Jeonghoon Park
BME MS Thesis Defense Presentation
Date: 2026-04-02
Time: 1:00PM
Location / Meeting Link: https://teams.microsoft.com/meet/28289867684066?p=PEFDDCO2AcOlZGDWn6

Committee Members:
Dr. Mason Borzin; Dr. Shella Keilholz; Dr. John Oshinski


Title: STANDARDIZING ENTROPY MEASURES TO CHARACTERIZE BRAIN DYANMICS IN FMRI

Abstract:
Entropy-based methods are increasingly used to quantify neural complexity, yet their application to functional magnetic resonance imaging (fMRI) remains inconsistent due to substantial variability in preprocessing, parameter selection, and reporting practices. This thesis investigates whether entropy analysis can be standardized and reliably employed in fMRI to characterize brain dynamics and discriminate cognitive states. First, a systematic literature review following PRISMA 2020 guidelines was conducted to assess how entropy has been applied to BOLD fMRI data. Across 203 studies in healthy populations, the review identified major methodological heterogeneity and a lack of consensus in entropy implementation, limiting reproducibility and cross-study comparability. Motivated by these findings, a standardized entropy-analysis framework was developed and evaluated using Human Connectome Project resting-state fMRI data. Four entropy measures were examined: sample entropy, multiscale entropy, Shannon entropy, and transfer entropy. Sample entropy and multiscale entropy revealed biologically meaningful, network-specific differences in resting-state dynamics, while transfer entropy captured directed information flow consistent with known large-scale brain organization. In contrast, Shannon entropy primarily reflected signal-distribution regularity and showed limited biological specificity. Comparisons with conventional fMRI measures, including static functional connectivity and sliding-window correlation, demonstrated that entropy-based metrics provide complementary, non-redundant information. Split-half analyses further showed high reliability, particularly for sample entropy, multiscale entropy, and transfer entropy. Finally, the standardized framework was applied to brain-state classification using resting-state and task-based fMRI from the Human Connectome Project. Using only sample entropy- and transfer entropy-derived features, a convolutional neural network classified eight cognitive states with strong performance, achieving approximately 81.8% accuracy using Yeo7 networks and 91.1% accuracy using Yeo17 networks, with macro AUC values approaching 0.97 and 0.99, respectively. These findings demonstrate that entropy-based features encode highly discriminative task-specific signatures of brain dynamics. Overall, this work establishes entropy as a reliable and interpretable analytic framework for fMRI and supports its broader use in studying large-scale brain organization, cognitive-state differences, and future clinical applications.