Logo of the Physikalisch-Technische Bundesanstalt


Development and quality assessment of neuroimaging-based biomarkers for brain disorders


This project aims to advance the field of neuroimaging based biomarker identification for psychiatric diseases using multivariate machine learning methods by 1) developing standardized performance assessment protocols not only for prediction accuracy but also additional dimensions, 2) defining reference data for relevant prediction tasks in clinical neuroimaging, 3) developing novel multi-modal machine learning and deep learning methodologies that provide competitive performance in the above-mentioned quality dimensions.  


The success of deep learning architectures in solving difficult problems in imaging and natural language processing has raised expectations that similar approaches might be applicable to the (early) diagnosis and prognosis of neurological and psychiatric disorders based on neuroimaging data, which may pave the way for timely therapeutic interventions. However, the field must be considered still in its infancy. Published approaches are usually bound to single measurement modalities and evaluated on single datasets with moderate sample sizes. The majority of studies focuses on structural (e.g., T1-weighted MRI) data, while systems based on functional (e.g., fMRI, EEG or MEG) data are rare, as are multi-modal prediction models. Besides small sample sizes, the development of diagnostic markers is further challenged by the strong heterogeneity of mental disorders. Normative models of healthy persons are, thus, needed to derive meaningful notions of abnormality. Finally, there exists no agreed-upon protocol to benchmark neuro-biomarkers with respect to quality dimensions such as fairness, robustness, uncertainty calibration and interpretation, which are all crucial regarding prospective clinical applications. 


  • University of Zurich