AF-HQCNN: Adaptive Federated Hybrid Quantum Convolutional Neural Network for Privacy-Preserving Medical Image Classification

Velkumar K, Dr.Mathalairaj J, Bhavani M, Archana R, Pavithra M, Venkatalakshmi M

Medical image classification is an important task in clinical decision support systems. Although classical deep learning models can represent massive amounts of data, they are not very scalable DEEP LEARNING REVIEW and stand each in their own right for privacy loss and computation inefficiency. Quantum machine learning (QML) has been proposed as an exciting field to complement classical machine learning by using quantum superposition and entanglement for increasing the expressibility of the models. Since the hybrid quantum-classical methods are subject to high qubit complexity [10], training instability (barren plateau problem) [13, 14], non-existent privacy protection mechanisms and poor generalisability over trained datasets [7]. Herein, we propose AF-HQCNN: an Adaptive Federated Hybrid Quantum Convolutional Neural Network (1) a lightweight classical feature extractor using only MobileNetV2, (2) variational quantum circuit (VQC) with angle and amplitude encoding with 4-8 qubits, (3) novel adaptive momentum-based quantum optimizer to mitigate gradient vanishing and gradient explosion issues, (4) federated learning aggregation layer layer to ensure protection of patient data privacy,(5), multi-dataset validation on MedMNIST, HAM10000, Brain MRI & Chest X-Ray datasets. Our experimental results show that AF-HQCNN leads to the mean classification accuracy of 97.6%, outperforming existing hybrid models by a margin of 2.7% with 87% fewer trainable parameters, faster convergence and robustness against privacy attacks.