Radiomics and AI assisted analysis
In Radiomics, our research focuses on using radiological data to establish correlations between imaging biomarkers and their clinical outcomes to improve diagnosis, treatment monitoring, and prediction of treatment response. These biomarkers are conventional imaging biomarker classes, i.e., intensity-based, textural, and wavelet-based features, as well as biomarkers describing morphological and functional characteristics. The latter are often based on our own developed fully automated segmentation software tools that we run on central processing unit (CPU) and Graphic Processing Unit (GPUs). The fast GPU-accelerated interactive segmentation operations and precise rendering make our tool particularly suitable for efficient analysis of multimodal image datasets.
The field of computer vision (CV) could fundamentally change medicine by providing access to expert knowledge through CV. However, current machine learning methods fall short of expectations because training these methods is often done with datasets that are too small, incomplete, and heterogeneous.us data sets. Moreover, aspects such as robustness and explainability are very important to bring this new class of tools into clinical routine.
As part of our Artificial Intelligence (AI) work, we are focusing on Generative Adversarial Networks (GAN) to generate a nearly unlimited amount of high-resolution synthetic medical data that can be used to train other AI methods. Another focus of our research is to unmask the machine learning decision process. We use adversarial trained models that can significantly improve pathology detection compared to their standard methods. To enable successful training with a lower amount of training data, we are using hybrid modelling which combines complex physical and empirical models with machine learning methods.
Selected publications: