Modeling Multi-Site Neuroimaging Data Using Hierarchical Bayesian Neural Networks to Study Structural Brain Variability

Huijsdens, Hester

Modeling Multi-Site Neuroimaging Data Using Hierarchical Bayesian Neural Networks to Study Structural Brain Variability

Files

S4480562_HuijsdensH.pdf (1.68 MB)

Authors

Huijsdens, Hester

Issue Date

2020-10-01

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/10632

Abstract

Brain disorders are heterogeneous in their nature, therefore making it problematic to di er- entiate between multiple brain disorders using a traditional case-control approach. In contrary to case-control studies, where subjects are being classi ed into distinct groups, normative modeling can be used to model the variation in a large population. Normative modeling is a statistical method aiming to learn both the mean of a response variable and its normative range, therefore being able to handle the heterogeneous nature of highly variant groups. How- ever, large amounts of neuroimaging data are required for normative modeling. Hence, datasets coming from di erent imaging centers need to be combined, introducing site-related variance generated by among others variability in acquisition devices across centers. We must account for this variability in the analysis, as these site-related variations are usually larger than those that we are interested in when trying to detect individual outliers from the normative range. Hierarchical Bayesian regression (HBR) models have been applied for multi-site normative modeling previously. However, it was limited in that they only had the options to t either a linear, or a simple parametric, form. To allow for more exibility and make less assumptions about the data, we will extend the HBR approach with a hierarchical Bayesian neural net- work. Using three simulated datasets, we show the model's ability to handle site-e ects when the estimated e ect is non-linear in both the mean and variance, while being computationally scalable. Moreover, the model can also deal with, and hence does not over t on, sites with limited data. Although the estimates t well to the data, the parameter sampling is unstable. In future studies, this is something that would need to be solved rst, for example by assuming di erent priors or experimenting with di erent chain initialization methods. Finally, we tted and evaluated the model on image-derived phenotypes from the UK Biobank dataset. In this test case, most brain measures showed a linear trend in their mean. Thus, we have observed limited improvements of our model compared to the previously described linear HBR model. However, when investigating the estimated variance, the HBNN showed a good performance in estimating a non-linear heteroscedastic variance in the data. Nevertheless, this non-linearity in the variance was minor, resulting in limited improvements of the neural network in comparison to the linear HBR model. Nevertheless, our model is able to model the mean and variance in datasets, also when there is non-linearity in the mean or variance of the data, or when the variance follows a skew Gaussian distribution. Compared to the linear HBR model, our model o ers additional exibility to handle site-related variation and non-linearity in neuroimaging datasets. Despite the limited improvements that we found, we expect that using a dataset containing more non-linearity in the response variable would allow us to di erentiate between both models to a greater extent.