Deep learning for fully automated echocardiographic measurements of left ventricular wall thickness and chamber dimensions in the parasternal long-axis view

Abstract

Abstract Background Accurate quantification of left ventricular (LV) wall thickness and chamber dimensions in the echocardiographic parasternal long-axis view (PLAX) is crucial for clinical decisions in patients with heart disease. However, there is a need for more reproducible and time-efficient methods. Fully automated deep learning (DL)-based methods could address this need. Purpose To develop a DL-based method for fully automated measurements of LV wall thickness and chamber dimensions in PLAX, validate its performance in a large population cohort, and evaluate its capacity to differentiate between hypertensive and normotensive groups. Methods DL networks were trained on PLAX recordings from 207 subjects. The DL method performed cardiac segmentation, timing of end-systole and end-diastole, and measured LV wall thickness and chamber dimensions in PLAX. The model was validated in a large populational echocardiography dataset including a subset with test-retest evaluation. Agreement with expert reference measurements, test-retest reproducibility, and comparison of normotensive and hypertensive (defined as those on hypertensive medication or with systolic blood pressure >160 mmHg) groups were evaluated. Results In the populational study of 2047 subjects, the novel DL method demonstrated a feasibility of 97.6%. Strong agreement was observed between the automated DL method and expert readers in measuring end-diastolic septal thickness, posterior wall thickness, and LV diameter (bias 0.1 mm, 0.2 mm, and 2.2 mm, respectively, Figure 1). In a test-retest dataset (n=40, with 2 recordings per patient and 4 readers), the novel DL method displayed improved reproducibility compared to inter-reader measurements, with minimal detectable change for interventricular septal diameter 2.2 mm versus 2.6 mm, posterior wall diameter 1.2 mm versus 2.5 mm, and LV diameter 5.0 mm versus 6.3 mm (Figure 2). The DL method measured significantly wider septal diameter in the hypertensive group compared to normotensive group (p<0.001). Mean analysis time for wall thickness and chamber dimensions using the DL method was 1.2 seconds. Conclusion The novel DL method accurately measured LV wall thickness and chamber dimensions in the PLAX view, and demonstrated improved test-retest reproducibility compared to experienced echocardiographers. Additionally, it was able to detect significantly wider septal diameters in the hypertensive population. By providing fast and reliable measurements in retrospective data, and with potential for real-time measurements during image acquisition, this DL method has the potential to improve the yield, efficiency, and workflow in echocardiography.Bland-Altman posterior wall diameterTest-retest posterior wall diameter