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

The number one cause of death in the United States is consistently cardiovascular disease (CVD). Studies have proven that the buildup of cardiac adipose tissue (CAT) around the heart is a biomarker of CVD. MRI is the gold standard for imaging CAT but is expensive and not widely available. Ultrasound is less expensive and portable, but the images are noisy, and it is difficult to identify or quantify CAT. The aim of this project is to use spectral analysis of raw radiofrequency (RF) ultrasound data as input for a machine learning classifier to automatically classify regions-of-interest (ROIs) around the heart as containing CAT or not. ROIs are labeled using corresponding MRI images of the same patients. A previous study used a random forest classifier with 9 spectral parameters as input to classify tissue types in echocardiograms. This project focuses on improving this classifier by experimenting with properties of the chosen ROIs. Experiments were performed independently varying the ROI circumference (length), width, the threshold CAT thickness used for labeling an ROI as CAT, and the signal level required for valid processing. Additional experiments explored the impact of the anatomical location of each ROI as an input. The addition of two parameters indicating the distance of each ROI from the two left and right myocardium intersections as well as the use of the optimal ROI parameters as determined from experimentation resulted in an accuracy of 75.5%. This demonstrates feasibility of this approach for identifying CAT around the heart and will lead to future work in estimating the thickness of fat in each ROI.