MEDQUA: A NISQ-Aware Quantum Adapter for Medical Vision-Language Models

Vision-language models (VLMs) are promising for medical image classification but still face generalization limits from visual encoders and cross-site distribution shift. Fully quantum VLMs could offer richer representations, yet NISQ hardware makes end-to-end quantum training impractical. We introduce MEDQUA, a NISQ-aware quantum adapter that attaches to a pretrained VLM decoder. An entropydriven router sparsely selects tokens for a shallow variational quantum bottleneck, while a lightweight LoRA-based classical path processes all tokens to ensure stability and low cost. On MIMIC-CXR and ChestMNIST, MEDQUA consistently improves accuracy and AUROC over classical VLM baselines (including SFT) with modest overhead, showing that adaptively integrated quantum modules already yield practical gains. As coherence, error rates, and compilation advance, the same adapter can scale to deeper circuits and larger qubit counts without redesigning the classical backbone, providing a pragmatic route to quantum-enhanced medical VLMs.