Logo Logo
Help
Contact
Switch language to German
Enabling high-throughput image analysis with deep learning-based tools
Enabling high-throughput image analysis with deep learning-based tools
Microscopes are a valuable tool in biological research, facilitating information gathering with different magnification scales, samples and markers in single-cell and whole-population studies. However, image acquisition and analysis are very time-consuming, so efficient solutions are needed for the required speed-up to allow high-throughput microscopy. Throughout the work presented in this thesis, I developed new computational methods and software packages to facilitate high-throughput microscopy. My work comprised not only the development of these methods themselves but also their integration into the workflow of the lab, starting from automating the microscopy acquisition to deploying scalable analysis services and providing user-friendly local user interfaces. The main focus of my thesis was YeastMate, a tool for automatic detection and segmentation of yeast cells and sub-type classification of their life-cycle transitions. Development of YeastMate was mainly driven by research on quality control mechanisms of the mitochondrial genome in S. cerevisiae, where yeast cells are imaged during their sexual and asexual reproduction life-cycle stages. YeastMate can automatically detect both single cells and life-cycle transitions, perform segmentation and enable pedigree analysis by determining origin and offspring cells. I developed a novel adaptation of the Mask R-CNN object detection model to integrate the classification of inter-cell connections into the usual detection and segmentation analysis pipelines. Another part of my work focused on the automation of microscopes themselves using deep learning models to detect wings of D. melanogaster. A microscope was programmed to acquire large overview images and then to acquire detailed images at higher magnification on the detected coordinates of each wing. The implementation of this workflow replaced the process of manually imaging slides, usually taking hours to do so, with a fully automated, end-to-end solution.
Deep Learning, AI, High-throughput imaging, Computer Vision
Bunk, David Olaf Matthias
2023
English
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Bunk, David Olaf Matthias (2023): Enabling high-throughput image analysis with deep learning-based tools. Dissertation, LMU München: Faculty of Biology
[thumbnail of Bunk_David.pdf]
Preview
Licence: Creative Commons: Attribution-NonCommercial 4.0 (CC-BY-NC)
PDF
Bunk_David.pdf

54MB

Abstract

Microscopes are a valuable tool in biological research, facilitating information gathering with different magnification scales, samples and markers in single-cell and whole-population studies. However, image acquisition and analysis are very time-consuming, so efficient solutions are needed for the required speed-up to allow high-throughput microscopy. Throughout the work presented in this thesis, I developed new computational methods and software packages to facilitate high-throughput microscopy. My work comprised not only the development of these methods themselves but also their integration into the workflow of the lab, starting from automating the microscopy acquisition to deploying scalable analysis services and providing user-friendly local user interfaces. The main focus of my thesis was YeastMate, a tool for automatic detection and segmentation of yeast cells and sub-type classification of their life-cycle transitions. Development of YeastMate was mainly driven by research on quality control mechanisms of the mitochondrial genome in S. cerevisiae, where yeast cells are imaged during their sexual and asexual reproduction life-cycle stages. YeastMate can automatically detect both single cells and life-cycle transitions, perform segmentation and enable pedigree analysis by determining origin and offspring cells. I developed a novel adaptation of the Mask R-CNN object detection model to integrate the classification of inter-cell connections into the usual detection and segmentation analysis pipelines. Another part of my work focused on the automation of microscopes themselves using deep learning models to detect wings of D. melanogaster. A microscope was programmed to acquire large overview images and then to acquire detailed images at higher magnification on the detected coordinates of each wing. The implementation of this workflow replaced the process of manually imaging slides, usually taking hours to do so, with a fully automated, end-to-end solution.