Automated Quality-Controlled Left Heart Segmentation from 2D Echocardiography

Segmentation of 2D echocardiography (2DE) images is an important prerequisite for quantifying cardiac function. Although deep learning can automate analysis, variability in image quality and limitations in model generalisability can result in inaccurate segmentations. We present an automated quality control (QC) methodology to identify invalid segmentations, and propose post-processing techniques to automatically correct erroneous segmentations. A workflow was developed to utilise a deep learning model, trained using the CAMUS dataset, for segmenting all frames within apical two-chamber and four-chamber 2DE images from an independent dataset containing 91 participants (28 females; 51 healthy controls and 40 patients with mixed cardiac pathologies). Single- and multi-frame QC and post-processing techniques were applied, and subsequently validated against manual QC in a sample of 50 randomly selected participants. Cardiac indices derived from the automated segmentations using 2DE were compared to reference values obtained through expert manual analysis on the same subjects. Single-frame QC improved the proportion of usable frames from 76% to 96%. Multi-frame QC indicated failures in 53% of the images, and while the resulting specificity was 96%, correction only achieved a sensitivity of 42% with respect to manual assessment. The exclusion of the rejected images resulted in improvements in the reliability between predicted and manual measurements. These results demonstrated that applying automated QC to deep learning segmentation methods can enhance the reliability of 2DE segmentations.