The AutoRV deep learning pipeline automatically detected right ventricular landmarks with a distance of 4.2 ± 3.4 mm from ground truth, yielding errors comparable to inter-observer variability.
Does AutoRV accurately assess right ventricular structure and function compared to manual measurements in mechanically ventilated patients?
An automated deep learning pipeline (AutoRV) can accurately and rapidly quantify right ventricular function from transesophageal echocardiograms in mechanically ventilated patients, with agreement paralleling inter-observer variability.
Objective Right ventricle (RV) dysfunction has therapeutic implications for the management of mechanically ventilated patients in intensive care units. However, RV function is usually qualitatively evaluated by clinicians, as quantitative manual assessment is time-consuming and imprecise. Therefore, we developed AutoRV, an automated monitoring pipeline to assess RV function from 2-D transesophageal echocardiography using deep learning-based landmark detection. Methods Multiple deep convolutional neural networks (CNNs) were implemented to automatically detect the apex of the RV and the tricuspid annulus points in 2-D transesophageal echocardiography images. All networks were trained and tested on images acquired from 51 patients. Multiple parameters were computed based on the apex and tricuspid annulus landmarks, including tricuspid annular plane systolic excursion, RV fractional area change and linear RV strain ( RVS global ). Automated quantifications were compared with manual measurements performed by an experienced clinician. Results The U-Net architecture yielded the most promising results—processing images at 228 frames/s and detecting landmarks at a distance of 4.2 ± 3.4 mm from the corresponding ground truth. The resulting errors in tricuspid annular plane systolic excursion, RV fractional area change and R V S global were equal to 1.1 ± 2.3 mm , 0.2 ± 6.6 % and 3.9 ± 3.6 % , respectively. The agreement between our automatic solution and manual measurements paralleled the inter-observer variability of manual measurements between two clinicians. Conclusion Our fully automatic end-to-end solution, AutoRV, proved to be acceptably accurate and extremely time-efficient, suggesting its potential for automating RV monitoring in intensive care unit patients.
Missana et al. (Mon,) conducted a other in Mechanically ventilated patients in intensive care units (n=51). AutoRV (automated deep learning-based landmark detection) vs. Manual measurements by an experienced clinician was evaluated on Distance of detected landmarks from ground truth and errors in RV functional parameters. The AutoRV deep learning pipeline automatically detected right ventricular landmarks with a distance of 4.2 ± 3.4 mm from ground truth, yielding errors comparable to inter-observer variability.