Mapping Cardiac Adipose Tissue Using Spectral Analysis of Ultrasound Radiofrequency Data and Machine Learning

Author

Julian Rene Cuellar Buritica, Southern Illinois University CarbondaleFollow

Date of Award

12-1-2025

Degree Name

Doctor of Philosophy

Department

Engineering Science

First Advisor

Klingensmith, Jon

Second Advisor

Qin, Jun

Abstract

Cardiovascular disease (CVD) remains the leading cause of mortality worldwide, accounting for over 20 million deaths annually. Among these, coronary artery disease (CAD) is responsible for approximately 9.44 million deaths in 2023 alone, and its prevalence is expected to increase from 48% in 2022 to over 50% in 2030. With CVD affecting nearly one in two individuals in the United States, there is an urgent need for accessible, non-invasive diagnostic tools to support early detection and risk stratification.Cardiac adipose tissue (CAT), the visceral fat surrounding the myocardium, has emerged as a promising biomarker for both CVD and CAD. Elevated CAT depositions near the coronary arteries is linked to plaque formation and adverse cardiovascular events. Although magnetic resonance imaging (MRI) offers accurate visualization of CAT, its high cost and limited availability restrict its widespread clinical use. In contrast, standard echocardiography or ultrasound (US) is more accessible and cost-effective but currently limited to linear measurements of CAT on the right ventricular free wall, lacking the capability for comprehensive assessment. This dissertation investigates the hypothesis that spectral analysis of raw US radiofrequency (RF) data can be used to identify and map CAT in echocardiograms. To test this hypothesis, we propose a fully automated, low-cost imaging framework that integrates RF data analysis with machine learning (ML) and deep learning (DL) techniques. The approach enables segmentation of epicardial contours from parasternal short-axis (PSAX) US images, classification of CAT regions using spectral features, and mapping of CAT distribution around the heart’s perimeter. To validate the segmentation component, state-of-the-art DL models—including the Segment Anything Model (SAM), Detectron2, and U-Net with ResNet backbones—were evaluated. These models achieved an average Dice Similarity Coefficient of 0.82 and a Hausdorff Distance of 4.93, indicating strong performance in delineating epicardial boundaries. For CAT mapping, ground truth labels were derived by intersecting 3D heart models reconstructed from MRI data with corresponding US scans. The heart perimeter was divided into regions of interest (ROIs), which were labeled based on varying CAT thresholds from MRI-derived maps. ML classifiers were trained to predict CAT presence in each ROI using spectral parameters extracted from the RF data. The highest classification precision of 74.22% was achieved using RUSBoosted tree ensembles at a CAT threshold of 85%. Our findings demonstrate that DL models are effective for segmenting left and right ventricular contours in PSAX echocardiograms. Specifically, models trained on domain-specific datasets, such as U-Net with ResNet, outperform general-domain models when applied to small, locally acquired datasets. Furthermore, the proposed CAT mapping approach significantly enhances the identification of CAT in regions not typically visible in grayscale US images. By leveraging spectral information, this method enables a more comprehensive assessment of CAT using US alone. This research contributes a radiation-free, MRI-independent technique for CAT quantification, making it suitable for patients with metal implants (e.g., pacemakers, joint replacements) and those in resource-limited settings. Ultimately, the proposed framework supports improved risk stratification for heart attacks and lays the groundwork for future studies exploring the relationship between CAT distribution and coronary plaque characteristics in CAD progression.