AURA: A Novel ResNet50 Approach to Predicting Karnofsky Performance Score (KPS) and Survival Time of Glioblastoma Patients

Overview

Glioblastoma (GBM) is an incredibly aggressive brain cancer, accounting for nearly half the number of malignant brain tumors each year. With one of the lowest five-year survival rates of all cancers (only about 6.8%), GBM has proved to be a challenge for doctors to treat and diagnose. This project leverages artificial intelligence to predict GBM disease progression with the metrics of survival time and Karnofsky Performance Score (KPS). After comparing several state-of-the-art neural networks, ResNet50 was determined to have the highest predictive performance for these measures of GBM disease progression. The most effective hyperparameter configuration and optimizer were found to increase model accuracy. The survival time and KPS ResNet50 models both scored a training accuracy of 100% and testing accuracies of 85.67% and 82.43% respectively, reaching accuracies previously unseen by other GBM predictive progression models. The models were finally integrated into the mobile app AURA, which displays patients’ personalized GBM prognosis and provides end-of-life care resources. Inexpensive and highly accessible, AURA is a high-fidelity prototype that brings essential resources to every GBM patient. AURA can help patients plan their treatments, as the information on disease prognosis it provides can be used to organize intervention and prepare the best care for the patient.