Introduction
Quantum Machine Learning (QML) has traditionally focused on gate-based architectures, often utilizing Variational Quantum Circuits (VQCs) for hybrid classical-quantum tasks. While promising, these methods are currently constrained by the limitations of the Noisy Intermediate-Scale Quantum (NISQ) era, including low qubit counts and high decoherence rates. This research explores an alternative path: leveraging annealing-based Quantum Boltzmann Machines (QBMs) within a transfer learning framework to tackle complex, real-world image classification tasks.
By integrating the feature-extraction power of classical Convolutional Neural Networks (CNNs) with the probabilistic modeling capabilities of QBMs, the authors demonstrate a robust pipeline for medical diagnostic imaging, specifically the identification of COVID-19 from Computed Tomography (CT) scans.
The Hybrid Pipeline
The proposed architecture follows a multi-stage process designed to bridge the gap between high-dimensional image data and the binary nature of Boltzmann Machines:
- Classical Feature Extraction: A pre-trained ResNet-18 model, originally trained on ImageNet, serves as the backbone. It processes raw CT slices to extract high-level visual features.
- Dimensionality Reduction: The resulting 512-dimensional feature vector is passed through a classical linear layer, compressing it to 64 dimensions to accommodate the constraints of the quantum classifier.
- Binarization: Because Boltzmann Machines operate on discrete states, the compressed features are converted into bit-vectors.
- Quantum Classification: The final classification is performed by a QBM. Unlike feed-forward networks, QBMs are undirected energy-based models. To train this hybrid system, the authors utilize the SEQUENT (SEQuential Unsupervised-to-supervised Training) approach. This involves pre-training the classical layers with a surrogate classifier before introducing the QBM, which is then optimized using gradient descent based on Kullback-Leibler divergence.
Due to the limited availability of physical Quantum Processing Unit (QPU) time, the researchers employed Simulated Annealing (SA) as a classical proxy. SA mimics the thermal fluctuations of the annealing process, providing a reliable baseline for how a physical quantum annealer might perform.
Experimental Setup and Data
The study utilized the COVID-CT-MD dataset, a collection of lung CT scan slices categorized into three classes: COVID-19, Community-Acquired Pneumonia (Cap), and Healthy (Normal). The dataset was balanced to include 905 training images and 275 testing images per class. All images were preprocessed to 224x224 pixels and converted to grayscale to maintain compatibility with the ResNet-18 input requirements.
Key Findings
The performance of the SA-based QBM was compared against a classical Feed-forward Neural Network (FNN) of equivalent size using sigmoidal units. Several key advantages emerged:
- Superior Accuracy and AUC-ROC: The annealing-based models consistently outperformed the classical baselines, achieving test accuracies near 70% and AUC-ROC scores of approximately 80%.
- Accelerated Convergence: The QBM-based approach demonstrated significantly higher training efficiency. The best SA models reached peak performance within 6 to 11 epochs, whereas the classical FNN required 13 to 18 epochs to stabilize.
- Robustness: One of the most notable results was the stability of the QBM. The variance in performance across different random training seeds was significantly lower for the quantum-inspired model than for the classical baseline, suggesting that the energy-based landscape of the QBM may be less susceptible to getting stuck in poor local optima during training.
Discussion and Future Directions
While the results are promising, the authors identify several areas for further development. The current reliance on Simulated Annealing means that true quantum effects, such as quantum tunneling, were not present in this study. Transitioning the pipeline to physical hardware, such as D-Wave quantum annealers, is a necessary next step to determine if physical quantum properties offer additional performance gains or introduce new noise-related challenges.
Furthermore, the binarization step remains a potential bottleneck, as converting continuous feature vectors into bits likely discards nuanced information. Future research will focus on optimizing this interface and comparing the annealing-based approach directly against gate-model VQCs to establish which quantum paradigm is best suited for large-scale medical image analysis. Despite these hurdles, this work provides a critical proof-of-concept for moving Quantum Transfer Learning beyond the gate-model status quo.
Summary updated