Quantum Convolutional Neural Network- Based Classification of Brain Tumor vs. No- Tumor MRI Images

Abstract

The problem of detecting brain tumors in Magnetic Resonance Imaging (MRI) is still a difficult one within the field of  neuro-oncology. Although Classical Convolutional Neural Networks (CNNs) have been successful for the past decade, they  still suffer from sub-optimal parameter efficiency and are unable to build a sufficiently expressive feature representation for  high dimensional medical volumes. In this paper, a Hybrid Quantum Convolutional Neural Network (HQCNN) is proposed  to solve this issue, where there is a possibility to directly leverage quantum superposition and entanglement in the feature  processing pipeline without any classical preprocessing. The architecture involves two key components: an efficient ResNet 8 network as the backbone to compress the MRI input, and an angle embedding layer to encode the compressed feature  vector into a variational quantum layer. The test accuracy of the proposed model is shown to be as high as 98.76%, with a  sensitivity of 98.92%, specificity of 98.61% and an AUC-ROC of 0.9941, significantly outperforming both classical CNN  based models and the latest hybrid quantum-classical models on the publicly available Br35H Kaggle data set consisting of  3000 images, with balanced classes of tumor and non-tumour. Quantum encoding and strongly entangling layers are  quantified in an ablation study, which is then validated externally on BraTS 2021 to show that the performance benefit  persists for an out-of-distribution clinical cohort. Overall, these results indicate that hybrid QCNNs are a feasible and  promising route to future strong medical image classification with near-term quantum computers.