Content-Aware AI Driven Driven Super-Resolution Microscopy

Super-Resolution microscopy relies on a finely balanced optimization of the optical configuration and analytical data treatment. These 2 parameters need to be correctly adjusted and matched to the properties of the biological sample to be imaged. Deviations from their optimal combination limits the quality of the data via the introduction of artefacts. Up to now, these optimisations have been based on educated guesses by researchers with some help from empirical criteria.

Recently we have established a new Super-Resolution approach named SRRF, capable of achieving unprecedented live-cell imaging capacity at the nanoscale [1]. In tandem, we have developed SQUIRREL, an analytical approach capable of calculating the quality and resolution of images generated in Super-Resolution Microscopy [2]. We further demonstrated for the first time the capacity of Deep Learning to massively improve microscopy data [3], by restoring corrupted or under-sampled cell imaging data - an approach named CARE.

In this PhD project, we propose to combine SRRF, CARE and SQUIRREL in a real-time analysis framework, integrated into a microscope itself designed to study long term cell cycle. This ability immediately opens the door to establish a Machine Learning approach (using Deep Reinforcement Learning) that adapts the microscope acquisition and analysis settings to maximise resolution and image quality. This will enable microscopy systems for the first time to: i) learn from the sample about how to improve imaging in real time, and ii) in a dynamic manner adapt to changes in the imaging properties of a living sample.