QXRNet: A Hybrid CNN–QNN Model with Dynamic Feature Extraction and Variational Quantum Circuit

Hybrid Quantum Neural Networks (QNNs) have emerged as a new direction in the domain of medical image analysis, especially for low dimension images with less expressive power. In this work, we propose QXRNet, a hybrid architecture that employs a dynamic feature extraction pipeline using pretrained CNNs integrated with the high expressive power of a quantum processor. Experiments with multiple image sizes from 16 px to 512 px show that QXRNet achieves AUC scores that are on par with or better than their classical counterparts, while saving on the trainable parameters by an order of magnitude. With this performance, a 6-qubit QNN block contributes only 18 parameters, highlighting the efficiency of the hybrid design. Between the various hybrid frameworks employed, we find that ResNet-QNN performs better at smaller image sizes, while EfficientNet-QNN works at higher dimensions. Motivated by this, QXRNet uses their complementary strengths by creating a dynamic feature extraction pipeline. Additional experiments suggest that \((R_y)\) encoding results in better performance with minimal circuit overhead, while qubit scaling identifies 6 qubits as the optimal balance between stability, performance, and complexity. The hybrid QXRNet architecture is trained using classical gradient descent by unifying the gradient propagation using the chain rule and integrating the quantum parameter shift rules. This work combines parameter-efficient hybrid design and systematic benchmarking and demonstrates the potential of QXRNet as a compact and scalable alternative in medical image analysis.