Background
Machine learning (ML) can model clinical outcomes in neuro-oncology by considering image factors such as shape, intensity, and texture. These features are typically manually extracted and require human input to define the region of interest (i.e. tumour edge), which can inadvertently bias observed outputs.
Aims
To develop radiomics-based ML models predicting survival in individuals with glioma from preoperative MRI scans only. Then, model variability attributed to features extracted from a tumour volume versus whole brain was evaluated.
Methods
Preoperative MRI scans from 134 patients with were identified from our institutional database taking T1-weighted, postcontrast T1-weighted (T1c), T2-weighted, and T2 Fluid Attenuated Inversion Recovery (FLAIR) sequences as a minimum. Tumour edge and whole brain volumes were delineated on FLAIR. Radiomic features extraction was performed using in-house Matlab code, including wavelet transformation, feature reduction with ANOVA test, LASSO analysis and finally recursive elimination. Support Vector Machine (SVM) models to predict survival at 6, 12, 18 and 24 months were constructed. Survival was calculated from date of pre-operative MRI to death. 20% of each dataset was saved for model testing.
Results
For the entire cohort, there were 75.4, 53.7, 39.6, and 31.3% of the cohort alive at 6, 12, 18 and 24 months. Using features extracted from the tumour, AUC for the ROC curve predicting survival at these timepoints were 0.925, 0.789, 0.877 and 0.826 compared to 0.904, 0.834, 0.816 and 0.830 when using whole brain features. Both models had a similar feature type distribution, with shape features rarely represented.
Conclusions
Radiomics-driven SVM classification models based on preoperative MRI scans can predict survival in this population. Using FLAIR datasets, the performance of whole brain-based radiomic models was at least equivalent to tumour volume-based models in this cohort. These findings might influence model-development in neuro-oncology more broadly.