Mapping adipose tissue in short-axis echocardiograms using spectral analysis.

Area of Science:

• Biomedical Engineering
• Cardiovascular Imaging
• Machine Learning in Medicine

Background:

• Cardiovascular disease (CVD) is a leading cause of death, with cardiac adipose tissue (CAT) buildup identified as a key biomarker.
• Magnetic Resonance Imaging (MRI) is the gold standard for CAT imaging but is costly and inaccessible.
• Ultrasound offers a portable and affordable alternative, but image noise hinders accurate CAT identification and quantification.

Purpose of the Study:

• To develop a machine learning classifier utilizing spectral analysis of raw radiofrequency (RF) ultrasound data to automatically detect cardiac adipose tissue (CAT).
• To improve upon previous random forest models by optimizing region-of-interest (ROI) selection parameters and incorporating anatomical location data.

Main Methods:

• Employed spectral analysis of raw RF ultrasound data as input for a machine learning classifier.
• Experimented with varying ROI parameters (circumference, width, CAT thickness threshold, signal level) and anatomical location (distance from myocardium intersections).
• Utilized MRI-labeled ROIs from the same patients for ground truth classification.

Main Results:

• Achieved a classification accuracy of 75.5% by incorporating optimal ROI parameters and anatomical location data (distance from myocardium intersections).
• Demonstrated the feasibility of using spectral ultrasound features for CAT identification.
• Identified specific ROI properties and anatomical considerations that significantly impact classification performance.

Conclusions:

• This machine learning approach using spectral ultrasound data shows promise for non-invasive CAT detection, offering a more accessible alternative to MRI.
• Optimized ROI selection and inclusion of anatomical information are crucial for improving classification accuracy.
• The findings pave the way for future research into quantifying CAT thickness using ultrasound.