MRI radiomics to differentiate between low grade glioma and glioblastoma peritumoral region

• Published in: Journal of neuro-oncology 2021-11, Vol.155 (2), p.181-191
• Main Authors: Malik, Nauman, Geraghty, Benjamin, Dasgupta, Archya, Maralani, Pejman Jabehdar, Sandhu, Michael, Detsky, Jay, Tseng, Chia-Lin, Soliman, Hany, Myrehaug, Sten, Husain, Zain, Perry, James, Lau, Angus, Sahgal, Arjun, Czarnota, Gregory J.
• Format: Article
• Language: English
• Subjects: Brain cancer; Brain Neoplasms - diagnostic imaging; Brain tumors; Clinical Study; Diagnosis, Differential; Edema; Feature selection; Glioblastoma; Glioblastoma - diagnostic imaging; Glioma; Glioma - diagnostic imaging; Glioma cells; Humans; Learning algorithms; Machine learning; Magnetic resonance imaging; Magnetic Resonance Imaging - methods; Medicine; Medicine & Public Health; Neoplasm Grading; Neurology; Oncology; Radiomics; Reproducibility of Results; Tumor cells
• ISSN: 0167-594X; 1573-7373
• DOI: 10.1007/s11060-021-03866-9

Background The peritumoral region (PTR) of glioblastoma (GBM) appears as a T2W-hyperintensity and is composed of microscopic tumor and edema. Infiltrative low grade glioma (LGG) comprises tumor cells that seem similar to GBM PTR on MRI. The work here explored if a radiomics-based approach can distinguish between the two groups (tumor and edema versus tumor alone). Methods Patients with GBM and LGG imaged using a 1.5 T MRI were included in the study. Image data from cases of GBM PTR, and LGG were manually segmented guided by T2W hyperintensity. A set of 91 first-order and texture features were determined from each of T1W-contrast, and T2W-FLAIR, diffusion-weighted imaging sequences. Applying filtration techniques, a total of 3822 features were obtained. Different feature reduction techniques were employed, and a subsequent model was constructed using four machine learning classifiers. Leave-one-out cross-validation was used to assess classifier performance. Results The analysis included 42 GBM and 36 LGG. The best performance was obtained using AdaBoost classifier using all the features with a sensitivity, specificity, accuracy, and area of curve (AUC) of 91%, 86%, 89%, and 0.96, respectively. Amongst the feature selection techniques, the recursive feature elimination technique had the best results, with an AUC ranging from 0.87 to 0.92. Evaluation with the F-test resulted in the most consistent feature selection with 3 T1W-contrast texture features chosen in over 90% of instances. Conclusions Quantitative analysis of conventional MRI sequences can effectively demarcate GBM PTR from LGG, which is otherwise indistinguishable on visual estimation.