Beyond binary classification: a pilot study of imaging-derived glioma severity modeling using T1-weighted and...

Gliomas are heterogeneous brain tumors with variable biology and treatment response. Accurate, non-invasive assessment of tumor aggressiveness is essential for prognosis and treatment planning. Conventional machine learning (ML) approaches typically frame glioma grading as a discrete classification task, which may overlook substantial intra-grade heterogeneity. This pilot study explores a regression-based framework to derive a continuous imaging-derived severity score, providing a relative assessment of tumor aggressiveness anchored to, but not redefining, established WHO grades.

36 glioma patients (low-grade glioma; LGG: 58.33%, high-grade glioma; HGG: 41.67%) underwent 3D T1-weighted and DTI MRI on a 1.5 T scanner. Diffusivity maps were derived from DTI, and radiomic features were extracted. Sequential Feature Selection with a Random Forest regressor identified the most informative features. Fifteen ML regression models estimated the relative severity score, evaluated using MSE, MAE, and R², with fivefold nested cross-validation repeated 10 times. Model interpretability was examined using SHAP analysis.

The random forest model yielded the best performance (MSE = 0.066 ± 0.044, MAE = 0.182 ± 0.069, R² = 0.307 ± 0.432). T1-weighted features predominated, while DTI-derived measures, notably the robust mean absolute deviation of axial diffusivity, enhanced performance.

Regression-based modeling of radiomic features from T1-weighted and diffusion MRI enables the estimation of a relative, imaging-derived glioma severity score that captures intra-grade heterogeneity beyond categorical classification. While this model provides a continuous scale of radiological severity, it is intended as a complementary tool for assessing intra-grade heterogeneity and is not a replacement for the definitive WHO 2021 clinical grading.