Utilization of Texture Analysis to Identify Echocardiogram Features in Patients with Reduced Left Ventricular Ejection Fraction

Introduction: Heart failure (HF) accounts for 86,177 deaths in the United States every year. Left-sided HF, the most common form of HF, is characterized by the weakening of the left ventricle, which often results in decreased left ventricular ejection fraction (LVEF). LVEF is the volume of blood pumped out of the left ventricle during systole of the heart and is expressed as a percentage. Existing literature has established the correlation between dilation of the ventricular chambers, ventricular wall thinning, and stiffening of the ventricular walls with HF [Konstam, JACC, 2011]. These abnormalities in the heart chambers are routinely diagnosed using point-of-care ultrasound (POCUS) imaging. Furthermore, current point-of-care ultrasound-based blood ejection fraction (EF) calculations rely on the qualitative assessment of a clinician [Quinones, Circulation, 1981]. Quantitative ultrasound image texture analysis utilizes a statistical approach to describe individual grayscale values within a medical image [Paris, Ultrasound in Med & Biol, 2020]. From this, echocardiogram texture can be characterized through a series of calculations that describe the spatial distribution of signal intensities within an image. We propose that the application of image texture analysis to POCUS imaging can provide diagnostic accuracy for the objective assessment of left ventricular heart failure. Therefore, we hypothesize that echocardiogram texture parameters can distinguish between patients with preserved and reduced LVEF.

Materials and

Methods: Cardiac ultrasound images were obtained through the publicly available CAMUS dataset [Leclerc, IEEE Medical Imaging, 2019], consisting of data from 450 patients (263 males and 145 females, average age 65.2 +/- 14.5). Each patient had 2D apical four-chamber view sequences obtained in B-mode for one full cardiac cycle. The US images were manually segmented during systole and diastole by three cardiologists, independent of our research team. Prior to analysis, data from 42 of the 450 patients were excluded due to poor image quality as determined by the cardiologists. The study population was then categorized into five populations based on ejection fraction: hyperdynamic, normal, mild, moderate, and severe dysfunction. Echocardiogram image texture analysis (TA) was then performed using a custom implementation of the Python library pyRadiomics. 90 radiomic features were included in the texture analysis consisting of first, second, and higher-order parameters. The region of interest (ROI) selected for TA was defined as the region between the epicardium and endocardium as segmented by the cardiologists (Figure 1). An ANOVA with sex and cardiac phase as covariates was performed to determine which texture parameters are associated with LVEF. To account for multiple hypotheses, a Bonferroni correction was implemented. A stepwise regression model was then built using the statistically significant texture analysis parameters. All statistical analyses were performed in JMP 16.0.

Results, Conclusions, and Discussions: Following Bonferroni correction, four texture analysis parameters were found to show a statistically significant correlation with left ventricular ejection fraction (p < 0.0005). These included an increase in GLDM Dependence Non-uniformity (p < 0.0001), an increase in GLCM Inverse Difference Normalized (p < 0.0001), a decrease in GLCM Informational Measure of Correlation (p = 0.0003), and an increase in GLRLM Run Entropy (p = 0.0003). Further analysis revealed that one texture parameter (GLCM Dependence Non-uniformity) varied by sex (p = 0.001), whereas the other three did not show a statistically significant variance with sex (all p-values > 0.0623). Additionally, these three parameters varied by cardiac phase (end-systole vs end-diastole p < 0.001). Age did not have a statistically significant effect on any of the parameters (all p-values > 0.05). The stepwise models built using the statistically significant parameters all demonstrated low R-squared values ( < 0.03) and require further investigation.

The current analysis is among the first to apply sonography-based radiomic analysis in the setting of left ventricular heart failure. Our results suggest that increased heterogeneity of the left ventricular wall may be indicative of left ventricular heart failure. Previous studies have characterized physical changes (e.g. interstitial fibrosis, wall thinning and elongation) in the left ventricular wall seen in HF patients using histology and cardiac magnetic resonance imaging. The identification of three texture parameters (that do not vary significantly with sex or age) that can classify left ventricular heart failure demonstrates the encouraging potential of the diagnostic value of radiomic features. However, additional analysis is needed to incorporate these parameters into a statistical model that can account for the variance in the samples. Further studies should seek to implement an automated segmentation algorithm of the ventricular walls for use in the texture analysis. The incorporation of such an algorithm would facilitate the automated extraction of texture analysis features, thus broadening the potential applications of this toolset to those who do not have expertise in interpreting echocardiograms.

Acknowledgements (Optional):