Abstract
The aim of the work described in this thesis was to advance the implementation of diffusion MRI in SVD research by exploring methods that improve its technical validity. The two key objectives were 1) to improve the validity of multicentre dMRI, by removing variability related to acquisition hardware in dMRI across sites with harmonization of the raw diffusion signal and 2) to improve the validity of dMRI-based network analyses, in the presence of pathology, by removing false positive connections with thresholding.
We addressed our first key objective in Chapters 2 and 5, where we showed proof of concept of effective harmonization with data from five cohorts of patients with SVD and controls, using rotation invariant spherical harmonic (RISH) features. In chapter 2 we showed that the RISH method removes acquisition-related differences in scalar dMRI metrics between matched controls of different sites, while preserving disease-related effect sizes (i.e., differences between patients and controls and associations between dMRI metrics and markers of SVD). Consequently, the harmonized data could be pooled to increase sample size and infer associations between dMRI metrics and markers of SVD with improved power. In chapter 5 we applied RISH harmonization beyond the scalar metrics explored in Chapter 2, by investigating if it also improves cross-site consistency of brain networks. We demonstrated that harmonization helps to achieve more similar network architectures across sites. Furthermore, harmonization facilitated data pooling to infer patterns of network injury with improved sensitivity as compared to single centre datasets.
We addressed our second key objective in chapters 3, 4 and 5 where we examined whether network thresholding increases consistency of brain networks in studies with cross-sectional design, longitudinal design and in multicentre analysis. In chapter 3 we proposed fixed-density thresholding as method to control for differences in network density (i.e., number of detected connections) when comparing networks of patients and controls. We showed that thresholding networks to a fixed density across all subjects preserves the original network architecture over a large range of thresholds and maintains the sensitivity to detected group-differences between patients with SVD and controls in global and local metrics. In this manner, networks can be compared across groups without the risk of measures being affected by density bias. In chapter 4, we tackled the effect of false positive connections causing low reproducibility of brain networks in longitudinal studies. We showed that weight-based thresholding improves scan-rescan network consistency in healthy young subjects (with relatively short-rescan inter¬vals) but also in patients with SVD scanned over long time periods. Importantly, we proved that thresholding preserves sensitivity to detect statistical group-differences between patients with low and high SVD burden. Finally, in chapter 5 we demonstrated that thresholding, in combination with RISH harmonization, helps to generate more consistent networks in multicentre data. In this analysis we showed that thresholding complements harmonization by reducing the number of false positives in the network, ultimately improving precision to detect patterns of network injury in multicentre data of patients SVD.
Altogether, the work presented in this thesis shows the benefits of dMRI harmoni¬zation by making data more comparable across centres and enabling data pooling to increase sample size. Our work also shows that advantages of network thresholding by improving precision of network analysis, paving the way for robust large scale dMRI network analysis in SVD.
We addressed our first key objective in Chapters 2 and 5, where we showed proof of concept of effective harmonization with data from five cohorts of patients with SVD and controls, using rotation invariant spherical harmonic (RISH) features. In chapter 2 we showed that the RISH method removes acquisition-related differences in scalar dMRI metrics between matched controls of different sites, while preserving disease-related effect sizes (i.e., differences between patients and controls and associations between dMRI metrics and markers of SVD). Consequently, the harmonized data could be pooled to increase sample size and infer associations between dMRI metrics and markers of SVD with improved power. In chapter 5 we applied RISH harmonization beyond the scalar metrics explored in Chapter 2, by investigating if it also improves cross-site consistency of brain networks. We demonstrated that harmonization helps to achieve more similar network architectures across sites. Furthermore, harmonization facilitated data pooling to infer patterns of network injury with improved sensitivity as compared to single centre datasets.
We addressed our second key objective in chapters 3, 4 and 5 where we examined whether network thresholding increases consistency of brain networks in studies with cross-sectional design, longitudinal design and in multicentre analysis. In chapter 3 we proposed fixed-density thresholding as method to control for differences in network density (i.e., number of detected connections) when comparing networks of patients and controls. We showed that thresholding networks to a fixed density across all subjects preserves the original network architecture over a large range of thresholds and maintains the sensitivity to detected group-differences between patients with SVD and controls in global and local metrics. In this manner, networks can be compared across groups without the risk of measures being affected by density bias. In chapter 4, we tackled the effect of false positive connections causing low reproducibility of brain networks in longitudinal studies. We showed that weight-based thresholding improves scan-rescan network consistency in healthy young subjects (with relatively short-rescan inter¬vals) but also in patients with SVD scanned over long time periods. Importantly, we proved that thresholding preserves sensitivity to detect statistical group-differences between patients with low and high SVD burden. Finally, in chapter 5 we demonstrated that thresholding, in combination with RISH harmonization, helps to generate more consistent networks in multicentre data. In this analysis we showed that thresholding complements harmonization by reducing the number of false positives in the network, ultimately improving precision to detect patterns of network injury in multicentre data of patients SVD.
Altogether, the work presented in this thesis shows the benefits of dMRI harmoni¬zation by making data more comparable across centres and enabling data pooling to increase sample size. Our work also shows that advantages of network thresholding by improving precision of network analysis, paving the way for robust large scale dMRI network analysis in SVD.
Original language | English |
---|---|
Awarding Institution |
|
Supervisors/Advisors |
|
Award date | 16 Nov 2022 |
Publisher | |
Print ISBNs | 978-94-6458-678-7 |
DOIs | |
Publication status | Published - 16 Nov 2022 |
Keywords
- Diffusion MRI
- Cerebral small vessel disease
- brain connectivity
- harmonization
- thresholding