A fully automatic method combining deep learning and a deformable model achieved high accuracy for left ventricular endocardium segmentation on 3D echocardiography, with a mean surface distance of 2.20 mm at end-diastole.
Segmentation of left ventricular (LV) endocardium from 3D echocardiography is important for clinical diagnosis because it not only can provide some clinical indices (e.g. ventricular volume and ejection fraction) but also can be used for the analysis of anatomic structure of ventricle. In this work, we proposed a new full-automatic method, combining the deep learning and deformable model, for the segmentation of LV endocardium. We trained convolutional neural networks to generate a binary cuboid to locate the region of interest (ROI). And then, using ROI as the input, we trained stacked autoencoder to infer the LV initial shape. At last, we adopted snake model initiated by inferred shape to segment the LV endocardium. In the experiments, we used 3DE data, from CETUS challenge 2014 for training and testing by segmentation accuracy and clinical indices. The results demonstrated the proposed method is accuracy and efficiency respect to expert's measurements.
Dong et al. (Wed,) conducted a other in Left ventricular segmentation (n=45). Deep learning and GVF-Snake model vs. Expert's measurements (ground truth) was evaluated on Mean surface distance (dM) for end-diastole (ED). A fully automatic method combining deep learning and a deformable model achieved high accuracy for left ventricular endocardium segmentation on 3D echocardiography, with a mean surface distance of 2.20 mm at end-diastole.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: