Segmentation of 3D RF echocardiography using a joint spatio-temporal predictor and signal intensity model

We present an approach for left ventricular segmentation of radio-frequency (RF) ultrasound sequences. Our method employs an independent identically distributed (iid) spatial model for RF voxel intensity and a conditional model relating a subsequent frame in the image sequence to the frame being segmented by means of a linear predictor that exploits spatio-temporal coherence in the data. The conditional model overcomes segmentation problems due to image inhomogeneity issues, while the spatial model overcomes a tendency of the conditional model to underestimate the blood pool. The method is validated by comparison with manual tracings, segmentations performed using Chan-Vese level sets, and by segmentations leveraging only the linear predictor.