FedQDFU: Exploring Federated Quantum Learning for Diabetic Foot Ulcer Classification

Diabetic foot ulcers are a serious complication of diabetes and can lead to infection, hospitalization, and amputation if not detected and managed early. At the same time, developing reliable AI models for medical imaging is challenging because clinical data are often distributed across institutions and cannot always be centralized.

Our recent work, FedQDFU: Non-IID Federated Quantum Learning for Diabetic Foot Ulcer Classification from Deep Image Features, explores whether a compact quantum machine learning model can operate within a federated learning framework while keeping data decentralized.

Why Federated Quantum Learning?

Federated learning allows multiple clients to collaboratively train a shared model without transferring their raw data to a central server. Instead, each client trains locally and only model parameters or updates are aggregated.

FedQDFU adds a second question: can a compact variational quantum classifier (VQC) serve as an efficient local model inside this federated setting?

The goal of this study is not to claim quantum advantage. Rather, it evaluates the practical trade-offs between compact quantum models and classical neural networks under controlled federated and non-IID conditions.

Deep Features from Three CNN Backbones

The study uses deep image features extracted from three pretrained convolutional neural network backbones:

• VGG16: 4096-dimensional features
• ResNet50: 2048-dimensional features
• DenseNet: 1024-dimensional features

To ensure a fair evaluation, duplicate samples were identified and handled before creating a shared development and locked test split. This prevents the same or near-identical samples from leaking across training and test partitions.

The extracted features are standardized and projected into a six-dimensional representation before being encoded into a 6-qubit VQC.

Federated Experimental Design

The experiments simulate five federated clients and compare:

• Centralized learning
• FedAvg
• FedProx

Both IID and non-IID client distributions are evaluated. Non-IID heterogeneity is controlled using Dirichlet partitions with:

α = 10, 1, 0.5, and 0.1

Smaller α values create stronger differences between client data distributions and therefore represent more difficult federated learning conditions.

Performance is evaluated using several metrics, including:

• Matthews Correlation Coefficient (MCC)
• ROC-AUC
• PR-AUC
• F1-score
• Sensitivity
• Specificity
• Balanced accuracy

What Did We Learn?

The results reveal an important trade-off.

The VQC is substantially more compact than the main classical MLP baseline. In the primary configuration, the quantum model uses only 43 trainable parameters, compared with 129 parameters for the MLP.

This translates into an approximately threefold reduction in federated update payload, which may be useful in communication-constrained environments.

However, compactness does not automatically mean better predictive performance. Under the α = 0.5 non-IID setting, the VQC produced lower mean MCC than the MLP across the three feature backbones, although the difference was statistically significant only for ResNet50.

Interestingly, the study did not detect a significant degradation in VQC MCC when moving from centralized learning to FedAvg. This suggests that the compact quantum classifier can remain relatively stable under federated aggregation, even though its overall predictive performance still trails the classical baseline.

Communication Efficiency vs. Computational Cost

One of the clearest findings is the difference between communication and computation.

Because the VQC has fewer trainable parameters, it requires less information to be exchanged between federated clients and the server.

At the same time, noiseless CPU-based quantum simulation remains expensive. In our experiments, VQC training took approximately 18–22 times longer than the corresponding classical model.

This highlights the current reality of quantum machine learning research: parameter efficiency can be attractive, but simulation cost remains a major limitation.

Toward Practical Hybrid AI Systems

FedQDFU provides a controlled evaluation of federated quantum learning for medical image classification using deep features.

The results do not establish quantum superiority. Instead, they show where compact quantum models may offer value—particularly in communication-efficient federated settings—and where classical models still remain stronger.

Future work will explore larger datasets, alternative quantum encodings, more efficient circuits, hardware-based execution, and stronger privacy-preserving federated learning strategies.

The broader objective is to understand how quantum and classical AI can complement each other in secure, distributed, and resource-aware healthcare applications.