A view-flexible deep learning framework for automated analysis of 2D echocardiography

Figures

• 

  Created in BioRender. Malins, J. (2026) https://BioRender.com/lhw4pb1.

• 

  Fig. 1. LVEF estimation and classification across datasets. LVEF estimation and classification for the A Minnesota-Wisconsin, B Arizona, C Florida, and D EchoNet datasets. In each subplot, the left plot shows the correlation for LVEF estimation (i.e., comparison of model-estimated LVEF to clinically-calculated LVEF), whereas the right plot shows the ROC curve and confusion matrix for reduced versus normal LVEF classification.

• 

  Fig. 2. Bland-Altman analysis of LVEF estimation error. Bland-Altman plots showing the discrepancy between clinically-calculated and model-estimated LVEF for the A Minnesota-Wisconsin, B Arizona, C Florida, and D EchoNet datasets. The text in blue shows the number of cases with errors within certain bands (e.g., less than 5%, between 5 and 10%, and so on).

• 

  Fig. 3. The effect of the number of clips on LVEF estimation performance. The effect of the number of clips on LVEF estimation performance for the Minnesota-Wisconsin testing dataset. The leftmost five boxplots within each panel are for one through five random clips selected from the full set of valid-class clips for a patient’s exam, whereas “All valid clips” represents all valid-class clips from a patient’s exam. In each boxplot, the center line corresponds to the median, the box limits correspond to the upper and lower quartiles, the whiskers correspond to 1.5 times the interquartile range, and the points correspond to outliers.

• 

  Fig. 4. LVEF model performance for the four ‘valid’ views and combinations of these views. For combinations of views, model performance was evaluated using the average model estimate per patient across views, with only one clip taken per view (i.e., the clip with the highest score from the view classifier for that view). The legend on the left was plotted using version 0.9.0 of the upsetplot package51 in Python version 3.9.20.

• 

  Fig. 5. Visualization of regional model attention in echocardiographic videos using Grad-CAM. The Grad-CAM color overlays highlight the regions most influential in the model’s prediction, with frame numbers shown at the top of each tile. Each panel displays 12 frames taken at regular intervals from a randomly selected valid echocardiographic view with a random starting frame within the available temporal window. A An example patient with a clinically-calculated LVEF of 30% and a model-estimated LVEF of 41.31%. B An example patient with a clinically-calculated LVEF of 63% and a model-estimated LVEF of 61.75%. The colorbar indicates the relative intensity of model attention, with cooler colors (blue–green) denoting areas of low influence and warmer colors (yellow–red) denoting regions contributing most strongly to the prediction.

• 

  Fig. 6. LVEF estimation and classification in the simultaneous TTE-HCU dataset and the HCU dataset collected by experts and novice users. LVEF estimation and classification performance for (A) the simultaneously collected TTE-HCU dataset, and (B) the HCU dataset collected by both experts and novice users using UltraSight real-time AI guidance technology. The green lines between each set of points in the leftmost plots (i) correspond to the difference between estimates. The red points in the middle plots (ii) show the outliers for which the difference between estimates is greater than or equal to 10. For the outliers, in the legend, the true LVEF value is shown with “T” and the difference is shown with “D”. The rightmost plots (iii) show the ROC curves and metrics for reduced versus normal LVEF classification.

• 

  Fig. 7. Model performance for the view-flexible age estimation and sex classification models. The left three plots show age estimation (A) and (B) sex classification results for the TTE cohorts for both the internal (i) and external ((ii) and (iii)) testing datasets. The rightmost plot (iv) shows the overlapping correlation for HCU and TTE data collected simultaneously from the same set of patients at Mayo-Rochester, with green line between each corresponding set of estimates showing the difference between HCU versus TTE model-estimated age. The 95% confidence interval for each corresponding metric is shown in square brackets.

• 

  Fig. 8. Patient cohorts used for model development and evaluation. TTE transthoracic echocardiography, HCU handheld cardiac ultrasound, RH right heart, LVEF left ventricular ejection fraction.

• 

  Fig. 9. An overview of the view-flexible deep learning framework employed in the current study. Note that the LVEF model estimates are hypothetical and do not constitute actual data points from the study. The icon used to denote the deep learning model was sourced from BioRender.com.