Abstract
Practical quantum computing (QC) is still in its infancy, often restricted to small-scale problems. Image processing, which requires handling vast amounts of features, presents a significant challenge for contemporary Noisy Intermediate-Scale Quantum (NISQ) hardware. This paper explores a hybrid quantum machine learning approach to classify high-resolution lung CT scans into COVID-19, Community-Acquired Pneumonia (CAP), or Normal categories. By leveraging hybrid quantum transfer learning, we demonstrate that quantum circuits can process real-world clinical data with performance comparable to classical baselines.
Introduction: The Feature Bottleneck in NISQ
Current quantum hardware operates in the NISQ era, characterized by limited qubit counts and high error rates. This creates a "feature bottleneck" when dealing with high-dimensional data like medical images. For instance, a standard 512x512 CT scan contains over 260,000 pixels, far exceeding the direct processing capacity of current quantum computers. To bridge this gap, researchers are investigating hybrid architectures that combine the feature-extraction power of classical deep learning with the potential representational advantages of quantum circuits.
Methodology: Hybrid Quantum Transfer Learning
The study utilizes Hybrid Quantum Transfer Learning (QTL), where a pre-trained classical model acts as a feature extractor, and a variational quantum circuit (VQC) serves as the classifier.
Feature Extraction
A pre-trained ResNet18 model, originally trained on ImageNet, is used to process the CT scans. This reduces the high-dimensional image data into a compact vector of 512 features.
Quantum Architectures
Two primary hybrid architectures were evaluated:
- Dressed Quantum Circuit (DQC): In this configuration, the VQC is "sandwiched" between two classical fully connected layers. A dimension reduction layer compresses the 512 features to match the qubit count (e.g., 4 or 8), and a final classification layer maps the quantum measurements to the target labels.
- Pure VQC with Amplitude Embedding: This approach eliminates the classical "dressing" layers. Instead, it uses Amplitude Embedding to encode the 512 features directly into the quantum state. This technique is highly efficient, allowing $N$ features to be represented using only $\log_2(N)$ qubits (e.g., 512 features in 9 qubits).
Results and Explainability
The hybrid models were tested on the COVID-CT-MD dataset. For the three-way classification (COVID-19 vs. CAP vs. Normal), the DQC models achieved accuracies between 0.55 and 0.60, matching classical baselines. In binary classification (COVID-19 vs. Normal), performance improved significantly, reaching accuracies of 0.70–0.75 and an AUROC of 0.80–0.85.
To ensure clinical relevance, the researchers employed Grad-CAM++ to generate heatmaps. These visualizations confirmed that the models correctly focused on relevant lung pathologies, such as ground-glass opacities, rather than background noise.
Discussion and Conclusion
While hybrid quantum models currently perform on par with classical counterparts, they face challenges. The "dressing" layers in DQC architectures contain many classical parameters, making it difficult to isolate the specific contribution of the quantum component. Furthermore, small dataset sizes and hardware constraints remain significant hurdles.
Despite these limitations, this work demonstrates a viable path for applying quantum machine learning to complex, real-world medical imaging. Future research will focus on more sophisticated quantum architectures and techniques to overcome training issues like "barren plateaus," moving closer to a demonstrable quantum advantage in healthcare.
Summary updated