Connectivity mapping with resting state functional magnetic resonance imaging (rs-fMRI) is rapidly developing and has shown great promise for clinical applications. Before successful implementation in clinical setting, it is key to evaluate the long-term reproducibility of the functional connectivity profiles. To this end, the reproducibility of rs-fMRI data is studied in this work. The main research question revolves around the improvement of the overall reproducibility by selectively omitting single components (either nodes or elements) from BOLD rs-fMRI connectivity matrices (CM’s). The scans of the subjects are parcellated using four different schemes, which are all analysed throughout this work. A reproducibility study is carried out on a dataset of 37 subjects that are scanned twice within 2 weeks on average. The inter-subject intraclass correlation coefficient (ICC) is used to quantify component reproducibility within the dataset. An algorithm is designed to quantify which component has the lowest inter-subject ICC, which is then eliminated from all CM’s in the dataset. After every single component elimination, the intra-subject ICC is computed for every subject to quantify the reproducibility, and a matching accuracy (MA) test is performed on the set to quantify the distinctive power of the CM’s.

The order in which components are eliminated and its effect on the overall reproducibility is tested by applying this to a larger test set of longitudinal data. To this end, a dataset of 521 subjects is used to quantify the reproducibility of the CM’s after iteratively removing components in the order that is found in the reproducibility study. This larger dataset of 521 subjects is analysed, along with 4 subsets, namely: sex based, age based, interscan time based and based on the grounds for exclusion. The latter is a subset where the quality of the rs-fMRI scans could not be assured due to pathologies or excessive motion during image acquisition. No significant difference is found within the sex-based subsets, and no relation between the reproducibility and the interscan time (within the range that is assessed in this work, namely 5 years) is found. Significantly lower intra-subject ICC’s are found for the subjects whose scan quality was subpar, due to excessive motion or pathology. For the age-based subset analysis, it is reported that reproducibility decreases with age.

The node removal algorithm clearly outperforms the element removal algorithm when looking at the intra-subject ICC. As the element removal algorithm can increase the intra-subject ICC by roughly 0.1, whereas the node removal algorithm manages to increase the intra-subject ICC of roughly 0.3. The MA, which is used as to quantify the distinguishing power between various CM’s, is seen to increase from 82.4% to the maximum of 98.7% correctly matched subjects for the RSS100 parcellation scheme within the reproducibility study. Aside from the element removal within the reproducibility study, the matching accuracy is not improved for any of the other analyses. The component elimination algorithm can increase the intra-subject ICC’s of the subjects of the longitudinal set. The MA is not found to increase with the component elimination algorithm.