Quantum-Classical Hybrid Neural Network with Graph Attention for Multi-Class Brain Tumor Classification from MRI Images

Abstract

Brain tumors continue to be a significant challenge for clinical neuro-oncology, as there is a significant intra-class variability and there is significant inter-class morphological overlap. The conventional convolutional neural network (CNN) and transfer learning methods have shown good performance, but they have a tendency to reach saturation in discriminative power when considering spectrally similar tumor types, such as glioma and meningioma. In this study, a novel Quantum-Classical Hybrid Neural Architecture (QCHNA) for 4-class brain tumor classification (glioma, meningioma, pituitary adenoma and healthy brain tissue) of T1-weighted Contrast Enhanced Magnetic Resonance Imaging (MRIs) is proposed. It features a frozen ResNet18 feature backbone, a learnable adjacency matrix in a Graph Neural Network (GNN) layer, dual Variational Quantum Circuits (VQCs), and a deep hierarchical classifier head, channel-wise self-attention mechanism. A class-balanced dataset of 5,600 MRI images (1,400 images per class) was curated from the publicly available Brain Tumor MRI Repository on Kaggle and used to train the model, and a held-out test set of 1,600 images was used to evaluate the model. The architecture proposed in the present invention had an overall accuracy of 93.06%, an F1-score of 92.90% and macro-average precision & recall of 93.36% and 93.06% respectively. The macro-average AUC-ROC was 0.9887. The ablation study results validate the positive impact of every architectural component on performance, and show that the proposed QCHNA architecture is better than the classical CNN baselines (p < 0.05, McNemar's test).