QViT: A Noise-Resilient Quantum-Classical Vision Transformer with Quantum State Tomography for Lung Cancer Classification

Abstract

Lung cancer has always been the primary cause of cancer mortality in the world, and the exact identification of its subtypes is still vital in the treatment of patients to enhance their survival. Over the past years, Vision Transformers (ViT) have shown spectacular results in the field of medical image analysis and specifically in detecting and classifying lung cancer. They have not, however, been fully explored in terms of possible integration with quantum computing thus leaving a significant gap to take advantage of quantumadvantages of any real clinical significance. To improve the feature learning process, this study proposes a Quantum enhanced Vision Transformer (QViT) framework, in which it incorporates variational quantum circuits (VQC) into attention and feed-forward layers. Basic VQC and Quantum Approximate Optimization Algorithm (QAOA) VQC two-circuit families were tested at two depths resulting in four QViT configurations. The assessments with the help of quantum state tomography (QST) and analysis of noise demonstrated the scalability to deploy on Noisy Intermediate-Scale Quantum (NISQ) devices. This was applied to 3,150 CT images, where the QViT-QAOA-D1 configuration gave 98.52%accuracy, improved to 99.31% with Bayesian hyperparameter optimization without a corresponding increase in circuit resources. Circuits realized in Basic VQC could generate stronger entanglement but had to be deeper-designed whereas QAOA circuits could preserve high-purity states with shallow designs. The results make QViT viable and scalable, hardware-compatible, and make quantum-assisted ViTs an attractive medical AI prospect.