Abstract
Tissue Doppler Imaging is an essential echocardiographic technique for the non-invasive assessment of myocardial blood velocity. Interpretation by trained experts is time-consuming and disruptive to workflow. This study presents an automated deep learning model, trained and tested on Doppler strips of arbitrary length, capable of rapid beat detection and Cartesian coordinate localisation of peak velocities with accuracy indistinguishable from human experts, but with greater speed.
Introduction
Tissue Doppler imaging (TDI) is a relatively new echocardiographic technique that uses Doppler principles to measure the velocity of myocardial motion. Clinical guidelines recommend averaging peak velocity measurements over a minimum of three consecutive beats. However, echocardiographers often select beats they consider an average representative sample which may contribute to test-retest variability, leading to diagnostic errors. A reliable and objective automated system would save valuable resources for health services and has potential to improve patient outcomes by averaging measurements over more beats. By removing manual detection, specialists’ time can be better spent acquiring more high-quality beats, reducing subjectivity and cost.
In the example image, each cardiac cycle is indicated by a red vertical line with annotations from two trained experts for each keypoint. Expert 1 is a cross, with expert 2 a square.
Method
TDI traces were acquired from 48 patients with a mean age of 64±11 years, from both the
septal and lateral annuli. Information about the dataset and patient characteristics can be
found in. Six recordings were acquired for each patient and reconstructed into a continuous Doppler strip with a resolution of 900 x 1300 pixels. Information
about the reconstruction methods can be found in. The dataset
comprises 280 Doppler strips (5,327 beats). Annotations are from three expert clinicians;
ground-truth labels for Model training and evaluation were calculated as the expert consensus. Additionally, for the purpose of investigating inter-observer variability, three additional
networks were trained on individual expert labels, named Model-1, Model-2 and Model-3, respectively.
The network architecture is two-fold:
A. Heartbeats are detected/isolated
(without the need for ECG signal)
as a ROI by the Mask R-CNN architecture with a ResNet101 backbone. Images are resized and
zero padded to 1024x1024 pixels
B1-B3. ROI is cropped and resized to 192 x 192 pixels and input
to a convolutional heatmap regression model to predict Cartesian coordinates for systolic and diastolic
peak velocities (S’, E’ and A’), as
shown in the figure to the right.
Results and Discussion
Computation time for manual peak velocity annotations by human experts, compared to the automated model, was calculated over an average sample of 25 heartbeats; 4.76 seconds and 0.18 milliseconds, respectively. Cartesian coordinates in pixels were converted into Velocity measurements in cm/s.
The figure above shows mean septal S’, E’ and A’ velocity estimates and standard deviations using the experts consensus (red) and the Model (black) for each patient. Circular markers represent the mean and vertical bars represent the standard deviation. Patients have been placed in ascending order of the average velocity. Table 1 details Bland–Altman bias and 95% limits of agreement when comparing expert annotations and Model predictions for peak tissue Doppler velocity measurements at the septal and lateral annulus. We demonstrate the performance of the proposed Model is akin to human experts; detection error is within the range of calculated inter-observer variability, however processing time is greatly reduced.
| Model/Expert | Septal annulus | Lateral annulus | ||||
|---|---|---|---|---|---|---|
| S' | E' | A' | S' | E' | A' | |
| Human performance | ||||||
| Exp 1,2 vs. Expert-3 | 0.13±0.59 | -0.18±0.59 | -0.06±0.84 | 0.33±0.91 | 0.15±0.92 | 0.11±0.84 |
| Exp 1,3 vs. Expert-2 | 0.06±0.50 | -0.22±0.60 | 0.29±0.70 | 0.33±0.87 | 0.12±0.80 | 0.17±0.75 |
| Exp 2,3 vs. Expert-1 | -0.19±0.63 | -0.04±0.56 | -0.24±0.69 | -0.66±0.94 | -0.27±0.77 | -0.27±0.78 |
| Expert consensus | -0.14±0.67 | 0.06±0.70 | -0.08±0.90 | -0.44±1.10 | -0.19±0.97 | -0.17±0.92 |
| Machine performance | ||||||
| Exp 1, 2 vs. Model-3 | -0.01±0.82 | -0.42±1.00 | -0.11±0.85 | 0.59±0.93 | 0.08±1.32 | 0.38±1.87 |
| Exp 1, 3 vs. Model-2 | 0.04±0.93 | 0.15±0.94 | 0.44±0.98 | 0.50±0.93 | 0.19±1.10 | 0.21±1.66 |
| Exp 2, 3 vs. Model-1 | -0.12±0.97 | -0.17±0.94 | -0.15±0.99 | -0.11±1.02 | -0.04±0.93 | -0.04±1.35 |
| Expert consensus vs. Model | -0.07±0.78 | -0.22±0.92 | -0.02±0.88 | -0.38±0.81 | -0.06±0.84 | 0.19±1.38 |