Real-time guiding by deep learning during echocardiography to reduce left ventricular foreshortening and measurement variability

Figures

• 

  Figure 1. Flow chart of the study population. Participant inclusion process is illustrated in the left part of the figure. The two panels indicate the two separate study periods (Period 1 and Period 2). The second examination (thickened frame) in Period 2 illustrates the sonographers using the real-time deep-learning algorithm. DL, deep learning; 2D, two-dimensional.

• 

  Figure 2. The real-time deep-learning feedback of view-specific left ventricular length and apical foreshortening. (A) The real-time DL-guiding tool (right screen) adjacent to the ultrasound machine (left screen). The tool is also visualized in Supplementary data online, Video S1. (B and C) Screenshots of the DL real-time foreshortening detection from four-chamber (A4C) recordings of the same patient. (B) Foreshortened A4C. (C) A4C recording using the DL-guiding tool without foreshortening (no significant difference between the A4C LV length and corresponding A2C LV length in the LV length indicator to the right in the image).

• 

  Figure 3. LV length measurement method. Screenshots of calipers used for measuring LV length at the end-diastole. First, a caliper is placed between the two mitral annular points. Secondly, the LV length is estimated by placing another caliper from the subendocardial apical point to the centrum of the first mitral annular caliper. (A) Apical two-chamber recording. (B) Apical four-chamber recording. (C) Cardiologist’s tri-plane recording used for reference LV length. The mean LV length of the three planes was used as reference. The minor differences between views are partly explained by mispositioning of the apical point in four-chamber view, while the mitral annular curvature may also explain some variability between views.

• 

  Figure 4. Box plots of apical foreshortening for each view and operator. Box plots for LV foreshortening grouped by study period, operator group, and use of DL guiding. Box plot centre line is the median with whisker lengths extending 1.5 times the interquartile range above Q3 and below Q1. a = P < 0.05 between study periods within the same operator group; b = P < 0.05 vs. cardiologist within same study period (except b*; P = 0.075), c = P < 0.05 vs. other sonographer group within same study period (cardiologists excluded). LV, left ventricular; P1, Period 1; P2. Period 2; DL, deep learning.

• 

  Figure 5. Scatter plot with a linear mixed model regression line. LV length difference in relation to LV foreshortening in sonographer examinations. LV foreshortening (x-axis) was calculated as the reference LV length obtained from the same 2D recording minus the reference LV length from the cardiologist’s tri-plane recording. LV length difference (y-axis) was calculated as LV length extracted from the view-specific end-diastolic volume measurement minus the reference length obtained from the same 2D recording. LV, left ventricular.