Risk Prediction of Brain Glioblastoma Multiforme Recurrence Using Deep Neural Networks

Abstract

Brain Glioblastoma Multiforme (GBM) is one of the most dangerous types of primary malignancy, with a terrible 5-year survival rate of about 4% to 5% and a recurrence rate of up to 90%. Recently, the development of tumour-treating fields has shown positive clinical trial results for further survival extension but no superior efficacy has been observed in the treatment of recurrent GBM. The goal of this research is to discover the Machine Learning (ML) technology used to predict recurrence risk in glioblastoma patients before and after surgery. Radiomics is extensively being applied to advanced and conventional neuro-oncologic imaging data for glial tumours’ infiltrating margin detection, postoperative recurrence risk, and overall survival prediction is performed utilizing the rapid evolution of computational methods. Pre-operative Multi-Parametric Magnetic Resonance Imaging (MP-MRI) scans may be used to predict future tumour recurrence and to characterise tumour infiltration. Due to data inhomogeneity, Z-score normalisation and spatial resampling are initially used to address the MR image pre-processing. Subsequently, to address the problem of unbalanced data in medical image semantic segmentation, Recurrent Generative Adversarial Architecture (RNN-GAN) was developed. To construct a stable and validatable preoperative from the tumour area and the peritumoral edema area the research work utilized the CE-T1WI (contrast-enhanced T1-weighted imaging) model to objective response rate (ORR) as well as predict progression-free survival (PFS) in recurrent GBM patients treated with the combination of Bevacizumab and Nivolumab. Additional, to forecast glioblastoma recurrence, the research work proposed a Random forest (RF) model and a Deep neural network is utilized. The system is successfully trained and internally validated, and the patients at high risk of early recurrence are also identified. Subsequently, Inheritable Bi-objective Combinatorial Genetic Algorithm is presented as a feature optimization algorithm to select the relevant factors. The proposed approach has excellent accuracy in predicting GBM patient survival with recurrence rate. The proposed method is evaluated using Python and the proposed method is compared with existing SVM and LR models. The accuracy, specificity, and sensitivity of the proposed method are 3%, 4%, and 5% higher than the existing methods. Subsequently, this research demonstrates that predicted individual patient survival and time to recurrence produces high sensitivity, specificity and accuracy in a retrospective patient cohort.