Unsupervised Model Discovery for Stereotypical Organisms (acronym UMDISTO)

• Applicants: Dr. Dagmar Kainmüller; Dr. Bogdan Savchynskyy
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 498181230

We aim at a novel methodology for solving large-scale multi-matching problems in an unsupervised way. Such problems emerge in applications in biology in the context of organisms that are stereotypical at the level of individual cells. These organisms have a fixed body plan, i.e., each specimen has the same number and arrangement of cells. This property of cell-level stereotypicity provides biologists with the unique opportunity to make multiple observations in distinct specimens, and merge these observations into a common reference frame at the level of individual cells. In particular, our project addresses the popular model organism Caenorhabditis elegans (further C. elegans), a roundworm that has 558 cells at its first larval stage.Observations can be made of the individual expression of almost all of its genes via fluorescence microscopy. Merging such observations into a cell-level reference frame yields an atlas of gene expression. Such an atlas would be an important contribution to reverse-engineering the “program” encoded into the worm’s DNA.A crucial step for creating such an atlas consists in establishing cell-by-cell correspondences between microscopy images of individual specimens. In particular, we address a dataset consisting of 265 images of different instances of the C. elegans worm. These yield a huge multi-matching problem both in terms of the size of each object (558 cells) and their number (265 worms). This is by far the largest and the most difficult problem of this type among considered in computer vision. However, the true challenge is not only optimizing a matching problem of this size, but training the matching costs for the optimization, since generating the ground truth matchings required for training is practically infeasible for theconsidered bio-imaging applications. Previous work in this direction includes partial (357 cells out of 558) annotations of 30 worm instances. However, complexity and duration of this partial annotation work prevents from extending it to the whole annotation as well as from repeating it for other stereotypical organisms and other development stages of C. elegans.Since to date there is no suitable methodology to solve this type of problems in such scale, we aim to close this gap. Specifically, our methodological goal is a novel unsupervised training and optimization pipeline for large-scale multi-matching. As a case study we consider the C. elegans worm in its first larval stage and aim to create a complete atlas of this organism based on microscopy images of 265 instances. To evaluate accuracy of our approach we will use the 30 instances for which the partial ground truth matchings are available.

• DFG Programme: Research Grants