@phdthesis{Segebarth2021,
  author    = {Segebarth, Dennis},
  title     = {Evaluation and validation of deep learning strategies for bioimage analyses},
  doi       = {10.25972/OPUS-24372},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-243728},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2021},
  abstract  = {Significant advances in fluorescence imaging techniques enable life scientists today to gain insights into biological systems at an unprecedented scale. The interpretation of image features in such bioimage datasets and their subsequent quantitative analysis is referred to as bioimage analysis. A substantial proportion of bioimage analyses is still performed manually by a human expert - a tedious process that is long known to be subjective. Particularly in tasks that require the annotation of image features with a low signal-to-noise ratio, like in fluorescence images of tissue samples, the inter-rater agreement drops. However, like any other scientific analysis, also bioimage analysis has to meet the general quality criteria of quantitative research, which are objectivity, reliability, and validity. Thus, the automation of bioimage analysis with computer-aided approaches is highly desirable. Albeit conventional hard-coded algorithms are fully unbiased, a human user has to set its respective feature extraction parameters. Thus, also these approaches can be considered subjective. Recently, deep learning (DL) has enabled impressive advances in computer vision research. The predominant difference between DL and conventional algorithms is the capability of DL models to learn the respective task on base of an annotated training dataset, instead of following user-defined rules for feature extraction. This thesis hypothesized that DL can be used to increase the objectivity, reliability, and validity of bioimage analyses, thus going beyond mere automation. However, in absence of ground truth annotations, DL models have to be trained on manual and thus subjective annotations, which could cause the model to incorporate such a bias. Moreover, model training is stochastic and even training on the same data could result in models with divergent outputs. Consequently, both the training on subjective annotations and the model-to-model variability could impair the quality of DL-based bioimage analyses. This thesis systematically assessed the impacts of these two limitations experimentally by analyzing fluorescence signals of a protein called cFOS in mouse brain sections. Since the abundance of cFOS correlates with mouse behavior, behavioral analyses could be used for cross-validation of the bioimage analysis results. Furthermore, this thesis showed that pooling the input of multiple human experts during model training and integration of multiple trained models in a model ensemble can mitigate the impact of these limitations. In summary, the present study establishes guidelines for how DL can be used to increase the general quality of bioimage analyses.},
  subject      = {Deeplearning},
  language  = {en}
}
@phdthesis{Griebel2022,
  author    = {Griebel, Matthias},
  title     = {Applied Deep Learning: from Data to Deployment},
  doi       = {10.25972/OPUS-27765},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-277650},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2022},
  abstract  = {Novel deep learning (DL) architectures, better data availability, and a significant increase in computing power have enabled scientists to solve problems that were considered unassailable for many years. A case in point is the "protein folding problem", a 50-year-old grand challenge in biology that was recently solved by the DL-system AlphaFold. Other examples comprise the development of large DL-based language models that, for instance, generate newspaper articles that hardly differ from those written by humans. However, developing unbiased, reliable, and accurate DL models for various practical applications remains a major challenge - and many promising DL projects get stuck in the piloting stage, never to be completed. In light of these observations, this thesis investigates the practical challenges encountered throughout the life cycle of DL projects and proposes solutions to develop and deploy rigorous DL models. The first part of the thesis is concerned with prototyping DL solutions in different domains. First, we conceptualize guidelines for applied image recognition and showcase their application in a biomedical research project. Next, we illustrate the bottom-up development of a DL backend for an augmented intelligence system in the manufacturing sector. We then turn to the fashion domain and present an artificial curation system for individual fashion outfit recommendations that leverages DL techniques and unstructured data from social media and fashion blogs. After that, we showcase how DL solutions can assist fashion designers in the creative process. Finally, we present our award-winning DL solution for the segmentation of glomeruli in human kidney tissue images that was developed for the Kaggle data science competition HuBMAP - Hacking the Kidney. The second part continues the development path of the biomedical research project beyond the prototyping stage. Using data from five laboratories, we show that ground truth estimation from multiple human annotators and training of DL model ensembles help to establish objectivity, reliability, and validity in DL-based bioimage analyses. In the third part, we present deepflash2, a DL solution that addresses the typical challenges encountered during training, evaluation, and application of DL models in bioimaging. The tool facilitates the objective and reliable segmentation of ambiguous bioimages through multi-expert annotations and integrated quality assurance. It is embedded in an easy-to-use graphical user interface and offers best-in-class predictive performance for semantic and instance segmentation under economical usage of computational resources.},
  language  = {en}
}