Do emulated quantum circuits change what CNNs look at? Performance and explainability comparison in medical image classification

As hybrid quantum-classical neural networks gain traction in medical imaging, it remains unclear whether their performance gains are genuine or merely artifacts of model design. This study investigates whether replacing a standard dense layer in a CNN with a classically emulated quantum circuit (EQC) provides a meaningful advantage, ensuring a fair comparison by using identical convolutional backbones and parameter-matched architectures.

The authors developed a Hybrid Quantum-inspired Convolutional Neural Network (HQiCNN) and compared it against a classical CNN baseline. Both models share the same feature-extraction backbone, differing only in the intermediate dense layer. The researchers evaluated these models on two medical datasets (retinal OCT and brain MRI) across a wide range of hyperparameters, including training set size, learning rate, and circuit configuration. To ensure clinical relevance, they introduced two SHAP-based explainability tools, |SHAP|IoU and EMD_pos, to verify that both models focus on anatomically plausible regions of the images.

The results demonstrate that neither architecture consistently outperforms the other. HQiCNNs show superior performance in intermediate-data regimes, whereas classical CNNs achieve higher accuracy as training data increases. Interestingly, the study found that entanglement—a core quantum feature—is not strictly necessary for performance in these tasks, as removing it maintained comparable accuracy while significantly improving the scalability of the simulations. Furthermore, the explainability analysis confirmed that both models attend to similar, clinically relevant anatomical features, suggesting that the quantum-inspired layer does not fundamentally alter the model's focus but provides a different way to process learned features.