@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}
}