2. Methods
The present study was conducted according to the Declaration of Helsinki and its later
amendments. The project was approved by the local ethics committee (Approval number
2020-751-f-S).
2.1. Imaging Database
Echocardiograms of patients with congenital and structural heart disease were selected
retrospectively from the routine clinical imaging database of the Department of Cardiology
III—Adult Congenital and Valvular Heart Disease at the University Hospital Muenster.
Echocardiograms were chosen for diversity of underlying disease etiology (see Table
1
),
comprehensiveness of echocardiographic views and quality of acquired loops. In addition,
echocardiograms of patients without a cardiac abnormality were prospectively included
according to the aforementioned criteria. The examinations were performed on different
echocardiography machines from different vendors (especially GE Vivid E9, Vivid E95,
Vivid 7; Philips EPIC 7C, EPIC 7G, iE33). Two-dimensional (2D) echocardiographic studies
performed according to current guideline recommendations [
1
] were anonymized, exported
and converted into individual frames in a PNG format for automated analysis. In total,
individual frames of 17 separate TTE views were obtained. Figure
1
details the utilized
echocardiography views.
J. Clin. Med. 2022, 11, 690
3 of 11
Table 1.
Characteristics of the patient population with congenital or structural heart disease.
Diagnosis
Number (%)
Mean Age in Years
Male (%)
Ebstein anomaly
18 (7%)
40
14 (78%)
Hypoplastic left heart
3 (1%)
23
2 (67%)
Tricuspid atresia
1 (<1%)
36
1 (100%)
Non-compaction
cardiomyopathy
9 (3%)
37
7 (78%)
Transposition of the great
arteries (TGA)
48 (18%)
39
25 (52%)
Tetralogy of Fallot
30 (11%)
47
15 (50%)
Incomplete
atrio-ventricular septal
defect
6 (2%)
39
1 (17%)
Complete
atrio-ventricular septal
defect
5 (2%)
55
2 (40%)
Double outlet right
ventricle
1 (<1%)
50
1 (100%)
Atrial septal defect
7 (3%)
37
4 (57%)
Amyloidosis
23 (9%)
68
20 (87%)
Hypertrophic obstructive
cardiomyopathy
7 (3%)
58
3 (43%)
Fabry disease
11 (4%)
60
7 (64%)
Dilatative
cardiomyopathy
54 (21%)
58
35 (65%)
Congenitally corrected
TGA
34 (13%)
48
18 (53%)
Muscular ventricular
septal defect
5 (2%)
42
1 (20%)
All patients
262
49
±
17
156 (60%)
2.2. Convolutional Neural Networks
The source code and model weights of the DNN trained with a general dataset were
obtained from
https://bitbucket.org/rahuldeo/echocv
(accessed on 20 March 2021) and
the training and validation methodology was previously published by Zhang et al. [
12
]. To
summarize, a 13-layer convolutional neural network (VGG 13) was trained with images
assigned an individual view label and five-fold cross-validation was used to assess accuracy.
Because of a lack of images for some uncommon views in the C/SHD-cohort, we decided
to use 17 instead of the original 23 different views. Additionally, for ease of comparison,
we assessed single echocardiographic frames individually instead of averaging accuracy
across frames of the same image loop.
For the DNN trained with a C/SHD-specific dataset, our echocardiographic dataset
was split into a training/validation group (80%) and a test group (20%). Frames from
patients of the test group were not used for model training to ensure the external validity
of the new model. Image resolution was reduced to 150
×
150 pixels and a greyscale of
256 shades.
J. Clin. Med. 2022, 11, 690
4 of 11
J. Clin. Med. 2022, 11, x FOR PEER REVIEW
4 of 12
J. Clin. Med. 2022, 11, x. https://doi.org/10.3390/xxxxx www.mdpi.com/journal/jcm
Dostları ilə paylaş: |