Hybrid Quantum Machine Learning Assisted Classification of COVID-19 from Computed Tomography Scans

Leo Sünkel, Darya Martyniuk, Julia J. Reichwald, Andrei Morariu, Raja Havish Seggoju, Philipp Altmann, Christoph Roch, Adrian Paschke

TL;DR

Researchers have developed a hybrid quantum-classical pipeline to classify COVID-19 from high-resolution CT scans, bridging the gap between today’s limited quantum hardware and complex medical imaging. By using a pre-trained classical neural network to "distill" features for a variational quantum circuit, the team demonstrated that quantum machine learning can tackle real-world clinical data with accuracy comparable to purely classical approaches.

What Problem Does This Paper Address?

We are currently in the NISQ era (Noisy Intermediate-Scale Quantum), where quantum computers have very few qubits and are prone to errors. This creates a massive "feature bottleneck." A single 512x512 CT scan contains over 260,000 pixels—far too much data for a quantum computer to process directly.

While classical deep learning excels at medical image classification, researchers are eager to find a "quantum advantage"—a specific task where quantum algorithms might eventually outperform classical ones. This paper tackles the practical challenge of how to apply these nascent quantum techniques to high-dimensional, real-world medical data, specifically the classification of lung CT scans into three categories: COVID-19, Community-Acquired Pneumonia (CAP), or Normal.

Methods

The authors employ Hybrid Quantum Transfer Learning (QTL). The core idea is to let a classical model do the heavy lifting of "seeing" the image, while a quantum circuit handles the final "decision-making."

  1. Feature Extraction: They use a pre-trained ResNet18 (a standard classical convolutional neural network) to process the CT scans. This reduces the massive image data into a manageable vector of 512 features.
  2. The Quantum Part: These 512 features are fed into a Variational Quantum Circuit (VQC). The authors tested two main architectures:
  3. Dressed Quantum Circuit (DQC): The quantum circuit is "sandwiched" between two classical fully connected layers. These classical layers help compress the 512 features down to 4 or 8 qubits.
  4. Pure VQC with Amplitude Embedding: This approach removes the classical "sandwich." It uses Amplitude Embedding, a technique that mathematically squeezes data into qubits logarithmically (e.g., 512 features can fit into just 9 qubits).
  5. Embedding Techniques: They compared Angle Encoding (where features are represented by the rotation of a qubit) and Dense Angle Encoding (fitting two features per qubit) against Amplitude Embedding.

Datasets

The study utilized three datasets to validate their approach:

  • COVID-CT-MD: The primary dataset containing 512x512 volumetric chest CT scans from 169 COVID-19 patients, 60 CAP patients, and 76 healthy controls.
  • PneumoniaMNIST: A benchmark dataset of 28x28 images used for binary pneumonia classification.
  • OrganAMNIST: A multi-class benchmark dataset (28x28 images) focused on organ classification.

Key Insights and Ideas

  • The "Dressing" Dilemma: The "Dressed" circuits (with classical layers surrounding the quantum part) performed significantly better than the pure quantum circuits. However, the authors point out a catch: these classical layers contain many more parameters than the quantum circuit, making it difficult to determine if the "quantumness" is actually contributing to the performance or if the classical layers are doing all the work.
  • Logarithmic Scaling is Key: Amplitude embedding is highlighted as a vital tool for the NISQ era. Being able to represent 512 features in only 9 qubits is the only way to process complex images without massive classical pre-compression.
  • Explainability Matters: Using a technique called Grad-CAM++, the researchers generated heatmaps to show what the hybrid model was "looking at." They found the model correctly focused on lung pathologies (like ground-glass opacities) rather than random noise, which is crucial for clinical trust.

Key Results

  • COVID-19 Classification: On the multi-class task (COVID vs. CAP vs. Normal), the hybrid DQC models achieved an accuracy between 0.55 and 0.60, matching the performance of a purely classical baseline.
  • Binary Task: For the simpler task of COVID vs. Normal, accuracy jumped to 0.70–0.75, with an Area Under the ROC Curve (AUROC) of 0.80–0.85.
  • Benchmark Success: On the smaller MNIST datasets, the hybrid models performed exceptionally well, though the pure amplitude embedding approach still struggled compared to the "dressed" versions.

Limitations and Future Work

The authors are transparent about several hurdles:

  • Dataset Size: The COVID-CT-MD dataset is relatively small, which can lead to overfitting and limits the model's ability to generalize.
  • The "Quantum Advantage" Gap: While the hybrid model works, it doesn't yet provide a clear "speedup" or accuracy boost over classical methods. The classical parameters in the DQC still dominate the learning process.
  • Hardware Constraints: The experiments were limited by the depth and qubit count of current simulators and hardware.

Future Directions: The team suggests exploring more complex quantum circuit architectures, investigating ways to avoid "barren plateaus" (a common training problem in QML), and integrating image segmentation to help the quantum circuit focus only on the most relevant parts of the lung.

Summary updated

Abstract

Practical quantum computing (QC) is still in its in-fancy 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.
BibTeX
@inproceedings{sunkel2023hybrid,
  author = {Leo Sünkel and Darya Martyniuk and Julia J. Reichwald and Andrei Morariu and Raja Havish Seggoju and Philipp Altmann and Christoph Roch and Adrian Paschke},
  title = {{Hybrid Quantum Machine Learning Assisted Classification of COVID-19 from Computed Tomography Scans}},
  booktitle = {QCE},
  year = {2023},
  url = {https://openreview.net/forum?id=RbPV9o62bOM},
}

Loading related content...

Figure