Pathological Maps Based on T1 Relaxometry: Robust Estimation and Clinical Validation

Background and Objective: Quantitative MRI (qMRI) is an essential tool to quantify brain changes due to inflammation, degeneration and repair in multiple sclerosis (MS). However, how thesequantified changes in the microstructure of the brain affect the individual patient and if they areclinically relevant remains an open question. This master thesis attempts to answer this questionby: (1) Making the method developed by Bonnier et al., 2019 more robust (i.e. applying it to moresubjects), (2) Correlate the measured changes in brain tissue with the Expanded Disability StatusScale (EDSS) to ascertain the statistical significance and therefore the clinical relevance, and (3)making the method developed by Bonnier et al., 2019 more lightweight and unified so it can be implemented to support clinical decisions.
Methods: To compute the reference distribution of qMRI metrics a group consisting of 80 healthy cohort was used. To obtain the qMRI metrics, all subjects underwent 3T MRI examinations including T1 relaxation, Magnetization Prepared 2 Rapid Acquisition Gradient Echoes (MP2RAGE) and High-Resolution 3D Fluid Attenuated Inversion Recovery (FLAIR) imaging. The Siemens Morphobox was used for brain segmentation and for tissue probability estimation. For 34 patients,their personal pathological maps (deviation or z-score maps) were computed. The selected parameters for the prediction of the EDSS score are: age, average z-score within the lesions, average z-scores within the healthy brain tissue, number of lesions in total, and number of voxels within lesions. A generalized linear model with backwards selection with was applied for the prediction of EDSS scores.
Results and Conclusion: The average z-scores within MS lesions in all 34 patients is 4.14 _ 1.06 [mean _ SD] and the average z-scores outside of the lesions, in the normal appearing tissue, is 0.51 _ 0.40 [mean _ SD]. The generalized linear model with backwards selection determined the most significant model. The parameters of this model are: age, average z-score within the lesions, and number of voxels within lesions. The resulting model has a p-value of 0.000837 and therefore is statistically significant.