A team from Origin Quantum Computing Technology and collaborating hospitals integrates a variational quantum circuit into a Swin Transformer-based network, enhancing breast cancer screening accuracy and generalization by mitigating overfitting via quantum entanglement and superposition in a hybrid classical-quantum framework.

Key points

  • Integration of a 16-qubit variational quantum circuit replaces the Swin Transformer’s dense classifier to reduce parameter count by 62.5%.
  • Angle embedding encodes 8–16 normalized features directly into Y and Z rotations for depth-efficient implementation on NISQ hardware.
  • QEST achieves up to 3.62% balanced accuracy improvement in external validation and mitigates overfitting as shown by lower validation loss.

Why it matters: Embedding quantum circuits into deep learning models offers a scalable approach to reduce overfitting and parameter counts, paving the way for practical quantum-enhanced medical imaging applications.

Q&A

  • What is a variational quantum circuit?
  • How does angle embedding work?
  • Why replace the fully connected layer with a quantum circuit?
  • What is Balanced Accuracy (BACC)?
  • What hardware validated these experiments?
Copy link
Facebook X LinkedIn WhatsApp
Share post via...


Read full article

Quantum Machine Learning

Introduction: Quantum machine learning combines quantum computing principles with classical machine learning algorithms to harness quantum entanglement and superposition for data processing. It aims to accelerate pattern recognition and optimization tasks beyond current classical capabilities.

In traditional machine learning, data is represented in high-dimensional feature spaces and processed by neural networks or kernel methods on classical hardware. Quantum machine learning uses qubits, which can exist in superposition states, allowing a single quantum register of n qubits to encode 2n amplitudes simultaneously. This exponential state space offers potential speedups for certain tasks.

Key Concepts

  • Qubits: Quantum bits that hold complex probability amplitudes in superposition, enabling parallelism.
  • Entanglement: A quantum correlation where the state of one qubit depends on another, providing richer representational power.
  • Variational Quantum Circuits (VQC): Parameterized quantum circuits trained via classical optimization to perform tasks like classification.
  • Angle Embedding: Encoding classical features into rotation angles of quantum gates, mapping data onto the Bloch sphere with minimal circuit depth.
  • Measurement: Collapsing qubits to classical bits, where expectation values form the basis for loss functions and outputs.

Implementing a Quantum Classifier

  1. Data Preprocessing: Normalize classical data features, e.g., via arctangent functions or L2 normalization.
  2. Embedding Layer: Use angle embedding to map each feature to qubit rotations (RY, RZ gates).
  3. Variational Layers: Apply alternating entangling gates (e.g., CZ) and single-qubit rotations (U3 gates) with trainable parameters.
  4. Measurement: Read out qubit states and compute expectation values for classification logits.
  5. Classical Optimization: Use gradient descent (e.g., SGD) to update circuit parameters to minimize classification loss.

Applications in Medical Imaging

Quantum-enhanced models like QEST integrate a VQC into a Swin Transformer backbone for breast cancer screening. The VQC acts as a compact classifier, reducing overfitting and parameters. Experiments on real NISQ hardware confirm that quantum transfer learning can match or outperform classical methods, enabling potential quantum-first diagnostic tools.

Challenges and Considerations

Current quantum hardware faces limitations such as qubit decoherence, gate errors, and connectivity constraints. Strategies like error mitigation, qubit mapping, and shallow circuit design (e.g., angle embedding and minimal variational layers) help adapt machine learning models to Noisy Intermediate-Scale Quantum (NISQ) devices.

Future Directions

As quantum hardware scales and error rates drop, more qubits and deeper circuits will allow complex data processing, such as 3D medical scans. Research into efficient encoding, error mitigation, and hybrid architectures will drive practical quantum machine learning in healthcare.

Summary

Quantum machine learning offers a promising path for healthcare applications, leveraging qubit superposition and entanglement to reduce overfitting and parameter counts. Hybrid models, combining classical backbones with VQCs, demonstrate the feasibility of quantum-enhanced classification on genuine hardware, foreshadowing future breakthroughs as technology matures.

Quantum integration in swin transformer mitigates overfitting in breast cancer screening