As of today, there is a plethora of cyber-physical instruments consisting of physical sensing (e.g., microscopy imaging) and cyber (digital) Artificial Intelligence (AI)-based predictions. These instruments raise concerns about safety because they rely on black-box AI models and do not contain any guardrails if physical and/or digital parts of the instrument fail or are attacked by an adversary. We would like to address the safety concerns by researching a metrology for establishing digital references [2], safety zones (boundaries), validation methods for AI risk management, and baselines for traceability of physical and digital parts of AI-enabled instruments [1]. Our research is applied to instruments used in regenerative medicine and cancer research [3]. 

[1] OpenAI Microscope, a collection of visualizations of every significant layer and neuron of 13 important vision models, URL

[2] Peter Bajcsy et al., “AI Model Utilization Measurements For Finding Class Encoding Patterns”, arXiv, Dec. 2022, URL

[3] Nicholas J. Schaub et al., “Deep learning predicts function of live retinal pigment epithelium from quantitative microscopy,” Journal of Clinical Investigation. November 14, 2019. DOI