Prediction of low-grade glioma progression using MR imaging

Diffuse or infiltrative gliomas are a type of Central Nervous System (CNS) brain tumor. Among different types of primary CNS tumors, diffuse low-grade gliomas (LGG) are World Health Organization (WHO) Grade II and III gliomas. This study investigates the prediction of LGG progression using imaging features extracted from conventional MRI. First, we extract the imaging features from raw MRI including intensity, and fractal and multiresolution fractal representations the of the MRI tumor volume. This study uses a total of 108 LGG patients that is divided into 75% of the patients for training and the remaining 25% of the patients for testing from a pre-operative TCGA-LGG data. LGG progression prediction training model is performed using nested Leave-one-out cross-validation (LOOCV) on the training set. Recursive feature selection (RFS) method and LGG progression model training are performed in the inner cross-validation loop. The LGG progression prediction model is trained using Extreme Gradient Boosting technique. The performance of LGG progression prediction model is estimated using the outer cross-validation loop. Finally, we assess the predictive performance of the LGG progression model using the testing set. The training and testing procedures are repeated 10 times using 10 different training and testing sets. Our LGG progression prediction model achieves an AUC of 0.81±0.03, a sensitivity of 0.81±0.09, and a specificity of 0.81±0.10. Our results show promise of using non-invasive MRI in predicting LGG progression.