The IRIS Lab (Intelligent & Reliable Imaging Systems) develops AI-based technology that improves microscopy methods for biomedical use. The technology will be demonstrated on electron, optical, and ultrasound imaging but is applicable much wider. It will be able to unravel biological processes, from a molecular level up to a much larger scale.
Artificial intelligence (AI) is currently propelling nearly all computer vision applications in life science. Neural networks are trained to perform a certain task using very large sets of data, but which decisions these networks take, and why, is essentially a ‘black box’. For scientific applications this black box causes a serious dilemma: the learned knowledge is not transparent and cannot be reliably reused. Furthermore, how well a method performs its task is strongly tied to the specific data it was trained on and therefore cannot be easily used across different modalities of biological imaging that all produce different image data.
The aim of the IRIS lab is therefore to open the black box of AI. This can be accomplished by having the neural network approximate physical and system operations separately to enable the incorporation of known physical principles of image formation.
This approach has the potential to outperform more classical methods in core image processing tasks. The three goals are: efficient detection and tracking of objects, accurate labeling and classification of the objects and meaningful segmentation of the objects.
IRIS Lab will demonstrate their AI-methodology in electron- and fluorescence-microscopy and ultrasound imaging, but aims to become a campus-wide resource for researchers in any quantitative imaging- and visualization-intensive field.