2025 Theses Doctoral

Innovative Methods for Quantitative Medical Image Analysis in High Dimensional Spaces: Applications in Brain MRI and Lung CT

Arabshahi, Soroush

Quantitative medical image analysis plays a pivotal role in modern clinical practice by providing objective, reproducible, and detailed characterization of complex diseases.

This dissertation introduces novel computational and analytical frameworks to address critical challenges in two important biomedical imaging domains: functional brain imaging, specifically resting-state functional magnetic resonance imaging (rs-fMRI), and pulmonary emphysema subtyping using lung computed tomography (CT).

In the first part of this thesis, we introduce innovative methods to capture and characterize the spatiotemporal dynamics of resting-state fMRI data, addressing limitations inherent in traditional static functional connectivity analyses. Initially, we assess conventional static resting-state functional connectivity approaches in a cohort of mild traumatic brain injury (mTBI) patients.

The findings highlight the subtlety and complexity of connectivity disruptions, motivating the development of more sophisticated analytical methods capable of capturing dynamic brain function. Thus, we propose a novel technique, the dynamic functional Orientation Distribution Function (dfODF), designed to quantify local and transient Blood Oxygen Level-Dependent (BOLD) signal propagation at the voxel level. From the dfODF, we derive two new metrics: Generalized Fractional Anisotropy (GFA) for characterizing temporal anisotropy in signal propagation and the divergence-based Functional Propagation Directivity (FPD), which identifies regions acting as transient sources and sinks of neural activity. Rigorous validation across independent cohorts (NYU Adult mTBI Cohort and Midnight Scan Club dataset) demonstrated the reproducibility and robustness of these metrics. The dfODF framework provides a new toolset to study brain function as measured by BOLD fMRI for future studies on neurological and psychiatric conditions.

The second major contribution of this thesis addresses the clinical challenge of emphysema subtyping in chronic obstructive pulmonary disease (COPD). Traditional radiological classification methods—centrilobular, panlobular, and paraseptal emphysema—are subjective, variable, and limited in their clinical applicability. Recent advancements have proposed unsupervised computational methods yielding spatially-informed lung texture patterns (sLTPs), which revealed reproducible CT-based emphysema subtypes (CTES). However, such methods suffered from scalability limitations and computational complexity. To overcome these challenges, we introduce supervised deep-learning frameworks, specifically a supervised Deep Convolutional Disentangled Autoencoder (sDCDA) for accurate representation learning at the region-of-interest (ROI) level, and a multi-view aggregated 2.5D Residual U-Net (2.5D ResU-Net) for rapid and precise voxel-level segmentation of emphysema subtypes across entire lung CT volumes. Extensive experiments conducted on large datasets (SPIROMICS and MESA) validated the effectiveness, robustness, and clinical applicability of these models. The sDCDA achieved high classification accuracy and generated meaningful latent representations that facilitate future explorations of genetic, clinical, and environmental correlates. The 2.5D ResU-Net achieved rapid, accurate, and scalable subtype segmentation, substantially reducing computational processing time compared to previous methods.

Together, the methods developed and validated in this thesis offer advancements in quantitative medical image analysis over traditional approaches and set a foundation for enhanced clinical applications. The dynamic BOLD fMRI metrics provide promising new biomarkers for understanding and diagnosing subtle disruptions in brain function, while the deep-learning-based emphysema subtype classification models significantly improve the precision, reproducibility, and scalability of pulmonary imaging biomarkers. Future applications of these methodologies have broad implications across numerous neurological, psychiatric, and pulmonary diseases, promoting personalized medicine and improving patient outcomes through advanced quantitative imaging.