Deep learning automates detection of wall motion abnormalities via measurement of longitudinal strain from ECG-gated CT images

Figure 1. Flowchart of patient inclusion/exclusions.

Figure 2. Processing of ECG-gated CT for evaluation of LS. (A) ECG-gated volumes are analyzed using a deep-learning (DL) framework that provides the location of the 2-chamber, 3-chamber, and 4-chamber long axis planes and delineates the LV and LA blood pools. From this information, long-axis slices of the segmentations were created throughout the cardiac cycle. (B) The perimeter of the left ventricle and the LV/LA boundary pixels were identified and used to extract the LV perimeter. Method A did not perform any additional processing of the perimeter. However, a convex hull was applied to correct for papillary muscle artifacts (leading to Method B). Further, a cubic splint was fit to the result of the convex hull to correct for variations in texture (Method C). (C) For each long-axis view and each analysis method, the length of the perimeter was measured at end-diastole (the timeframe with largest LV volume) and end-systole (the timeframe with smallest LV volume) and used to calculate LS.

Figure 3. Measurement of endocardial perimeter based on the blood pool segmentation is susceptible to artifacts created by the papillary muscles. (A) The papillary muscles create indentations which impact the measurement of the perimeter. The end-diastolic (left) and end-systolic (right) perimeters for each of the views are shown. They are all overestimated. (B) By modeling the blood pool as a convex hull, we can correct for the errors from indentations created by papillary muscles. However, the perimeter measurement remains affected by the perimeter's texture. (C) Fitting of a curve to the perimeter avoids issues related to the surface texture.

Figure 4. Agreement between LS and FAC increases with use of the convex hull and perimeter curve fitting. The perimeter measured using our deep learning method is susceptible to artifacts due to the insertion points of the papillary muscles and by the texture of the endocardial surface. Use of a convex hull to “fill” in the papillary insertions and curve fitting of the surface improves agreement (R2) with fractional area change of the corresponding long-axis view. Dotted lines represent the 95% confidence interval of the linear fit.

Figure 5. WMA classification accuracy using LS in the training cohort. Receiver operating characteristic curves for the three long-axis views are shown for the three LS methods (blue: naive, red: convex hull, orange: convex hull + curve fitting). The optimal operating point for the convex hull with curve fitting is depicted by a black dot. The operating point of the convex hull with curve fitting in the testing cohort is shown by the black diamond.