Quantum-Inspired NAS With Attention-Based Search Spaces in Medical Applications

Recent advances in Deep Learning models have significantly contributed to the development of more accurate solutions for medical applications, but this progress has come with increased computational complexity. In high-demand scenarios, designing efficient architectures that maintain high accuracy without adding computational overhead is crucial. This study proposes an enhanced Quantum-Inspired Neural Architecture Search (Q-NAS) algorithm, incorporating attention-based search spaces to optimize convolutional neural networks for medical image classification. By integrating Squeeze-and-Excitation (SE) block and Convolutional Block Attention Modules (CBAM), we aim to improve model accuracy and computational efficiency. The proposed approach was evaluated on four MedMNIST datasets, each representing different multi-class classification challenges. The experimental results show that while improvements in AUC and accuracy were modest (up to 0.76% and 2.15%, respectively) compared to ResNet-18 and ResNet-50, the attention-integrated Q-NAS models required up to 90% fewer parameters. This substantial reduction in model size, combined with acceptable computational costs, makes Q-NAS-derived architectures highly competitive for medical imaging applications.