Groundbreaking Application of Hybrid Quantum Neural Networks in Breast Cancer Image Classification

This article focuses on the application research of Hybrid Quantum Neural Networks (Hybrid QNN) in breast cancer image classification. The core finding is: the hybrid architecture combining classical deep learning and quantum computing achieves a 90.5% classification accuracy on the BreastMNIST dataset, outperforming traditional Convolutional Neural Networks (CNN), providing a feasible technical path for the implementation of quantum computing in the medical AI field.

Breast cancer is the most common malignant tumor among women worldwide. Early screening relies on medical image analysis, but manual reading has subjectivity and the risk of missed diagnosis. Traditional CNNs face bottlenecks in processing complex medical image patterns, while the superposition and entanglement properties of quantum computing can theoretically capture complex correlations that traditional models struggle to learn. The hybrid architecture combines the feature extraction capabilities of classical models with the pattern learning advantages of quantum circuits, avoiding the need for large-scale qubits required by pure quantum models.

The study compares the performance of multiple classical and hybrid models:

• Classical baselines: ResNet50 (86.0%), EfficientNetV2-S (85.5%), EfficientNet-B5 (88.5%) accuracy rates on BreastMNIST
• Hybrid models: ResNet50+QNN (89.0%), EfficientNetV2-S+QNN (88.0%), EfficientNet-B5+QNN (90.5%, best) accuracy rates on BreastMNIST
• For EfficientNetV2-S on the BreakHis dataset: pure classical model achieves 77.5%, hybrid model achieves 72.5%.

Two public datasets are used:

1. BreastMNIST: A subset of MedMNIST v2, containing annotated breast cancer ultrasound images, suitable for rapid validation
2. BreakHis: Histopathological images, testing the generalization ability of the model Evaluation metrics include accuracy, precision, recall, and F1-score, ensuring balanced control of false negatives/positives in medical AI applications.

Core of the hybrid architecture: Embedding quantum circuits as differentiable layers into classical networks. Implementation tools are PyTorch (classical framework) + PennyLane (quantum library). Process:

1. Classical models (e.g., EfficientNet-B5) extract high-level features
2. Features are encoded into quantum states and input into Parameterized Quantum Circuits (PQC)
3. Quantum circuits learn patterns via Variational Quantum Algorithms (VQA)
4. Measurement results are post-processed classically to output classification predictions.

• Quantum enhancement effect: EfficientNet-B5+QNN accuracy increased by 2% (from 88.5% to 90.5%), which has clinical significance
• Model differences: ResNet50/EfficientNetV2-S+QNN had smaller improvement margins, possibly related to feature extraction capabilities and quantum circuit design
• Dataset challenges: Hybrid models performed worse on BreakHis, indicating that quantum components are not suitable for all scenarios.

The project is fully open-source, providing Jupyter Notebook code (including ResNet50/EfficientNet series baselines and hybrid model experiments) that supports Google Colab cloud execution. The PennyLane simulator can simulate quantum circuits on classical computers, lowering the research threshold.

Current limitations: Small quantum circuit scale (due to hardware constraints), high computational overhead (low efficiency of classical simulation), insufficient model interpretability (limiting clinical applications). Future directions:

1. Use larger-scale quantum hardware to improve expressive power
2. Optimize computational efficiency to balance performance and overhead
3. Develop interpretable quantum machine learning models.