Radiomics-Based Machine Learning Classification for Glioma Grading Using Diffusion- and Perfusion-Weighted Magnetic...

Area of Science:

• Neuroimaging
• Radiology
• Machine Learning

Background:

• Gliomas are primary brain tumors with distinct grades (low-grade vs. high-grade) that influence treatment and prognosis.
• Accurate differentiation between low-grade gliomas (LGGs) and high-grade gliomas (HGGs) is crucial for optimal patient management.
• Advanced imaging techniques like radiomics offer potential for non-invasive tumor characterization.

Purpose of the Study:

• To evaluate radiomics-based machine learning models for distinguishing LGGs from HGGs.
• To assess the utility of apparent diffusion coefficient (ADC) and cerebral blood flow (CBF) maps in glioma grading.
• To compare the performance of different machine learning classifiers in this diagnostic task.

Main Methods:

• Fifty-two glioma patients (18 LGGs, 34 HGGs) underwent 3.0-T MRI.
• Apparent diffusion coefficient (ADC) and cerebral blood flow (CBF) maps were generated.
• 91 radiomic features were extracted from ADC and CBF maps.
• Four machine learning models (LASSO-LR, RF, SVM-RBF, SVM-L) were trained and validated.

Main Results:

• ADC first-order-based skewness was a key feature across all models.
• All models achieved high areas under the curve (AUC) on the training set (0.965–1.000).
• Test set AUCs ranged from 0.717 to 0.917, indicating sufficient diagnostic performance for all models.

Conclusions:

• Radiomics using quantitative ADC and CBF maps show promise for differentiating HGGs from LGGs.
• Machine learning classifiers based on these radiomic features provide effective glioma grading.
• This approach can aid in non-invasive tumor classification and treatment planning.