Learning Semantic Part Segmentation of Model Organisms from Little to No Training Data

Dagmar Kainmüller, Berlin Institute of Health

October 11, 2018


Annotating cells in 3D light microscopic images of the nematode worm C. Elegans is an elementary task for cell-level studies of gene expression. Manually annotating individual cells with their unique biological names is hard even for trained anatomists. There exist, to date, thirty 3D volumes of L1 larvae in which most cells have been expert-annotated, and this effort took 5 years to complete. Such annotations do not exist for other stages of development of the worm. For an exhaustive study of gene expression at the cell level, thousands of worms will have to be annotated, which will only be possible if the annotation task can be automated. Our work explores how to best leverage a small set of annotated worm images for supervised training of automated annotation models. In particular, we compare (1) the current state of the art for automated annotation, Active Graph Matching, which matches a global point distribution model of cell locations via combinatorial optimization, to (2) a location-sensitive U-Net trained to predict hundreds of different cell names on the few avilable training volumes. To our surprise, we find that the U-Net gets close to the performance of the model based approach. Furthermore, we explore possibilities to alleviate the need for even a small training set, via unsupervised training of a model for nuclei annotation.


The event will take place on Thursday, 11 October, 2018 at 2pm at Room: B.128 (Mathematikon B, 3rd floor, Universität Heidelberg Berliner Str. 43). Please ring the bell at the entry door (towards Berliner Strasse) which says: Universität Heidelberg; HCI am IWR. Kindly help us plan ahead by registering for the evnent on our meetup page.