Rotation 1: Biological plausible models of visual perceptual learning

点击翻译按钮获取中文摘要

BBSRC strategic theme: Transformative technologiesThe objective of this PhD is to implement normative modelling in the field of connectomics with the aim of developing an AI model that can help in the diagnosis of neuropsychiatric disorders. Traditional MRI analysis involves registration and parcellation of scans, however the atlases and references used come from healthy brain models. When the scans are registered to these atlases, clinically relevant differences in certain regions can be masked. We will overcome this by implementing normative models. Normative modelling is a framework used to study differences between an individual and a control group (Marquand, 2016). It has already been implemented on structural MRI data to study heterogeneity in brain development through age (Bethlehem et al, 2022) and different neurodevelopmental conditions (Bedford et al, 2023). We propose a large-scale normative model derived from thousands of individuals that clinical studies can have access to, overcoming the issue with small sample sizes often encountered in these studies. We will incorporate this method to the field of functional MRI (fMRI) to study the link between heterogeneity in connectivity and disease. Many studies have shown that mental disorders occur via changes in brain structure and function, therefore understanding differences in an individual's connectivity can inform us of potential disorder progressions. So far, connectivity has been studied using graph theory approaches that assume the same brain locations are used across individuals, posing a harsh constraint on any analysis. Geometric deep learning is a new tool that can be used to analyse non fixed topological structures. Graphical neural networks (GNNs) represent a form of geometric deep learning tailored for data with a graph-like structure. They have already been used to analyse resting state fMRI data capturing significant features (Azevedo et al. 2022). We will use similar geometric deep learning approaches like graph-based networks and point cloud models to construct our model with the aim of encoding the brains architecture without the need of registration or parcellation. We hope to capture deviations in patterns of brain connectivity at the individual level that provide diagnostic information. This would take us one step further in understanding and diagnosing neuropsychiatric disorders.