Overcoming Bias and Data Inhomogeneity in Machine Learning: COVID-19 Imaging and Beyond

• Applicants: Professor Florian Bernard, Ph.D.; Professor Dr. Daniel Cremers
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Medical Informatics and Medical Bioinformatics
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 458610525

In recent years machine-learning techniques enabled significant advances in the field of medical image analysis. They promise a coherent, fast, and scalable way for diagnosis and prognosis of diseases, such as COVID-19, from radiological images. Nevertheless, the major road-blocks for a more systematic deployment of these techniques in clinical settings are bias (e.g. due to imbalanced population features) and data inhomogeneity (e.g. different data modalities). They lead to shortcomings regarding the universality and generalisation of machine learning models and thus limit the translation of scientific results into widely applicable clinical tools. In the proposed project, we will develop novel machine learning approaches to overcome these obstacles: (A) Confounder mining: a framework for the direct identification of confounders that induce bias in training machine learning models. (B) Large-scale learning from inhomogeneous data: deep learning methods to improve universality and generalisation by enabling training from massive cross-centre and cross-modality datasets. (C) Domain adaptation and transfer learning: systematisation of design patterns for neural network topologies for different sub-fields of domain adaptation and transfer learning to deal with biased data. As such, bias and data inhomogeneity will be tackled from three complementary angles that could be integrated into one stratified solution: bias identification in (A), adaptation to inhomogeneity in (B), and adaptation to bias in (C). By doing so, the negative effects of bias and inhomogeneity will be reduced substantially, enabling new potentials to bring machine-learning methods into the clinic at scale.

• DFG Programme: Research Grants
• International Connection: Luxembourg
• Partner Organisation: Fonds National de la Recherche
• Cooperation Partner: Andreas Husch, Ph.D.