Machine Learning
for Non-Textual Data

In particular, image data that has not yet been intensively researched using AI methods, such as echocardiography sequences and longitudinal image studies for monitoring patients, will be analyzed using new methods in order to extract relevant markers and use them for prediction models in cardiology. The different image modalities, for which often only limited training data is available, pose particular challenges. Due to the different domains, fields of application and sensor technologies, there are data sources that have similarities, but whose transfer between them can hardly be used. For example, training data can only be generated with great effort by (expensive) experts, e.g. for the pixel-precise segmentation of an image, but a new sensor technology cannot be trained with this data because the learned models do not generalize well enough. In this group, methods are to be developed to enable efficient domain transfer. For example, multi-task or self-supervised learning techniques will be used and Model Agnostic Meta Learning (MAML) will be implemented to learn representations from which very fast fine-tuning to a new domain (with far less data) is possible.