Do Emulated Quantum Circuits Change What CNNs Look At? Performance and Explainability Comparison in Medical Image Classification

This preprint systematically compares a parameter-matched Hybrid Quantum-inspired Convolutional Neural Network (HQiCNN) with a classical CNN whose only architectural difference is an intermediate dense layer. Evaluated across two real-world medical image datasets and varied hyperparameters, neither model consistently dominates. HQiCNN performs best in intermediate-data regimes, while the classical CNN reaches the highest accuracy with the largest training sets. Removing entanglement yields comparable performance and substantially improves the scalability of classical quantum-circuit simulation. The authors also introduce SHAP-based comparison metrics, |SHAP|IoU and EMD_pos, finding that both architectures attend to anatomically plausible regions.