Poster S11 - Machine Learning Framework for Glioblastoma Survival Prediction Using Multi Contrast MRI

Introduction: Glioblastoma (GBM) is the most common treatment-resistant cancer of the central nervous system with an overall survival rate of less than 2 years. Although, there are various treatment options for GBM, when we consider the aggressive nature of the tumor and the increased prevalence with age, early prognosis and advance treatment plans become crucial for patients and health care providers. The predictive capability of machine learning (ML) can be leveraged as a tool for patient risk assessments during the preoperative period. As magnetic resonance imaging (MRI) is the gold standard for GBM diagnosis, ML can be used to examine patterns in imaging and clinical data to aid clinicians in their decision making through predicting GBM prognosis. We hypothesize that the performance of classical ML models in predicting GBM overall survival would improve if a multimodal framework consisting of both clinical and MRI-derived imaging features is used for model training. The current study aims to evaluate the performance of classical ML models in their ability to classify GBM overall survival into three survival groups based on preoperative brain MRI sequences, clinically relevant imaging features, and non-imaging features such as patient demographics, tumor genomics, post-surgical outcomes, and overall prognosis.

Materials and

Methods: De-identified MRI cases (n = 458) from the 2021 Brain Tumor Segmentation (BraTS) Challenge were used for model training. A 5-fold cross-validation was used to ensure model generalizability. Three GBM sub-regions consisting of the gadolinium contrast-enhancing tumor (ET), the peritumoral edema (ED) region, and the non-enhancing necrotic tumor core (TC) were segmented using Re-Diffinet, a deep learning MRI tumor segmentation framework. Radiomic features of segmented tumors were extracted using the PyRadiomics toolbox. Clinically relevant imaging features, such as tumor burden, tumor diffusibility, and midline shift quantification were derived from tumor segmentations. Non-imaging features including demographic features (e.g., age at diagnosis and gender), tumor genomic features (e.g., IDH and MGMT gene mutations), and clinical features related to post-surgical outcomes (e.g., tumor resection status and survival days), were used to train ML models.

Overall survival prediction was evaluated using classical ML models such as random forest (RF), support vector machine (SVM), and multilayer perceptron (MLP). All models were trained using a standard ML workflow, consisting of feature selection to remove uninformative features, training to classify cases into three survival groups ( < 1 year, 1 – 1.5 years, and > 1.5 years), and post-hoc evaluations of overall survival prediction. The strongest predictors of GBM overall survival were identified using permutation feature importance.

All materials and methods are summarized in the graphical abstract shown in Figure 1.

Results, Conclusions, and Discussions: Results

The overall survival prediction accuracy for all models after 5-fold cross-validation is summarized in Table 1. Model prediction accuracies were improved when features of tumor diffusibility were incorporated into model training. However, adding tumor radiomic features did not improve the model accuracy and model overfitting was observed. The RF classifier outperforms all models when tumor segmentation and diffusibility features were ingested during the training and validation process, achieving a 60.20% accuracy in the validation set. Permutation feature analysis reveals that age at diagnosis, enhancing tumor volume, and tumor diffusibility (tumor volume / convex hull tumor volume) were the strongest predictors of overall survival.

Discussion

This study investigates the feasibility of applying a multimodal approach for predicting the overall survival of newly diagnosed GBM patients. Our results identified candidate predictors of overall survival. Though these features may be useful for clinicians, the models did not achieve high predictive accuracy for survival group classification. As GBM tumors have a heterogeneous microenvironment and a complex biological disease progression, extracting additional quantitative tumor features using convolutional neural networks may address sub-optimal survival prediction accuracy. While the MRI cases used for model training spanned numerous academic medical institutions, survival prediction performance should be re-evaluated with more cases to address the potential case-by-case variability in the equipment and post-processing pipelines used during MRI acquisition. This would improve the generalizability of our findings and potentially improve classification accuracy. Nonetheless, our results demonstrate the potential for adopting a multimodal ML framework for advancing GBM disease stage classification based on overall survival.

Conclusion

The capability of ML models to predict overall survival in GBM patients based on initial neuroimaging diagnosis presents an opportunity for clinician-aide diagnostic tools for GBM patient risk stratification to be adopted in the clinical setting. These predictive tools would give clinicians unique insights into each patient’s disease stage, enabling them to improve diagnostic accuracy, personalize disease management, track disease progression, and determine patient eligibility for clinical trials. With greater refinement, multimodal neuroimaging ML frameworks may reveal imaging correlates of overall survival and may serve as a diagnostic aid for clinicians in the clinical setting.

Acknowledgements (Optional):