Deserno, Maurice ORCID: 0000-0001-6583-7360 (2025). Learning Spatiotemporal Representations of C. elegans From Bright-Field Microscopy Data Using Deep Learning Methods. PhD thesis, Universität zu Köln.

[img] PDF
Dissertation.pdf - Accepted Version
Bereitstellung unter der CC-Lizenz: Creative Commons Attribution.

Download (25MB)

Abstract

Recent biomedical research has led to a deep level of understanding of biological mecha- nisms we never reached before, enabling us to develop novel tests and cures for disease and improve our lives. However, there are still many disease and mechanisms to be researched and fully understood. One important element of this research are experiments with model organisms. Model organisms play a crucial role in biomedical research and thanks to their wide spread use in science, we understand these organisms in a level of detail that was not reached for any other organism yet. Researchers use model organisms in their experiments to study disease like Alzheimer’s and cancer, to understand aging and sleep and its underlying biomedical mechanisms. One model organism is the small roundworm Caenorhabditis elegans (C. elegans). Proposed as model organism by Sidney Brenner in the 1960s it quickly became a highly researched organism. The transparent body allows effortless observations of in vivo organs and inner processes, especially when applying stains that at- tach to specific biomolecules, highlighting them for improved observations. Using modern technologies, scientists can introduce all kinds of genetic mutations into an organism e.g. to understand the influence and interplay of specific genes in their experiments. Their findings do not only help to understand C. elegans but bring insights in human biology. Additionally, compared to other organisms C. elegans are easy to breed and cultivate under laboratory conditions making it a cheap and practical model organism. These and other factors result in C. elegans popularity in research and wide use in experiments. One of the main parts of the experiments conducted with C. elegans is quantifying its behavior. As behavior is an output of the organism’s neural network, it gives scientists valuable insights and helps them to understand the effects of their experiments. Together with the aforementioned benefits, behavior quantification of C. elegans enables broad possibilities for analysis and research. Unfortunately, traditional quantification of behavior is done by hand during time consuming observations of the nematodes under a microscope. Therefore, there is the need to automate this process with the promise to speed-up experiments, allowing scientists to spend more time on other tasks like interpretation of the results, obtained by the automated analysis, and conducting more experiments. Recently, Machine Learning (ML) and Deep Learning (DL) methods specialized on C. elegans have been proposed for tasks like detection, segmentation, tracking, pose estimation and behavior quantization. At the same time, more and more high-resolution recordings of C. elegans become available, thanks to the increased level of automatization in science. Although recent methods are implementing automation in this domain with increasing success, state-of-the-art approaches struggle when it comes to more challenging poses of C. elegans like coiling and self-intersecting or complex behavior like mating and aggregation. Additionally, many state-of-the-art approaches rely on hand-engineered features, omitting one of DLs strongest abilities: to find robust, discriminating, and possibly previously unknown features, suitable for the task. Based on this we see great potential in additional research into behavior quantization using DL, to find solutions for the aforementioned challenges and we aim to tackle them with this work. Here, we present our work, focusing on closing the gap between high-resolution behavior recording and time consuming and incomplete behavior quantification due to inaccessible poses and behavior of C. elegans. We present our novel instance segmentation approach, trained on synthetic data for segmentation of C. elegans in challenging scenes. We test our method on video data including C. elegans with coiling and heavy bending poses, as well as multiple individuals moving closely in parallel or overlapping each other. Ad- ditionally, we designed a tracking algorithm to present the abilities of our contribution. Our approach is capable of segmenting C. elegans in video frames depicting multiple individuals in challenging scenarios where previous methods failed to retrieve correct segmentation information. Our contribution allows for more detailed information required in downstream tasks and therefore enables more precise quantification and studies of the phenotype. Next, we present our self-supervised representation learning method for behavior sequences. By now, we focused on the spatial level of behavior by segmenting individual C. elegans in image data to extract pose information on a pixel level. In this work, we include the temporal component of behavior by learning how a pose changes in time using video recordings of C. elegans. First, we train a contrastive learning network to embed pose information without relying on curve or keypoint estimations. Second, we use the pre-trained contrastive learning network to learn representations of behavior sequences. We demonstrate the abilities of our approach by visualizing the embedding space and coloring it using hand-engineered features computed by state-of-the-art methods. These visualizations reveal that our network is able to capture hand-engineered features without explicitly enforcing them during training. Thanks to the absence of explicit features, our new approach is not limited to these but is rather able to capture properties previously inaccessible. Additionally, as our method is self-supervised, it does not require pose or behavior annotations and can directly be applied on videos of C. elegans, bridging the gap between fast data acquisition and slow data labeling. Combining both approaches allows to surpass the limitations of previous state-of-the-art methods and enables quantization of challenging behavior that other methods left unsolved or only partially solved.

Item Type: Thesis (PhD thesis)
Creators:
CreatorsEmailORCIDORCID Put Code
Deserno, MauriceUNSPECIFIEDorcid.org/0000-0001-6583-7360UNSPECIFIED
URN: urn:nbn:de:hbz:38-782938
Date: 2025
Language: English
Faculty: Faculty of Mathematics and Natural Sciences
Divisions: Faculty of Mathematics and Natural Sciences > Department of Mathematics and Computer Science > Institute of Computer Science
Subjects: Data processing Computer science
Life sciences
Uncontrolled Keywords:
KeywordsLanguage
c. elegansEnglish
deep learningEnglish
computer visionEnglish
instance segmentationEnglish
behavior quantificationEnglish
artificial intelligenceEnglish
Date of oral exam: 16 May 2025
Referee:
NameAcademic Title
Bozek, KatarzynaProf. Dr.
Beyan, OyaProf. Dr.
Refereed: Yes
URI: http://kups.ub.uni-koeln.de/id/eprint/78293

Downloads

Downloads per month over past year

Export

Actions (login required)

View Item View Item