Quantum transfer learning for breast cancer detection.

Area of Science:

• Quantum computing and machine learning intersection
• Quantum machine learning (QML) applications

Background:

• Machine learning (ML) can benefit from quantum computing (QC) in handling large datasets.
• Hybrid classical-quantum neural networks offer a novel approach to ML tasks.

Purpose of the Study:

• To explore a quantum approach for classifying mammograms using hybrid classical-quantum neural networks.
• To evaluate the effectiveness of transfer learning (TL) in QML for medical image analysis.

Main Methods:

• Training hybrid classical-quantum neural networks using transfer learning (TL).
• Classifying full-image mammograms (malignant vs. benign) from BCDR dataset.
• Utilizing heatmaps to visualize network attention and evaluating performance metrics.
• Comparing results from real quantum devices, simulators, and classical methods.

Main Results:

• Achieved 84% accuracy in mammogram classification, surpassing the state-of-the-art 76.9%.
• Demonstrated potential benefits of QML for complex data generalization.
• Identified that specific network architectures perform better depending on the task.
• Validated findings on a real quantum device, comparing against classical and simulator results.

Conclusions:

• Hybrid quantum-classical models with TL show significant potential for improving mammogram classification accuracy.
• Further research is needed to fully explore the benefits of QML in medical diagnostics.
• The choice of network architecture is crucial for optimizing performance in QML tasks.