3. Results 3.1. Deep Neural Network (DNN) Trained with a General Dataset
Transthoracic echocardiograms of 262 patients with C/SHD were identified for inclu-
sion. Patient characteristics and individual congenital or structural abnormalities of the
population are depicted in Table
1
. In addition, echocardiography studies of 62 patients
(mean age 45 years, 50.0% male) without a cardiac abnormality were identified and in-
cluded for automated view classification. In total, 9793 TTE loops were included for the
patient group with C/SHD. Of these, 8371 loops were acquired on GE ultrasound systems
(Vivid 7 or Vivid E95) and 1422 loops were acquired on Philips ultrasound systems (Epiq).
For the group with normal cardiac anatomy, 706 loops were included in the analysis. In
total, 284,250 individual frames were assessed for view classification by the DNN trained
with a general dataset in the present study. Overall, the accuracy of the DNN trained with
a general dataset concerning view classification was 48.3% in patients with C/SHD (see
Table
2
) on a frame by frame basis. The highest accuracy was achieved in the identification
of the parasternal long axis (76.5% correct) and the subcostal 4 chamber view (87.7% correct).
In contrast, the DNN had a low accuracy in distinguishing the different parasternal short
axis views and apical views (see Table
2
).
The DNN’s accuracy for view classification was 66.7% overall in patients without a
cardiac abnormality (see Table
3
). In this group of patients, identification of the parasternal
long axis and subcostal 4 chamber view remained very accurate (98.4% and 100%, respec-
tively), but the differentiation between separate parasternal short axis and apical views
was higher compared with C/SHD-frames. For example, a parasternal short axis view at
the level of the papillary muscles was correctly identified by the DNN in 63.0% of frames
depicting C/SHD compared with 79.4% of frames without cardiac abnormality and the
apical 4 chamber view was correctly identified in 52.7% of frames with C/SHD versus
77.5% of frames without cardiac abnormality.
3.2. DNN Trained with a Congenital or Structural Heart Disease (C/SHD)-Specific Dataset
A new convolutional neural network was independently trained on 139,910 frames
depicting C/SHD and subsequently tested on a dataset of 35,614 frames. Table
4
depicts a
cross matrix of this DNN’s accuracy in the identification of the 17 utilized echocardiographic
views. The overall accuracy across all views was 76.1%. Similar to the DNN trained with a
general dataset, the parasternal long axis and subcostal views were distinguished with a
high accuracy by the DNN trained with a C/SHD-specific dataset. However, this DNN
showed a higher accuracy over the DNN trained with a general dataset in the classification
of parasternal short axis and apical views. For example, the DNN trained with a C/SHD-
specific dataset was able to detect a parasternal short axis view of the mitral valve with
an accuracy of 52% compared to an accuracy of 11.3% by the DNN trained with a general
dataset. Additionally, the apical 2-, 3-, 4- and 5-chamber views were able to be distinguished
with a very high accuracy by the DNN trained with a C/SHD-specific dataset (80%, 88%,
78% and 91%, respectively) compared to the DNN by Zhang et al. (31.3%, 28.5%, 52.7%
and 25.5%, respectively). This resulted in a highly statistically significant difference in
the accuracy of the DNN trained with a C/SHD-specific dataset compared with the DNN
trained with a general dataset in the view classification of patients with C/SHD (p < 0.001).