Full-cycle left ventricular segmentation and tracking in 3D echocardiography using active appearance models

Fully automatic segmentation of the left ventricle in 3D echocardiography is a challenging task due to its low contrast-to-noise compared to other cardiac imaging modalities, its typical artifacts and the grainy appearance due to speckle. We previously demonstrated the use of active appearance models (AAM) for segmentation of the end-diastolic (ED) and end-systolic phases of the cardiac cycle, which results in a competitive performance with respect to state-of-the-art methods. In this work we propose an extension of the 3D AAM segmentation with tracking of the left ventricle (LV) over the entire cardiac cycle. We test and evaluate the methodology in the framework of the CETUS challenge on data of 30 patients with varying pathology, imaged with varying ultrasound equipment at three centers. The results confirm that tracking the endocardium from the initial detection in ED improves the detection of ES, as surface distances for ES drop from 2.8 ± 1.5 mm to 2.5 ± 0.8 mm. Furthermore, tracking the full cycle twice renders stable detections for ED and ES, resulting in further improvement in ED surface distances (from 2.0 ± 0.6 mm to 1.9 ± 0.8 mm) and ejection fraction. Our results show that tracking improves overall segmentation results and prediction of clinical parameters. Segmentation and tracking based on AAMs might be a solid base for future hybrid solutions where locally refined detection is added by different methodology.