[Paper Review] Brain Tumor Diagnosis Using Quantum Convolutional Neural Networks

Muhammad Al-Zafar Khan, Abdullah Al Omar Galib+2 more|arXiv (Cornell University)|Jan 28, 2024

Brain Tumor Detection and Classification6 citations

TL;DR

The paper develops a Quantum Convolutional Neural Network (QCNN) hybrid model to classify brain tumor MRI images, achieving 99.67% accuracy on a 3-class dataset and validating quantum-enhanced medical imaging.

ABSTRACT

Accurate classification of brain tumors from MRI scans is critical for effective treatment planning. This study presents a Hybrid Quantum Convolutional Neural Network (HQCNN) that integrates quantum feature-encoding circuits with depth-wise separable convolutional layers to analyze images from a publicly available brain tumor dataset. Evaluated on this dataset, the HQCNN achieved 99.16% training accuracy and 91.47% validation accuracy, demonstrating robust performance across varied imaging conditions. The quantum layers capture complex, non-linear relationships, while separable convolutions ensure computational efficiency. By reducing both parameter count and circuit depth, the architecture is compatible with near-term quantum hardware and resource-constrained clinical environments. These results establish a foundation for integrating quantum-enhanced models into medical-imaging workflows with minimal changes to existing software platforms. Future work will extend evaluation to multi-center cohorts, assess real-time inference on quantum simulators and hardware, and explore integration with surgical-planning systems.

Motivation & Objective

• Motivate quantum machine learning (QML) for medical image diagnostics and address limitations of classical CNNs.
• Propose and implement a QCNN-hybrid architecture for brain tumor classification.
• Demonstrate high accuracy and generalization on a sizeable brain tumor MRI dataset.
• Show feasibility of quantum image processing within a practical ML pipeline.
• Highlight potential clinical implications and future research directions in QCNNs for medical imaging.

Proposed method

• Describe QCNN architecture and data encoding of classical images into quantum states.
• Apply quantum convolution and quantum pooling using U*, RX gates, CRZ/CRX gates, CZ entanglement, and measurements.
• Use a 2x2 quantum patching scheme and a SWAP-test-based pooling mechanism.
• Integrate a classical CNN within the QCNN framework in a hybrid model.
• Train the resulting network on quantum-processed images with a CNN classifier (Conv2D, MaxPooling, dense layers) for final class scores.
• Evaluate performance over 20 training/validation epochs with accuracy and loss plots.

Figure 1 : Architecture of a classical CNN.

Experimental results

Research questions

• RQ1 Can a QCNN-hybrid model achieve high accuracy on brain tumor classification compared to classical approaches?
• RQ2 How does quantum image processing affect feature representation and generalization in medical imaging?
• RQ3 What is the impact of dataset size and class distribution on QCNN performance for brain tumors?
• RQ4 Is a hybrid QCNN-CNN architecture practical within current (NISQ) computational constraints?
• RQ5 Which tumor types are most challenging to distinguish with QCNN features?

Key findings

• The QCNN achieved a peak validation accuracy of 99.67%.
• The model demonstrated robust training with decreasing loss across 20 epochs.
• Pituitary tumor cases were perfectly classified in the reported confusion matrix.
• Misclassifications mainly occurred between meningioma and glioma classes, indicating areas for refinement.
• The approach uses a hybrid quantum-classical framework and a quantum simulator to realize QCNN performance improvements over classical baselines.
• The dataset comprised 3,064 T1-weighted MRI slices from 233 patients across three tumor types (meningioma, glioma, pituitary).

Figure 2 : Hypothetical architecture of QCNNs model.