Master thesis
Investigating the Relationship Between the Morphometric Data of MS Patients and a Selected Cognitive Game
Master’s Thesis by Aaisha Bah
Multiple sclerosis (MS) is a disease of the central nervous system that is characterized by demyelination and neuronal damage. Volumetric MRI changes in MS patients over the disease progression show a strong correlation with cognitive performance, which can be assessed using traditional neuropsychological evaluations and digital tools. Researchers at the Research Center for Clinical Neuroimmunology and Neuroscience (RC2NB) developed, Numbers, a cognitive game of number assortment, designed to measure information processing speed.
This thesis investigates the correlation between performance features derived from the Numbers game, traditional cognitive assessments, and volumetric measurements of 25 brain regions in 153 patients. Deep learning models were used to predict the game features and stacked ensemble models were developed to provide patient-level predictions.
The findings show that deep grey matter structures such as the putamen, thalamus, and pallidum, along with the cerebral cortex volume are strongly associated with standard cognitive assessments and key game features like the number of successful touches and completion time, which are both indicators of the patients’ cognitive performance. The ensemble models are able to effectively learn from patient specific features to provide improved final predictions.
Predicting Remyelination in Multiple Sclerosis Patients using Deep Learning
Master Thesis by Tejeswini Jayakumar
Multiple Sclerosis (MS) is a chronic inflammatory disease that targets the Central Nervous System whose effects can be seen as demyelination and axonal/neuronal damage. Previous studies by Rahmanzadeh et al. (2) have demonstrated that remyelination can be detected on Quantitative Susceptibility Mapping (QSM) and that monitoring remyelinated lesions might serve as a biomarker for disease progression.
Therefore, the aim of this Master’s thesis was to distinguish between remyelinated lesions and non-remyelinated lesions in patients affected by MS with the help of deep learning (DL). The objective extended beyond mere differentiation; it sought to predict if a lesion would undergo remyelination in the future. This objective was achieved by using retrospective advanced MRI in Imaging the Interplay Between Axonal Damage and Repair in Multiple Sclerosis (INsIDER) dataset (2) and clinical MRI in the Swiss Multiple Sclerosis Cohort (SMSC) dataset (3). The ground truth remyelinated masks were segmented on the QSM from the advanced MRI. Data from the two closest time points in the SMSC dataset separated by 5-17 months were utilized to help the network detect remyelination over time.
The DL method developed in this study was based on UNEt TRansformers (UNETR). Upon validating the results of the lesion segmentation, it became evident that using UNETR in its original segmentation configuration was not optimal for our specific needs. Consequently, we adapted the network for binary classification to distinguish between remyelinated and non-remyelinated lesions.