[Paper Review] Hybrid Quantum Machine Learning Assisted Classification of COVID-19 from Computed Tomography Scans

Leo Sünkel, Darya Martyniuk+6 more|arXiv (Cornell University)|Oct 4, 2023

Quantum Computing Algorithms and ArchitectureComputer Science41 references3 citations

TL;DR

This paper proposes a hybrid quantum machine learning approach for classifying lung CT scans into COVID-19, CAP, or normal using a pre-trained ResNet18 for feature extraction and variational quantum circuits for classification. Despite challenges from limited quantum hardware and small datasets, the method achieves moderate performance, with better results on smaller, more balanced datasets, highlighting the need for deeper circuits and improved embedding techniques for medical image classification in the NISQ era.

ABSTRACT

Practical quantum computing (QC) is still in its infancy and problems considered are usually fairly small, especially in quantum machine learning when compared to its classical counterpart. Image processing applications in particular require models that are able to handle a large amount of features, and while classical approaches can easily tackle this, it is a major challenge and a cause for harsh restrictions in contemporary QC. In this paper, we apply a hybrid quantum machine learning approach to a practically relevant problem with real world-data. That is, we apply hybrid quantum transfer learning to an image processing task in the field of medical image processing. More specifically, we classify large CT-scans of the lung into COVID-19, CAP, or Normal. We discuss quantum image embedding as well as hybrid quantum machine learning and evaluate several approaches to quantum transfer learning with various quantum circuits and embedding techniques.

Motivation & Objective

• To address the challenge of classifying large, high-dimensional lung CT scans for COVID-19 detection using near-term quantum hardware.
• To evaluate hybrid quantum transfer learning (QTL) approaches in a real-world medical imaging context with limited quantum resources.
• To compare different quantum circuit architectures and embedding techniques for medical image classification under NISQ constraints.
• To investigate whether separating classical and quantum components improves performance and interpretability in hybrid quantum-classical models.
• To benchmark the approach on related medical datasets (PneumoniaMNIST, OrganAMNIST) to assess generalization and scalability.

Proposed method

• A hybrid quantum-classical pipeline is used, where a pre-trained ResNet18 extracts features from large CT scans, reducing dimensionality to 512 features.
• The 512-dimensional feature vector is embedded into a quantum state using amplitude encoding or angle encoding techniques.
• Two QTL variants are evaluated: one using a dressed quantum circuit (DQC) with jointly optimized classical and quantum layers, and another with only trainable quantum parameters and no classical fine-tuning.
• Quantum circuits with 4, 8, and 16 qubits are tested, with depth 1, 2, and 4, optimizing only the rotation gate angles in the second approach.
• The model is trained end-to-end in the DQC setup and only the quantum parameters in the second setup, enabling clearer attribution of performance to the quantum component.
• Various quantum circuits and embedding strategies are evaluated for their expressibility, trainability, and performance on multi-class and binary classification tasks.

Experimental results

Research questions

• RQ1 Can hybrid quantum machine learning effectively classify large lung CT scans into COVID-19, CAP, or normal classes under current NISQ hardware constraints?
• RQ2 How does the performance of quantum transfer learning vary between architectures with and without jointly optimized classical layers?
• RQ3 What impact do different quantum embedding techniques (e.g., amplitude vs. angle encoding) have on classification accuracy for medical image data?
• RQ4 Does increasing the depth and number of qubits in the quantum circuit improve model performance on the COVID-19 CT classification task?
• RQ5 How do the results on the COVID-19 dataset compare to those on smaller, more balanced medical datasets like PneumoniaMNIST and OrganAMNIST?

Key findings

• The model achieved higher accuracy and AUROC scores on the PneumoniaMNIST and OrganAMNIST datasets compared to the COVID-19 multi-class classification task, indicating better generalization on larger, more balanced datasets.
• The quantum circuit using amplitude embedding yielded results similar to other methods on the COVID-19 dataset, but performance remained limited, likely due to data imbalance and small sample size.
• The DQC approach with jointly optimized classical and quantum layers showed better performance than the strictly separated classical-quantum setup, though it obscured the contribution of the quantum component.
• The second approach, which trained only the quantum circuit parameters and excluded classical fine-tuning, enabled clearer evaluation of quantum model performance, though results were still modest.
• Increasing the number of qubits and circuit depth may improve model expressivity and performance, but further investigation is needed to determine optimal architecture and avoid barren plateaus.
• The results suggest that current NISQ hardware and circuit depth limitations restrict the ability to effectively classify large, complex medical images, even with hybrid quantum-classical designs.