Federated Hierarchical Tensor Networks: a Collaborative Learning Quantum AI-Driven Framework for Healthcare

TL;DR

This paper tackles privacy and data-heterogeneity challenges in healthcare by introducing FedQTN, a federated framework that trains quantum tensor networks across multiple institutions using patch-based quantum encoding. The approach leverages MPS, TTN, and MERA architectures to capture complex correlations, with TTN/MERA showing superior accuracy and convergence under non-IID data and differential privacy constraints. Empirical results on RSNA, NIH, ADNI, and CT-kidney datasets demonstrate high AUCs and robust generalization, while communication efficiency and DP robustness are highlighted. The work suggests a viable route for privacy-preserving, collaborative medical imaging analysis with quantum-enhanced representations and scalable federated optimization.

Abstract

Healthcare industries frequently handle sensitive and proprietary data, and due to strict privacy regulations, they are often reluctant to share data directly. In today's context, Federated Learning (FL) stands out as a crucial remedy, facilitating the rapid advancement of distributed machine learning while effectively managing critical concerns regarding data privacy and governance. The fusion of federated learning and quantum computing represents a groundbreaking interdisciplinary approach with immense potential to revolutionize various industries, from healthcare to finance. In this work, we proposed a federated learning framework based on quantum tensor networks, which leverages the principles of many-body quantum physics. Currently, there are no known classical tensor networks implemented in federated settings. Furthermore, we investigated the effectiveness and feasibility of the proposed framework by conducting a differential privacy analysis to ensure the security of sensitive data across healthcare institutions. Experiments on popular medical image datasets show that the federated quantum tensor network model achieved a mean receiver-operator characteristic area under the curve (ROC-AUC) between 0.91-0.98. Experimental results demonstrate that the quantum federated global model, consisting of highly entangled tensor network structures, showed better generalization and robustness and achieved higher testing accuracy, surpassing the performance of locally trained clients under unbalanced data distributions among healthcare institutions.

Figures (8)

• Figure 1: Collaborative Quantum AI-based framework for healthcare advancement: Multiple hospitals participate in Quantum Federated Learning by collaboratively training models while preserving patient's privacy. Step 1: Hospitals locally process patient's image data using the quantum tensor networks (QTNs) model. Step 2: The quantum model updates of each hospital, not the raw data, are shared with the global server. Step 3: The global server aggregates the model updates from all participating hospitals and sends the updated state of the model back to all hospitals, ensuring that each participant benefits from the collective knowledge gained during the training process.
• Figure 2: Classical tensor networks and their quantum circuit representation counterparts. (a) Matrix product state (MPS): A one-dimensional chain of tensors that captures one-dimensional quantum systems efficiently. (b) Tree tensor network (TTN): It extends the idea of MPS to higher dimensions, using a tree-like structure to capture entanglement across multiple dimensions. (c) Multiscale entanglement renormalization ansatz (MERA): It captures logarithmic correlations while scaling with the length of 1D. (d-f) Implementation of the model as a quantum circuit. The tensor elements within the tensor networks are substituted with unitary operations to construct quantum circuits. Circles indicate eight qubit inputs prepared in a product state (left-hand side of the circuit), hash marks denote qubits that have not been observed beyond a specific stage in the circuit, and qubits are entangled via two-qubit unitaries (in square blocks). A preselected qubit is sampled (in red measurement operator), and the resulting distribution is considered as an output of the quantum model.
• Figure 3: Performance of federated quantum tensor networks on different datasets. (a) RSNA chest X-ray dataset: (b) NIH chest X-ray dataset: (c) ADNI MRI-scan dataset: (d) Kidney CT-scan dataset. An example of medical images of each dataset is displayed in the first row. The unequal distribution of all datasets among four hospitals/clients/participants is provided in the second row. The testing loss curves of three federated QTN models (FedMPS, FedTTN, and FedMERA) are presented in the third row. The illustration of the box plots shows the performance metrics: testing accuracy, precision, recall, and f-score of federated QTNs across all datasets.
• Figure 4: Confusion matrices and receiver operating characteristic curves (ROCs) of federated TTN model performance on different datasets. Left: RSNA chest X-ray for normal versus lung opacity classification. Middle: NIH chest X-ray for normal versus abnormal classification. ADNI MRI-scan for normal cognitive versus Alzheimer's disease (AD) classification. Right: Kidney CT-scan for normal versus malignant classification. ROCs depict the performance of quantum TTN on various datasets in federated settings. The maximum Area Under the Curve (AUC) reached 0.9818 for ADNI MRI-scans, while the minimum AUC was 0.9158 for the NIH chest X-ray dataset.
• Figure 5: Performance of individual local models and the global model. (a) Testing loss curves for the four locally trained hospitals on RSNA chest radiographs and global FedQTN. (b) Testing accuracy performance across four hospitals and FedQTN is illustrated through box plots, revealing the impact of insufficient data distribution (c-d) Testing loss curves for the four locally trained hospitals on ADNI MRI scans. The global FedQTN model significantly outperforms the locally trained models, achieving a testing accuracy of 94.5% along with a smoother convergence.