Improving Glioblastoma Multiforme Recurrence Prediction Through Integrated Radiomics and Deep Learning Techniques

Abstract

Glioblastoma multiforme (GBM) is one of the most lethal forms of brain cancer, with a five-year survival rate of only 4% to 5%. The recurrence rate is alarmingly high, reaching up to 90%. While tumor-treating fields have shown potential in extending survival, their efficacy in treating recurrent GBM remains limited. This study aims to leverage Deep Learning (DNN) to predict the recurrence of GBM in patients, both pre-and post-surgery. Utilizing advanced computational techniques, this research employs radiomics to analyze brain tumor images, aiding clinicians in identifying tumor spread, predicting post-surgical recurrence, and estimating patient survival. Pre-surgery, Multi-Parametric Magnetic Resonance Imaging (MP-MRI) scans are used to detect tumor locations and forecast potential recurrences. To enhance image processing, Z-score normalization and spatial resampling are applied. Additionally, a model was developed to address the issue of imbalanced data in medical imaging. The study utilized Contrast-Enhanced T1-Weighted Imaging (CE-T1WI) MRI to assess treatment effectiveness and predict recurrence-free survival. A Deep Neural Network was trained to forecast tumor recurrence, identifying patients at risk of early recurrence. Feature extraction from brain images was performed using the Inheritable Bi-Objective Combinatorial Genetic Algorithm. The accuracy of the recurrence predictions was validated and compared against other models, including CNN Inception-V3, CNN AlexNet, and VGG16, using the Python programming language. Results indicate that the proposed method surpasses existing techniques by 3%, 4%, and 5% in accuracy, specificity, and sensitivity, respectively. This research demonstrates that in a retrospective patient population, predictions of patient survival and time to recurrence exhibit high sensitivity, specificity, and accuracy, offering a promising tool for improving GBM management and patient outcomes. © 2024, International Publications. All rights reserved.