 patients diagnosed with LGG.
- Training Dataset: 72 patients.
- Test Dataset: 18 patients.
The datasets were split in an 8:2 ratio, with efforts made to balance clinical characteristics between the training and test sets. No new benchmarks were introduced; the study aimed to validate its model on existing patient data.
Key Insights and Ideas
- APTw and DWI are Powerful Predictors: The study highlights that advanced MRI techniques like APTw and DWI, when combined with radiomics, can extract subtle information that is highly predictive of 1p/19q co-deletion status, going beyond what conventional MRI can offer alone.
- Radiomics Outperforms Clinical Factors: The radiomics model and the combined model significantly outperformed the clinical model in predicting 1p/19q co-deletion, suggesting that imaging-derived features carry more discriminative power than traditional clinical variables for this specific task.
- AI as a Radiologist's Assistant: The radiomics model's performance was comparable to that of an experienced neuroradiologist and significantly better than resident physicians. This indicates that AI-driven radiomics can serve as a valuable tool to augment diagnostic capabilities, especially for less experienced clinicians.
- Texture Features are Key: The study found that texture features from T2WI and DWI, along with a histogram feature from APTw, were among the most contributing features to the radiomics model. This suggests that the spatial arrangement and intensity variations within the tumor, as captured by these features, are crucial for distinguishing between 1p/19q co-deleted and intact gliomas.
Key Results
- Excellent Predictive Performance:
- Training Set: Radiomics model achieved an AUC of 0.948, and the combined model achieved an AUC of 0.966.
- Test Set: Radiomics model achieved an AUC of 0.909, and the combined model achieved an AUC of 0.896.
- Superiority over Clinical Model: The radiomics and combined models significantly outperformed the clinical model, which had AUCs of 0.760 (training) and 0.766 (test).
- Comparison with Radiologists:
- In the training set, both radiomic and combined models performed significantly better than all three neuroradiologists.
- In the test set, the radiomics model significantly outperformed resident physicians (Reader B and C) and was comparable to the experienced neuroradiologist (Reader A).
Limitations and Future Work
- Single-Center, Retrospective Study: The study was conducted at a single center and retrospectively collected data. This limits generalizability, and the authors emphasize the need for multi-center data and external validation.
- Limited Sample Size: While 90 patients is a reasonable number, the test set of 18 patients is relatively small, which might have influenced the performance comparison between the radiomics and combined models in the test set.
- Exclusion of Other Modalities: The study did not incorporate perfusion-weighted imaging (PWI) or CT scans, which could provide additional functional and calcification information. Future work could explore integrating these modalities.
- Focus on 1p/19q Co-deletion: The current study focused solely on predicting 1p/19q co-deletion. The authors suggest that future research could extend this approach to predict other molecular subtypes, such as IDH1/2 mutations, and to understand diffuse gliomas more comprehensively.
Summary updated