An Integrated MobileNetV2 and DeepSurv Framework for Multimodal Glioblastoma Survival Prediction

Authors: 

Sadiya Marfua¹, Kshama Jain², Sriram Gutam³, Mr. R. Vinod Kumar⁴

Scholars¹²³, Department of Computer Engineering, Methodist College of Engineering and

Technology, Abids, Hyderabad, Telangana, 500001, India.

Professor⁴, Department of Computer Engineering, Methodist College of Engineering and

Technology, Abids, Hyderabad, Telangana, 500001, India

ABSTRACT

Glioblastoma multiforme (GBM) creates major obstacles which prevent physicians from predicting patient outcomes because of its diverse molecular and architectural characteristics. The research introduces a multimodal system which combines magnetic resonance imaging (MRI) with genomic expression data and clinical information to forecast patient survival rates. Our method achieves patient data matching by using feature-level fusion to combine data from separate processing methods which operate independently on different datasets. The MobileNetV2 network which has been pretrained processes MRI volumes to extract imaging features while the top 500 informative genes are selected from genomic data through variance-based filtering. The system processes clinical variables through a process which encodes and normalizes them together with other extracted data points. The DeepSurv neural network develops its hazard function through nonlinear learning by using the combined feature set which it receives as input. The framework achieves a 0.89 concordance index on TCGA-GBM and BraTS 2021 datasets which outperforms all single‑modality baseline methods. The Kaplan-Meier analysis shows that patients in the high-risk and low-risk groups develop distinct survival patterns which produce statistically significant results with p 0.001. The research findings confirm that multimodal data integration methods provide accurate GBM patient prognosis for clinical applications.

Keywords— Glioblastoma, Survival Prediction, DeepSurv, MobileNetV2, Multimodal Learning, Feature Fusion