Diagnostic performance of radiomics and machine learning algorithms in differentiating grade 2-3 gliomas from glioblastomas among adult-type diffuse gliomas

AIM: Glioma grading provides critical information for survival and prognosis. This study aims to determine the performance of machine learning (ML) algorithms using radiomic features to distinguish between grade 2-3 gliomas and glioblastomas (World Health Organization [WHO] grade 4) within adult-type diffuse gliomas, as defined by the recent central MATERIALS AND METHODS: During the period of 2017-2023, preoperative magnetic resonance imaging (MRI) of 92 patients who underwent surgery at our institution was retrospectively analysed. Tumour segmentations were independently performed by two radiologists using the '3D Slicer (version 5.2.2; Slicer Community, www.slicer.org)' software. Apparent diffusion coefficient (ADC), T2-weighted (T2W), and contrast-enhanced T1weighted (T1W-CE) images were used for segmentation. On the radiomics analysis, 107 features were extracted from each of the sequences. Reproducible features were determined by the intraclass correlation coefficient (ICC). Feature selection was performed using SelectKBest. Classification was done with 10 different ML algorithms. RESULTS: The highest area under the receiver operating characteristic curve (AUC) values in the classification were obtained from AdaBoost in the combined group (0.83, 95% confidence interval [CI]: 0.77-0.90) and random forest in the T1W-CE single-feature group (0.82, 95% CI: 0.70-0.93). CONCLUSION: This study showed that ML algorithms could distinguish grade 2-3 gliomas from glioblastomas in adult-type diffuse gliomas, with some algorithms performing better than others. However, the comparison was limited to glioblastomas within the grade 4 group and other grade 4 tumour subtypes were not included. Additionally, the lack of external validation is another important limitation of the study.