Trustworthy AI in Medical Imaging: A Formal Verification Framework for Security and Governance

Artificial intelligence has transformed medical imaging but remains vulnerable to cybersecurity and governance risks. Adversarial perturbations are small, targeted changes to input data such as pixels in an image that cause a machine learning model particularly a deep neural network to make incorrect predictions even though the changes are imperceptible to humans. Furthermore, insecure data transmission and unverified model updates can also compromise diagnostic integrity and regulatory compliance. We propose a new quantitative verification framework for evaluating the trustworthiness of AI in medical imaging systems. The diagnostic workflow from medical image acquisition and AI inference is modelled as a stochastic system using the PRISM model checker. Attacker actions, defensive controls, and governance policies are represented as probabilistic transitions, enabling formal verification of confidentiality, integrity, and compliance using temporal logic. A case study employing the Fast Gradient Sign Method (FGSM) attack shows that the verified configuration lowers the probability of a successful breach compared with the unverified baseline, while maintaining throughput. The framework provides a basis for analysing and improving the security and governance of trustworthy AI in medical imaging.