A Specialized Importance-Aware Quantum Convolutional Neural Network with Ring-Topology (IA-QCNN) for MGMT Promoter Methylation Prediction in Glioblastoma

TL;DR

This paper introduces a quantum-inspired neural network architecture tailored for glioblastoma methylation prediction, leveraging quantum principles to improve accuracy and robustness in high-dimensional MRI data analysis.

Contribution

The study presents a novel importance-aware quantum convolutional neural network architecture that enhances feature learning and generalization in radiogenomic glioblastoma data.

Findings

Achieves high accuracy with fewer trainable parameters.

T1Gd modality shows higher discriminative power than mpMRI.

Model maintains robustness in noisy environments.

Abstract

GBM is a highly aggressive primary malignancy in adults, necessitating personalized therapeutic strategies due to its inherent molecular heterogeneity. MGMT promoter methylation is a pivotal prognostic biomarker for anticipating response to temozolomide-based chemotherapy. Although various AI frameworks have been developed for non-invasive MGMT prediction, spatial heterogeneity of methylation status and the high-dimensional and correlated nature of MRI data frequently constrain discriminative feature learning and generalizability of classical models. To circumvent these limitations, a specialized IA-QCNN architecture is proposed, based on the principles of quantum mechanics, including superposition and entanglement, and enabling more efficient representation learning in high-dimensional Hilbert space. The framework establishes a methodological bridge between GBM radiogenomics and quantum deep learning by integrating energy-based slice selection, importance-aware weighting, ring-topology quantum convolution, and folding-based pooling layers. When the model predicts MGMT promoter methylation status using both mpMRI and T1Gd images, experimental results demonstrate that the IA-QCNN achieves high accuracy despite its low number of trainable parameters while effectively minimizing the overfitting problem observed in classical models. Quantitative analyses reveal that the T1Gd modality possesses higher discriminative power than mpMRI, establishing a clinically significant sequence preference. Furthermore, the model exhibits exceptional robustness in hybrid noise environments, effectively utilizing noise as a regularization mechanism to enhance predictive performance. Consequently, the specialized IA-QCNN architecture provides a robust and computationally efficient alternative to classical approaches in the analysis of heterogeneous radiogenomic data.