% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@PHDTHESIS{Eschweiler:993832,
author = {Eschweiler, Dennis},
othercontributors = {Stegmaier, Johannes and Truhn, Daniel},
title = {{I}mproving deep learning-based instance segmentation and
generative approaches for 3{D} microscopy image data},
school = {Rheinisch-Westfälische Technische Hochschule Aachen},
type = {Dissertation},
address = {Aachen},
publisher = {RWTH Aachen University},
reportid = {RWTH-2024-09006},
pages = {1 Online-Ressource : Illustrationen},
year = {2024},
note = {Veröffentlicht auf dem Publikationsserver der RWTH Aachen
University; Dissertation, Rheinisch-Westfälische Technische
Hochschule Aachen, 2024},
abstract = {Recent developments in microscopy imaging techniques have
significantly enhanced the capacity to capture highly
detailed image data. This capability has empowered
biological experts to conduct an increasing range of
insightful experiments, resulting in a surge of both the
volume and the size of microscopy image data that
necessitate processing. As a consequence, the demand for
automated approaches has grown substantially, with modern
deep learning-based approaches taking the lead in offering
precise results that surpass the outcomes achieved by
classical techniques. These deep learning-based approaches
usually require large annotated datasets for optimal
performance. Unfortunately, the general scarcity of such
annotated datasets serves as a bottleneck, constraining the
performance and broader applicability of these techniques.
Furthermore, the tedious process of manual annotation,
particularly in the context of 3D image data, is exceedingly
time-consuming and often impractical, presenting another
challenge in addressing the overall shortage of annotations.
Therefore, the shift from classical to deep learning-based
approaches necessitates the design of new processing
pipelines that either deal with the scarcity of annotations
or pose solutions to overcome it. Consequently, the major
contributions of this thesis focus on the development of
robust instance segmentation approaches and the design of
generative pipelines that help to mitigate the
aforementioned annotation scarcity. Ultimately, the proposed
approaches strive towards enhancing the applicability of
modern deep learning-based segmentation approaches to a
diverse range of microscopy datasets, with the vision to
eliminate the need for manually generated annotations.In
this context, instance segmentation approaches have been
designed to enhance robustness and generalizability,
ultimately improving the efficiency of existing annotated
datasets. Three distinct instance segmentation approaches
have been proposed, each concentrating on different aspects
of robustness. These range from generating intuitive results
for further processing, to exploiting redundancy from
information density and to leveraging model
knowledge.Moreover, data enrichment and generative pipelines
have been developed to directly address the vital need for
manually annotated datasets. A data enrichment technique
that relies on local image statistics has been employed to
create diverse image variations, which allows to diversify
existing datasets without compromising existing annotations
by inherently preserving local structures. Experiments
involving a data generation pipeline and simulation
approaches have laid the groundwork and determined crucial
requirements for setting up optimized generation pipelines.
Those include approaches based on adversarial training
concepts and diffusion models, which both focus on the
generation of realistic synthetic datasets based on existing
or simulated annotations, andoffer different levels of
control and intuitiveness. Importantly, these generated
datasets have been proven to be suitable for training
segmentation approaches, eliminating the reliance on manual
annotation efforts. To ensure broad accessibility, software
for these approaches has been made publicly available
through different ways, accommodating researchers with a
range for programming experience levels and backgrounds.
Lastly, several real-world application scenarios have been
conducted in collaboration with fellow researchers, which
serve to emphasize the adaptability and practicality of the
proposed methods.},
cin = {611710},
ddc = {621.3},
cid = {$I:(DE-82)611710_20140620$},
pnm = {DFG project 447699143 - On-the-fly Datensynthese für eine
auf Deep Learning basierende Analyse von 3D+t Mikroskopie
Experimenten (447699143)},
pid = {G:(GEPRIS)447699143},
typ = {PUB:(DE-HGF)11},
doi = {10.18154/RWTH-2024-09006},
url = {https://publications.rwth-aachen.de/record/993832},
}
guest ::