Segmentation of substantia nigra and red nucleus in MRI images : application to Parkinson’s disease. (2021)
Type of ContentTheses / Dissertations
Thesis DisciplineComputer Science
Degree NameDoctor of Philosophy
PublisherUniversity of Canterbury
Accurate segmentation of substantia nigra (SN) and red nucleus (RN) is challenging, yet important for understanding health problems like Parkinson’s disease (PD). This thesis proposes robust algorithms to segment SN and RN from T2*-weighted images and quantitative susceptibility mapping (QSM) MRI. After optimising the algorithms to produce reliable SN and RN segmentations, multiple MRI metrics extracted from the SN and RN are used to investigate association with cognitive and motor function in PD.
For T2*-weighted images, an adaptive k-means algorithm and a combination of level set method and wavelet transform are proposed for accurate segmentation of substantia nigra. The latter method is further improved with substantia nigra seeding, contrast enhancement and watershed transform to provide efficient segmentations of substantia nigra and red nucleus in QSM images. The algorithm-derived segments of SN and RN are compared to expert manually derived (‘gold standard’) segmentations using the Dice similarity index for quantitative assessment.
The QSM values extracted from manually drawn SN and RN are compared to values extracted from the proposed segmentations in a subsample of 40 participants. Additionally, QSM values obtained from both ‘gold standard’ and proposed segmentation method are compared between groups (PD and healthy controls), and the relationship with global cognitive ability and motor severity in PD are investigated using Bayesian regression models.
The proposed segmentation algorithms produce effective segmentations of SN and RN from T2*-weighted images and QSM images, validated against expert manual segmentation. This thesis shows moderate evidence of increased QSM values in SN in PD relative to healthy controls, with moderate evidence for association between QSM, global cognitive ability, and motor impairment in the SN in PD. RN QSM values also exhibit moderate evidence for association with motor impairments in PD.
After validating the proposed segmentation algorithm, SN and RN are segmented in a large cohort of 127 PD participants and 47 matched controls. Multiple MRI metrics are extracted from the SN and RN regions from different modalities and include QSM, R2*, diffusion tensor imaging (DTI) parameters, perfusion, and volume. The MRI metrics are investigated independently to assess their relationship with group (PD vs controls), global cognitive ability, and motor impairment, again using Bayesian regression models.
This thesis shows strong evidence of increased QSM, R2*, MD and RD values, and moderate evidence of decreased FA values in SN, moderate evidence of increased QSM, R2* and MD values in RN in PD relative to controls. SN MD values also demonstrate moderate evidence of association with motor impairments in PD.
Lastly, group status (PD/Control) is predicted based on multiple MRI metrics from the SN and RN. Four different classifiers are compared with 10-fold cross validation: linear support vector machine (SVM), non-linear SVM, logistic regression, and random forest.
In terms of predicting disease group, linear SVM gave the highest performance with area under the receiver operating characteristics curve (AUCROC) = 0.76 and area under the precision-recall curve (AUCPR) = 0.89, showing robust classification of disease based on MRI values from SN and RN.
When applying the proposed robust segmentation method to a cohort of healthy controls and PD patients, this thesis demonstrates interesting association between groups and clinical scores utilizing various MRI measurements from substantia nigra and red nucleus, suggesting that the combination of different MRI metrics may provide useful information about iron deposition, microstructural alterations, perfusion changes, and volume atrophy of the midbrain nuclei in Parkinson’s disease, and also may be helpful in distinguishing PD from controls. Therefore, we suggest that an efficient and improved midbrain segmentation algorithm may be utilized to monitor disease severity in Parkinson’s. Additionally, the segmentation method proposed in this thesis may also be applied to different scenarios and disease groups where SN and RN are of particular interest.