Retraining Convolutional Neural Networks for Specialized Cardiovascular Imaging Tasks: Lessons from Tetralogy of Fallot

Fig. 1. Cardiac magnetic resonance contours for a patient with repaired tetralogy of Fallot and right ventricular dilation, who showed significant improvement in RV contours after training. End diastole is shown, with manual contours on the left, contours derived from the initial MSN algorithm in the middle, and retrained MSN + rTOF algorithm on the right. Note slice selection errors with both MSN and MSN + rTOF methods. 3D representations of the ventricular contours are shown below. Left ventricular endocardial contours are in red, left ventricular epicardial in green, and right ventricular endocardial in yellow

Fig. 2. Spatial MSN performance on the training dataset. As hypothesized, the RV endocardial contours generated by the MSN algorithm were consistently worse than the LV contours. Representative ED AVD data are shown. Violin plots with LV endo contours in red; LV epi contours in green; and RV endo in yellow. Bars represent median and IQR. * represents p < 0.05; **** represents p ≤ 0.0001

Fig. 3. Volumetric MSN performance on the training dataset is worse for the RV. EDV is shown on the left, ESV in the middle, and EF on the right. LV volumes with red and RV volumes in yellow. Solid line represents median and dotted lines IQR. * represents p < 0.05

Fig. 4. MSN vs. MSN + rTOF algorithm spatial performance on the testing dataset. Example data are shown for RV endo contours, with DSC (top), HD (center), and AVD (bottom). Violin plots are shown with MSN on the left (orange) and MSN + rTOF on the right (purple), and changes for individual cases are shown in the middle. Solid line represents median and dotted lines IQR. * represents p < 0.05; ** represents p < 0.01; *** represents p < 0.001; **** represents p ≤ 0.0001

Fig. 5. Volumetric comparisons of MSN and MSN + rTOF algorithms. The top panel shows correlation of algorithmic RV EDV and manual RV EDV. The line of identity, best-fit lines, and best-fit line errors are shown. MSN + rTOF showed significantly improved correlation with manual volumes compared to MSN (p = 0.0459). The middle panel shows the Bland–Altman analysis of MSN and MSN + rTOF RV EDV compared to the manual contours. The bottom panel shows a violin plot with MSN on the left and MSN + rTOF on the right, and changes for individual cases shown in the middle. Solid line represents median and dotted lines represent IQR. ** represents p < 0.01

Fig. 6. Shown here are the MSN + rTOF contours at ED. This patient has a significantly dilated RV, with the “shoulder” extending more basally past the tricuspid valve plane (yellow arrows), necessitating extension of the short-axis slices basally past the mitral and tricuspid valve plane. The MSN + rTOF contours extend into the right atrium and into subcutaneous fat past the apex. The LV contour is drawn in the left atrium on the most basal slice and at least one slice past the apex. The volumetric calculations are done using all contours, interpolating for missing slices