Parallelisation of multi point-cloud registration via multi-core and GPU for localization microscopy

Optical nanoscopy is a powerful technique used in biology to study subcellular structures and functions via specifically targeted fluorescent labels. Localization microscopy in particular offers a much better resolution (~10-50 nm) than conventional microscopy (~250 nm) while being relatively undemanding on the experimental setup and the subsequent image analysis.

The next revolution in imaging towards 1 nm resolution must realize a big increase of the labelling density. Only then can subcellular structures be imaged at the molecular level to study the molecular machinery of the cell. Relaxing the required labelling density using a priori information by data fusion of many identical entities is one of the challenges in the field.

This idea, taken from cryo-Electron Microscopy, has already lead to new insights in the structure of the Nuclear Pore Complex. However, due to the computational complexity this work is just starting. More so, algorithms and image processing from the field of electron microscopy cannot be translated into light microscopy due to the very different image formation. The data here consist out of localizations of fluorophores per particle which yields a list of x, y, z positions, a point cloud, with associated (anisotropic) measurement uncertainties sx, sy and sz, where sx= sy but sx≠ sz due to the imaging modality. These measurement uncertainties are different for each point in the cloud, due to the fluctuations in the detected photon signal and thus cannot be ignored in the registration for optimal performance.

The computational complexity arises (1) from the large number of particles ~103-104, (2) the large number of points per particle ~103 and (3) from the evaluation of the most suited cost function, the Bhattacharyya distance, for registration that takes the measurement uncertainty into account.

Typically computation times for a full tree would be months on a multi-core server where acquisition times are at most one afternoon. This project will be scientifically embedded into an ERC consolidator grant (to BR) on optical nanoscopy.

We want to develop an image processing framework that is multi-core and GPU capable for fast computation of the whole workflow including multi point-cloud registration. In order to do so, and disseminate the algorithm and the code, we want to implement the multi-point-cloud template-free registration based on our existing and well-used image analysis library DIPlib. However, first we need to invest in updating the core of the library, now 20 years old, to support multi-threading and GPU computing in a flexible, easy to access, and long term supportable way.