BPCPPDYSOE21 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
BP: Fachverband Biologische Physik
BP 8: Bioimaging and Biospectroscopy
BP 8.2: Talk
Monday, March 22, 2021, 14:20–14:40, BPb
Motion-based segmentation for particle tracking: A fully-convolutional neuronal network that analyses movement — •Till Korten1, Walter de Back2, Christoph Robert Meinecke3, Danny Reuter3, 4, and Stefan Diez1 — 1B CUBE - Center for Molecular Bioengineering, Technische Universität Dresden, Dresden, Germany — 2Institute for Medical Informatics and Biometry (IMB), Carl Gustav Carus Faculty of Medicine, Technische Universität Dresden, Dresden, Germany — 3Center for Microtechnologies, TU-Chemnitz, Chemnitz, Germany — 4Fraunhofer Institute for Electronic Nanosystems (ENAS), Chemnitz, Germany
For single-particle tracking it is often necessary to separate particles of interest from background particles based on their movement pattern. Here we introduce a deep neuronal network that employed convolutional long-short-term-memory layers in order to be able to perform image segmentation based on the motion pattern of particles. Training was performed with ≈ 500 manually annotated 128x128 pixel frames. The segmentation result was used as input for a conventional single particle tracking algorithm. With this workflow 100% of all tracks belonged to microtubules that were propelled by kinesin-1 motor proteins along guiding channels and no tracks belonged to microtubules diffusing in the background. Furthermore, microtubules moving in a different orientation than the guiding channels during training, did not show up during inference. In conclusion, the deep-learning-based tracking resulted in almost twice as many (2800 vs. 1500) usable tracks that were 35 % longer compared to filtering after tracking.