Title: Content-Adaptive Biomedical Image Understanding
Duration: June 2020 – October 2025
Research Area: Applied Data Science and AI, Life Science and Medicine
Biomedical images contain rich contextual information in which structures were labeled and how these structures are spatially related. However, biomedical images are often 3D volumes, acquired by tomography or light-sheet microscopy, of large size. This prevents the use of standard computer-vision techniques, and the question of how to extract and characterize information about shapes and their mutual spatial relations from such images remains open. In this project, we build on our previous work in content-adaptive image representation [1] to enable content-adaptive understanding of shapes and spatial arrangements in biomedical images using deep learning methods. This also includes the visualization of large distributed volume data for visual analytics and validation by human observers.
There are mainly two issues with understanding spatial structure in biomedical images: First, the image data are large (100s of Gigabytes to Terabytes for a single image), preventing single-GPU use of standard machine-learning frameworks. Second, information is mainly contained in the spatiotemporal dynamics of the signal intensity field, as morphology and texture are often dominated by the optics of the imaging equipment. This project provides a mathematically proven framework (proofs of global optimality, asymptotic convergence, error bounds, and computational equivalence) for addressing these questions and implements the solutions in practically useful and scalable software.
We are currently using the first results of this project in a collaboration with the Wyss Center for Neuroscience in Geneva to enable 3D neuro-histology of large human brain samples at sub-cellular resolution (see Scholler et al., 2023).
Going forward, the project will expand along two orthogonal directions: First, we will research topological approaches to characterizing shapes in APRs of images. This will enable us to exploit the natural invariance properties of biomedical images. Second, we will bring the technology into applications, for example in large-scale neuro-histology or in high-content drug screening.
[1] B. L. Cheeseman, U. Günther, K. Gonciarz, M. Susik, and I. F. Sbalzarini. Adaptive particle representation of fluorescence microscopy images. Nat. Commun., 9:5160, 2018.