Machine learning in Clinical Neuroimaging
This lab analyses neuroimaging data in neurological and psychiatric diseases using advanced machine learning methods.
You are here:
In this lab, we use methods of artificial intelligence to diagnose and characterize neurological and psychiatric diseases based on neuroimaging data. Whereas previous disease decoding approaches mostly relied on expert-based extraction of features in combination with standard classification algorithms and thus strongly depend on the choice of data representation, deep learning approaches, in particular convolutional neural networks, are capable of learning hierarchical information directly from raw imaging data. By this, they have a great potential for finding unexpected and latent data characteristics and might perform as a real "second reader".
The research of this lab is supported by the DFG, the Charité Nachwuchsförderung, the Brain & Behavior Research Foundation, the DMSG, the Ursula und Manfred Müller-Stiftung and a NVIDIA GPU grant. This lab is associated with the Bernstein Center for Computational Neuroscience Berlin.
Deep learning for multiple sclerosis
Multiple sclerosis (MS) is an autoimmune disease of the central nervous system and often leads to substantial disability in patients. In this project, we use convolutional neural networks to diagnose MS and predict individual disease courses based on T2-weighted FLAIR images that are particular sensitive to lesions in MS. In particular, we look into differences between normal-appearing and lesional areas. This is a joint project with Prof. Dr. Friedemann Paul and PD Dr. Michael Scheel (NeuroCure Clinical Research Center). In an additional project, we aim to develop novel techniques for the identification of new MRI- derived clinically applicable biomarkers for the complex task of personalised prediction of disease activity in MS patients (please see the project webpage deepMS). We are looking here for potential scientific and data collaborations!
Deep learning for Alzheimer’s disease
Alzheimer’s disease is a neurodegenerative disease characterized in later stages by dementia. Even for experienced neuroradiologists the discrimination between atrophy mediated by Alzheimer’s disease and normal age-related atrophy is difficult. In contrast, computer programs can relatively easy catch up these differences. Since the clinical interpretation is of major importance in disease diagnostics and convolutional neural networks are usually considered as "black-box" models, we compare here different state-of-the-art visualization techniques including layer-wise relevance propagation. These methods allow for generating heatmaps that indicate for each individual patient the relevance for a particular classification decision. These analyses are based on T1-weighted MRI data of the open data base ADNI.
Funded by the DFG and the Ursula und Manfred Müller-Stiftung.
Machine learning for psychiatric diseases
In this research project we investigate whether machine learning methods help us to find subtle differences in neuroimaging data of patients with psychiatric diseases. For example, we took part in the PAC Depression Challenge 2018, in which structural MRI data had to be classified into patients with and without depression (in cooperation with Prof. Dr. Henrik Walter and Prof. Dr. Sebastian Stober). Together with Prof. Dr. Norbert Kathmann we analyse data of patients with obsessive-compulsive disorder. If you are interested in joint projects here, please contact us.
Studying mental health via research domain criteria, neuroimaging and convolutional neural networks
Due to a large heterogeneity in clinical symptoms and biological measures, the characterization of subjects regarding mental health is difficult. Therefore, it has been suggested to study mental health along the so-called Research Domain Criteria (RDoC) rather than clinical labels per se. In this project, we want to investigate mental health along the RDoC, e.g. in terms of neuroticism or happiness, based on neuroimaging data of the UK Biobank. We will compare deep learning analyses with standard machine learning analyses using predefined features such as cortical thickness or functional connectivity. This is a joint project between Prof. Dr. Dr. Henrik Walter and the research unit Mind and Brain, Prof. Dr. Andre Marquand and Prof. Dr. John-Dylan Haynes.
Funded by NARSAD Young Investigator Grant (Brain & Behavior Research Foundation).
Autoencoders, transfer learning and normative modelling
In this project we use autoencoders to learn representations from neuroimaging data, which can then be used to model deviations of certain populations (e.g. young vs. old, healthy vs. sick) in a normative modelling approach. For these analyses we use data of the UK Biobank and the Human Connectome Project. We methodologically cooperate here with Prof. Dr. André Marquand and Prof. Dr. Sebastian Stober.