Abstract Introduction Assessing left ventricular (LV) diastolic dysfunction is essential, because it can lead to heart failure with preserved ejection fraction (HFpEF), significantly impacting morbidity and mortality rates. Echocardiography is the primary imaging tool for detecting LV diastolic dysfunction, however its assessment is time consuming and requires experienced clinicians to interpret multiple echocardiographic parameters. Recent advancements in deep learning might offer the potential to reduce analysis time and standardize the evaluation of LV diastolic function. Purpose The aim of this study was to assess the potential of deep convolutional neural networks (CNNs) for the evaluation of LV diastolic dysfunction. Methods Prospective observational study enrolled 233 patients, a total of 3238 2D echocardiographic images were obtained. The study included patients with normal LV systolic function, who were suspected of HFpEF. The ensemble of four CNN’s was used for the evaluation of transmitral inflow waves (E and A peak velocities), mitral (septal and lateral) velocities (e’) and peak tricuspid regurgitation velocity (TR Vmax). Additional diastolic function parameters, such as the E/A and E/e’ ratios, were calculated to assess diastolic function classes based on the 2016 ASE/EACVI recommendations. Manual measurements were performed by an expert cardiologist. CNN’s results detecting diastolic dysfunction were compared to the evaluation of expert cardiologists. Results The analysis revealed high yields of CNN’s measurements across various parameters, ranging from 81.12% for the E/e' ratio to 89.27% for the A wave. CNNs perform well in measuring maximum velocities using different Doppler imaging modes. The root mean square error (RMSE) values indicated the degree of deviation between the automated system and the cardiologist measurements, with the lowest RMSE observed for the E/A ratio (0.4). Bias showed minor discrepancies for the E/A ratio (-0.05), with the largest bias observed in the A wave (5.95) (Table 1). Conclusion This study demonstrates that automated echocardiographic analysis shows high yield and good agreement with cardiologist measurements, though some parameters require further enhancement to ensure greater accuracy and consistency. The application of CNNs can potentially simplify the assessment process, reduce analysis time, and provide standardized evaluations. Table 1. Results of statistical analysis
Read full abstract