Kurtosis is An MRI Radiomics Feature Predictor of Poor Prognosis in Patients with GBM

• Published in: Brazilian journal of physics 2021-08, Vol.51 (4), p.1035-1042
• Main Authors: de Marco Borges, Pedro Henrique, Lizar, Jéssica Caroline, Faustino, Alexandre Ciuffi Correa, Arruda, Gustavo Viani, Pavoni, Juliana Fernandes
• Format: Article
• Language: English
• Subjects: Algorithms; Brain cancer; Classification; Edema; Feature extraction; General and Applied Physics; Kurtosis; Machine learning; Magnetic resonance imaging; Medical prognosis; Physics; Physics and Astronomy; Prognosis; Radiation therapy; Radiomics; Recall; Tumors; Workflow
• ISSN: 0103-9733; 1678-4448
• DOI: 10.1007/s13538-021-00912-9

Glioblastoma multiforme (GBM) is the most lethal and aggressive brain tumor. Magnetic resonance imaging (MRI) is currently used to diagnose and monitoring it. Radiomic features extracted from MRI are being used for correlations with the disease prognosis. A methodology to extract MRI radiomics features in a radiotherapy planning workflow, to assess features related to GBM poor prognosis patients, is presented. One hundred five radiomics features were extracted from T1 post-contrast MRIs of 43 GBM patients. The progression-free survival (PFS) time of all patients was also achieved. These patients were separated into two groups: PFS within 3 months (class value = 1) and PFS higher than 3 months (class value = 0). A machine learning model was built to predict the poor prognosis of patients using a random forest algorithm optimized for class = 1 classification, i. e. optimized for the recall score. Kurtosis was ranked as the most important feature for this classification. The final model presents a recall score of 1.00 ± 0.0 and an area under a receiver operating characteristic curve (AUROC) of 0.81 ± 0.04. The model used a kurtosis threshold value at 2.69 ± 0.03 for classification. The kurtosis values achieved for the PFS within 3 months indicate that these tumors are composed of almost the same amount of necrose, neoangiogenesis, edema, and/or tumor cells. In conclusion, a relation between the radiomics analysis of kurtosis in MRI with GBM’s poor prognosis was found, and the developed model may help guide the patient’s clinical management.