Adaptive processing of thin structures to augment segmentation of dual-channel structural MRI of the human brain
This thesis presents a method for the segmentation of dual-channel structural magnetic resonance imaging (MRI) volumes of the human brain into four tissue classes. The state-of-the-art FSL FAST segmentation software (Zhang et al., 2001) is in widespread clinical use, and so it is considered a benchmark. A significant proportion of FAST’s errors has been shown to be localised to cortical sulci and blood vessels; this issue has driven the developments in this thesis, rather than any particular clinical demand. The original theme lies in preserving and even restoring these thin structures, poorly resolved in typical clinical MRI. Bright plate-shaped sulci and dark tubular vessels are best contrasted from the other tissues using the T2- and PD-weighted data, respectively. A contrasting tube detector algorithm (based on Frangi et al., 1998) was adapted to detect both structures, with smoothing (based on Westin and Knutsson, 2006) of an intermediate tensor representation to ensure smoothness and fuller coverage of the maps. The segmentation strategy required the MRI volumes to be upscaled to an artificial high resolution where a small partial volume label set would be valid and the segmentation process would be simplified. A resolution enhancement process (based on Salvado et al., 2006) was significantly modified to smooth homogeneous regions and sharpen their boundaries in dual-channel data. In addition, it was able to preserve the mapped thin structures’ intensities or restore them to pure tissue values. Finally, the segmentation phase employed a relaxation-based labelling optimisation process (based on Li et al., 1997) to improve accuracy, rather than more efficient greedy methods which are typically used. The thin structure location prior maps and the resolution-enhanced data also helped improve the labelling accuracy, particularly around sulci and vessels. Testing was performed on the aged LBC1936 clinical dataset and on younger brain volumes acquired at the SHEFC Brain Imaging Centre (Western General Hospital, Edinburgh, UK), as well as the BrainWeb phantom. Overall, the proposed methods rivalled and often improved segmentation accuracy compared to FAST, where the ground truth was produced by a radiologist using software designed for this project. The performance in pathological and atrophied brain volumes, and the differences with the original segmentation algorithm on which it was based (van Leemput et al., 2003), were also examined. Among the suggestions for future development include a soft labelling consensus formation framework to mitigate rater bias in the ground truth, and contour-based models of the brain parenchyma to provide additional structural constraints.