Detection of cerebral small vessel disease signs on brain MR images

Thesis

Introduction: Accurate and automated detection of brain disease signs could aid in studying their clinical associations in large populations. This research aims to detect on magnetic resonance images two signs of cerebral small vessel disease- white matter hyperintensities (WMHs) and cerebral microbleeds (CMBs)- and to analyse their spatial distribution and characteristics.

Methods: We modelled the distribution of WMHs at the population-level with respect to various parametric factors (e.g. age) using Bayesian inference. For improving the performance of FSL's WMH segmentation tool, BIANCA, we used the above population-level WMH distribution probabilities as an additional feature, tested different classiers and proposed a locally adaptive thresholding method (LOCATE). Furthermore, we developed a triplanar U-Net ensemble model (TrUE-Net), based on deep learning (DL), spatially-weighted cross-entropy and Dice loss functions to improve WMH segmentation, and explored various domain adaptation techniques. For CMBs, we developed an end-to-end analysis: CMB candidate subject preselection pipeline using machine learning, CMB detection using DL and atlas-based automated CMB rating.

Results and conclusions: Our distribution modelling provided comparable error values, with lower computational load, with respect to existing methods on simulated data. On real data, periventricular WMH probabilities showed signicant increases with age, while the probabilities for deep WMHs were higher in hypertensive subjects, in line with literature. Among the proposed improvements for BIANCA, LOCATE gave the highest performance increase on 7 datasets with dierent lesion characteristics. TrUE-Net performed better than BIANCA on 5 dierent datasets, especially on low lesion load subjects. The CMB preselection pipeline provided robust performance despite dierent training modalities. The CMB detection method provided performance comparable with existing methods and higher precision than the state-of-the-art method. Future directions for this work could include improving the DL-based CMB/WMH detection methods and studying the clinical impact of CMBs at the population-level.

Authors: Sundaresan, V

• DOI: 10.5287/ora-7qbvxd8p8
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: structural MRI; cerebral microbleeds; small vessel disease; white matter hyperintensities; machine learning; image analysis; deep learning; segmentation