Optimizing MRI sequence selection for glioma classification: a radiomics-based analysis of WHO CNS grade, final pathological diagnosis, and IDH status

BACKGROUND: Multi-sequence MRI protocols for glioma classification are resource-intensive, prompting the need to identify optimal, potentially reduced imaging protocols. PURPOSE: To evaluate the diagnostic utility of individual MRI sequences (ASL, DWI, FLAIR, SWI, T1w, T2w) for classifying WHO CNS grade (II, III, IV), final pathological diagnosis (glioblastoma vs. astrocytoma), and IDH status (mutated, wildtype, IDH1) using radiomic features and machine learning. MATERIALS AND METHODS: We analyzed data from 501 patients with diagnosed brain cancer from the UCSF-PDGM dataset, by extracting radiomics features from all available imaging sequences and training uni- and multi-sequence machine learning models (ASL, DWI, FLAIR, SWI, T1w, T2w) for the three outcome tasks of interest. Performance metrics (accuracy, sensitivity, F1-score) were derived for each task, alongside feature selection frequencies. RESULTS: The uni-sequence machine learning models using FLAIR imaging features excelled for WHO CNS grade classification (57.4% accuracy), ASL for final pathological diagnosis (81.5% accuracy), and T1w for IDH status classification (66.7% accuracy). Interestingly, the combination of all sequences did not lead to significant improvements compared to the best performing uni-sequence machine learning models for the three tasks. Regardless of the input sequence, the models consistently relied on similar features, such as shape and demographic variables (e.g., patient age and tumor sphericity). CONCLUSION: Single sequences, such as FLAIR, ASL, and T1w, achieved the highest individual accuracies in the radiomics models for classification of tumor grade, final diagnosis, and IDH status, and may inform the design of more streamlined imaging protocols.