TY  - THES
AU  - Kröll, Jean-Philippe
TI  - Interpretability and Reliability in Neuroimaging
PB  - HHU Düsseldorf
VL  - Dissertation
M1  - FZJ-2025-04110
SP  - 91
PY  - 2025
N1  - Dissertation, HHU Düsseldorf, 2025
AB  - The development of magnetic resonance imaging (MRI) based biomarkers is a constant endeavorin the field of clinical neuroscience. Although these biomarkers hold great potential, only few havebeen adopted for routine clinical use. Primary challenges for the translation into clinical use areaccuracy, reliability and interpretability of a given biomarker. Consequently, this dissertationpresents a new machine learning (ML) framework that improves accuracy of diagnosis andprognosis of one of the most common neurological diseases, AlzheimerÕ Disease (AD), byconstructing complex representations of base features. Further, by using a context-free grammar(CFG), the constructed representations are forced to remain humanly interpretable, thus enablingthe validation of a relationship between the biomarker and the supposed underlying pathologiccorrelate. Additionally, it is investigated if naturalistic viewing (NV) paradigms are suited toimprove characteristics of MRI measurements that are important for biomarker development, suchas reliability, reduced intra-subject variability and enhanced detection of individual differences, incomparison with resting-state (RS). Therefore, the effect of NV stimuli with varying levels of socialcontent and different lengths is investigated in 14 functional brain networks. It is shown that, basedon network functional connectivity (NFC), NV stimuli improve the detection of individualdifferences in 10 out of 14 networks, with the stimuli with the highest level of social contentachieving the most improvement. A subsequent analysis confirms that movie stimuli with higherlevels of social content evoke similar NFC patterns that are distinct from RS and a stimulus lackingsocial interactions. Further, it is demonstrated that NV stimuli can reduce intra-subject variabilityin meta-analytic networks that are essential for perception and processing of action, behavior andemotions. In addition, it is shown that NV stimuli can increase the reliability of graph metricsextracted from NFC, over RS. However, the results also emphasize that NV stimuli do notunconditionally improve metrics of interest across the whole brain. In particular for networks thatare related to intrinsically oriented functions, RS proves to be the more favorable paradigm.Therefore, selecting the appropriate stimulus and functional network is essential for addressing thespecific research question at hand. Finally, this dissertation provides a new publicly available NVdataset to further analyze the effect of NV stimuli.
LB  - PUB:(DE-HGF)11
UR  - https://juser.fz-juelich.de/record/1047084
ER  -