Explicit and automatic ejection fraction assessment on 2D cardiac ultrasound

Funding Acknowledgements Type of funding sources: Private company. Main funding source(s): DESKi Background Left ventricular ejection fraction (EF) is a key parameter to assess cardiovascular functions in 2D cardiac ultrasound, but its manual assessment is time-consuming and subject to high inter and intra-observer variability. Deep learning-based approaches have been proposed to automatically predict EF but can suffer from a lack of explainability and interpretability, which limit the trust of the clinician and prevent errors from being detected. Purpose In this context, we propose a fully automatic method to reliably evaluate the biplane left ventricular EF on 2D echocardiography while following the modified Simpson’s rule calculation steps to provide explicit details to clinicians. Methods A deep learning model based on U-Net architecture was trained on internal apical 4 and 2-chamber echocardiographic images to segment the left ventricle and locate the mitral valve. Predicted segmentations are then validated with a statistical shape model which detects potential failures that could impact the final EF evaluation. Finally, end-diastolic and end-systolic frames are identified thanks to a spline interpolated from the remaining LV segmentations’ areas, and the EF is estimated on all available cardiac cycles. This approach was trained on a dataset of 783 patients. Its performances were evaluated on an internal dataset of 200 patients with a large diversity of EF and on an external openly available dataset containing 450 patients. Results On the internal dataset, EF assessment achieved a mean absolute error of 6.10% and a bias of 1.56 ± 7.58% using multiple cardiac cycles and removing failed segmentations. Regarding end-diastolic and end-systolic volumes, the mean absolute error was evaluated at 13.75mL and 10.70mL respectively. Segmentation performances were evaluated with the Dice and the Hausdorff distance, respectively at 0.92 ± 0.41 and 6.13 ± 2.9mm. On the external dataset, the approach predicted EF with a mean absolute error of 5.39% and a bias of -0.74 ± 7.12%. The mean absolute error for end-diastolic volume was 15.40mL and for end-systolic volume 8.18mL. Conclusions Following the recommended guidelines, we proposed an end-to-end fully automatic approach achieving state-of-the-art performances in EF evaluation in 2D cardiac ultrasound while giving explicit details to the clinicians at each step of the assessment. Abstract Figure. End-to-end fully automatic approach