Forschungsthemen
[DA] Analysis and Development of a Software Concept for Image Data Synthesizing based on Generative Adversarial Networks for Machine Learning Systems in Digital Health
Recently, data-driven applications, e.g. Machine Learning (ML)-based software solutions and concepts become increasingly important for different domains such as industrial optimization, traffic controlling, finance products, medicine, robots and much more. In contrast to classical ML based approaches, Deep Learning (DL) requires a vast amount of well annotated data to ensure result quality and stability. Especially these data and the according annotations are expensive to acquire and/or sensitive with regard to privacy or sovereignty aspects. Furthermore the data quality itself or its annotations are often inhomogeneous or
inconsistent, thus, becoming a challenge for ML systems. Finally, even the modern software systems are characterized by continuous changes of objectives and contexts, rendering it difficult to provide a stable data base for many potential future use cases.
On the other hand, Generative Adversarial Networks (GANs) is comparable new class of applications of Artificial Neural Networks (ANNs) <3> to generate synthetic data based on processed real data. For this purpose, two competitive ANNs (Generator and Discriminator) synthesize and validate new data from learned patterns in iterative manner, yielding convincing results in the areas of Computer Vision, Natural Language Processing, Time Series Synthesis or Semantic Segmentation. The technologies are comparable stable and available as open source solutions.
As the data challenges mentioned above apply especially for the domain of digital health, i.e. medical data is rare, heterogeneous and complex, and thus error-prone and crucial w.r.t. patients data privacy. Hence, the potential of synthetic data and its specific generation it should be investigated with the aim of a general-purpose software framework, based on distinct medical use cases.
Betreuer: Karsten Wendt