3D+t Terabyte Image Analysis

• Applicants: Professor Dr.-Ing. Ralf Mikut; Professor Dr. Peter Sanders
• Subject Area: Theoretical Computer Science ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: from 2015 to 2019
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 268484869

The project will develop methods for processing time series of high resolution 3D images. The goal is to obtain very high performance both with respect to result quality and computational efficiency on modern hardware. The exemplary application domain is tracking of objects (e.g., cell nuclei, cytoplasm, nanoparticles) in microscopic images where different object classes are labeled with particular fluorescent dyes. These tracking problems are highly important in biological research and challenging because the known computational approaches can only scratch the surface of the wealth of information available in the huge datasets currently being acquired. We also view this as a case study how algorithm engineering for large datasets can help in a concrete, important application domain.Our strategy is to use high performance (e.g., graph based) algorithms for 3D image processing to extract objects annotated with uncertainty information. Since these results in great reduction of data volumes, tracking algorithms can then afford to take a global view integrating uncertainty information into a consistent overall result. Fuzzy logic will be used as an overarching framework for uncertainty estimation and propagation. Achieving this will require significant progress in uncertainty modelling and propagation, in generalizing methods from 2D to 3D images, and designing parallel algorithms in a context of highly pronounced memory hierarchies. Some of the sub-problems have never been systematically solved before. For example, a rather classical pipeline of image preprocessing, foreground detection, and image segmentation will produce a graph of possible objects in each time frame (e.g., cell nuclei) and edges connecting objects expressing possible movements or splits of the object. For grasping the development of the whole organism, we have to extract likely subtrees from this graph. So far, it is unknown how to even phrase this problem as a well-defined optimization problem, let alone how to solve it with provably high quality.We will demonstrate the performance of our methods using large real world inputs of developing zebrafish embryos generated by light-sheet microscopy. These datasets reach more than 10 Terabytes per embryo and will be two orders of magnitude larger than previously reported tools can handle. Large realistic, simulated inputs including ground truth will be used to quantitatively assess result quality.

• DFG Programme: Research Grants