[Paper Review] Medical image classification via quantum neural networks

Natansh Mathur, Jonas Landman+5 more|arXiv (Cornell University)|Sep 4, 2021

Retinal Imaging and AnalysisMedicine21 references34 citations

TL;DR

The paper explores quantum neural network approaches for medical image classification using quantum-assisted neural networks and quantum orthogonal neural networks, benchmarking on PneumoniaMNIST and RetinaMNIST with IBM hardware and simulators.

ABSTRACT

Machine Learning provides powerful tools for a variety of applications, including disease diagnosis through medical image classification. In recent years, quantum machine learning techniques have been put forward as a way to potentially enhance performance in machine learning applications, both through quantum algorithms for linear algebra and quantum neural networks. In this work, we study two different quantum neural network techniques for medical image classification: first by employing quantum circuits in training of classical neural networks, and second, by designing and training quantum orthogonal neural networks. We benchmark our techniques on two different imaging modalities, retinal color fundus images and chest X-rays. The results show the promises of such techniques and the limitations of current quantum hardware.

Motivation & Objective

• Investigate whether quantum neural networks can aid medical image classification on current hardware.
• Evaluate two QNN paradigms: quantum-assisted training/inference of classical networks and quantum orthogonal neural networks.
• Benchmark performance on two MedMNIST datasets (PneumoniaMNIST and RetinaMNIST) using various data loaders and hardware configurations.
• Assess scalability, robustness, and hardware-induced limitations of quantum approaches versus classical baselines.

Proposed method

• Use quantum circuits to assist training and inference of classical neural networks via QNN techniques adapted for NISQ hardware.
• Implement data loading with unary amplitude encoding using Reconfigurable Beam Splitter (RBS) circuits.
• Estimate inner products and vector multiplications with quantum circuits to replace classical linear algebra subroutines.
• Develop and train quantum orthogonal neural networks where weight matrices are orthogonal, using a quantum pyramid circuit and a gradient-based training in gate-angles (QPC).
• Execute hardware experiments on IBM superconducting machines (5-, 7-, and 16-qubit) and simulators, with PCA-based dimensionality reduction to 4 or 8 features.

Experimental results

Research questions

• RQ1 Can quantum-assisted neural networks match classical neural networks in medical image classification on standard datasets?
• RQ2 Do quantum orthogonal neural networks provide comparable or superior performance with hardware-aware training?
• RQ3 How do quantum data loaders and circuit depths impact accuracy and robustness on near-term quantum hardware?
• RQ4 What are the practical limitations of current quantum hardware for QNNs in medical imaging tasks?

Key findings

• Quantum-assisted neural networks achieve comparable AUC and accuracy to classical networks on PneumoniaMNIST and RetinaMNIST in simulations and some hardware runs.
• Hardware results show quantum methods can reach similar accuracy to classical baselines on several tasks, but hardware noise leads to performance drops on more difficult tasks.
• 8-dimensional inputs often yield better metrics than 4-dimensional inputs, with simulations matching classical performance and hardware showing variance over time due to noise.
• Orthogonal quantum neural networks trained via angle-based optimization (QPC) can produce substantially different models and, in some cases, higher test accuracy than SVB-trained counterparts.
• Simulations indicate quantum-inspired training for orthogonal networks can outperform SVB-based training in accuracy on RetinaMNIST (example: 75.25% vs 58.25% ACC in a specific setup).
• Unaries encoding and hardware-aware circuit optimizations improve reliability and accuracy of quantum computations on current IBM devices.