Our brain is the most central part of the nervous system and crucial to our health and wellbeing. In Denmark, mental disorders make 25% of the total disease burden, with yet increasing prevalence. Across many medical domains, magnetic resonance imaging (MRI) is a widely used tool to study the anatomy and physiology of soft tissue, e.g. the brain, and is applied in diagnosis and monitoring of diseases, as well as a tool to investigate their underlying mechanisms. In psychiatry, research has yielded substantial evidence for structural brain changes at a group level, however these are typically subtle and currently, there is no clinical benefit from MRI for the individual patient. Previous research aiming to identify brain aberrations in patients with neuropsychiatric disorders struggle with relatively small and often inhomogeneous samples paired with complex clinical traits and weak pathological signals.
To unravel the intricacy of mental disorders and the brain, one approach is to apply powerful state-of-the-art machine learning algorithms, such as deep neural networks (DNNs). Large DNNs are able to extract high level features of images and other signals and able solve sophisticated tasks, outperforming traditional machine learning methods by far. On the downside, a DNN requires a large amount of ‘labelled data’, e.g., where each brain image has a meaningful notation, such as ‘patient’ or ‘control’. The amount of labelled data needed to train such a model is currently not available in conventional psychiatric research. In contrast, ‘unlabelled data’, e.g. normative brain images independent of a certain class or group, are often generously and publicly available. In the field of machine learning, scarcity of labelled data and richness of unlabeled data has given rise to self-supervised learning paradigms. In self-supervised learning one exploits rich unlabeled data to learn a general intermediate representation of the matter of interest. Scarce labelled data is used efficiently to fine-tune the intermediate representation to a specific task of interest.
The aim of this project is to develop a self-supervised DNN model of the brain using MRI data of large international, high-quality databases. This model will then be fine-tuned using highly specific data from psychiatry to address specific research questions regarding the mechanisms of aberrant neurodevelopment. We believe a self-supervised model is more robust and able to learn more meaningful features compared to conventional models. To explore these features and their relation to clinical traits present in psychiatric patients, we want to employ explainable artificial intelligence techniques.
To sum up, we want to use self-supervised learning paradigms and utilize its efficient use of scarce labelled data to develop a state-of-the art DNN model of brain images, bringing neuropsychiatric research to the forefront of machine learning research.