Real-time deep learning-based tracking assistance during acquisition improves image quality for right ventricular strain assessment
V Holmstrom, A Azad, E O Jakobsen, J Nyberg, A Chernyshov, S Stolen, H Dalen, E Holte, L Lovstakken, A Ostvik, B GrenneAbstract
Background
Right ventricular (RV) free-wall longitudinal strain (FWLS) is a sensitive marker of RV systolic dysfunction. However, imaging of the RV is challenged by its complex geometry, anterior position, and thin, trabeculated wall, which impair precise endocardial delineation and myocardial motion tracking, resulting in suboptimal RV-FWLS feasibility and operator-related variability.
Purpose
To develop a deep learning (DL) application that provides real-time RV free wall tracking assistance during image acquisition, and to prospectively assess its effect on RV-FWLS, interreader variability, feasibility and baseline myocardial drift.
Methods
A neural network (U-Net KP1) detected and tracked RV free-wall landmarks through the cardiac cycle. A dedicated workstation connected to the echocardiography scanner provided real-time visualization of tracking performance during acquisition without prior image storage (Image), allowing the operator to detect and correct suboptimal images. Real-time DL tracking acquisitions were compared with conventional acquisitions in a prospective test–retest study of 29 patients recruited from a clinical cardiology population. Seven experienced readers analysed images (EchoPAC version 206, GE HealthCare) blinded to acquisition method. Interreader variability was quantified as the per-patient coefficient of variation (CoV) across seven readers. Mixed-effects logistic regression with patient and reader clustering assessed feasibility, and odds ratios quantified differences between methods. Free-wall drift was calculated from the displacement of tracked points between end-diastolic frames. Student’s t-test or paired Wilcoxon rank-sum was used as appropriate.
Results
Mean age was 57 ± 16 years, and 69% were male. Comorbidities included hypertension (45%), heart failure (34%) and coronary artery disease (24%). Mean absolute RV-FWLS across the seven readers was higher using the real-time DL tracking assistance compared with conventional acquisition (22.6% vs. 20.4%, p < 0.01). Tracking assistance reduced interreader variability (CoV 6.0% vs. 8.3%, p = 0.03, Figure). Moreover, real-time DL tracking improved both segmental feasibility (89% vs. 84% free-wall segments analysed) and RV-FWLS feasibility (exams with ≥2 of 3 segments analysed: 93% vs. 88%), translating to 1.8-fold higher odds for obtaining free-wall segmental analysis (p < 0.01) and 2.5-fold higher odds for obtaining RV-FWLS (p = 0.03). Finally, images acquired with real-time tracking assistance showed lower median free-wall drift (1.56 mm [IQR 1.73] vs. 1.81 mm [IQR 1.95], p < 0.01).
Conclusion
DL-based real-time feedback on RV tracking performance during acquisition improved RV-FWLS assessment, as reflected by higher absolute RV-FWLS, improved interreader agreement, better feasibility, and reduced baseline drift. This DL-based workflow may facilitate more reliable RV-FWLS assessment and support its wider clinical implementation.Image: Deep learning workflowFigure: Interreader CoV