LQ-HybridNet: A Parameter-Compact Classical Architecture with Quantum-Inspired Feature Transformations for Skin Lesion Classification

Background/Objectives: This study evaluates LQ-HybridNet, a parameter-compact classical architecture for skin-lesion classification. Methods: MobileNetV2 is augmented with amplitude–phase channel gating and a four-qubit variational circuit feature transformation. Crucially, all evaluations enforce a rigorous, leakage-resistant stratified lesion-level split to prevent data contamination. Here, quantum-inspired denotes a classical mathematical inductive bias, not quantum hardware execution or computational advantage. Results: On HAM10000, LQ-HybridNet achieved 90.80±0.25% accuracy, 0.7714 macro precision, 0.7257 macro recall, and 0.7428 macro F1-score, compared with 88.50±0.30% accuracy for MobileNetV2. The 2.30 percentage-point gain required 45.9K additional parameters. Across the evaluated backbones, accuracy gains ranged from 1.10 to 2.40 percentage points on HAM10000 and 1.77 to 4.00 points on ISIC 2018. Latency increased from 12.4 to 48.7 ms per batch of 32 images. Conclusions: The structured classical transformations yielded incremental benchmark improvements with small parameter overhead but approximately 3.9-fold higher latency. These results concern the evaluated architectures and benchmark distributions; parameter compactness does not imply inference efficiency.