P3A-5 3D Segmentation of the Heart Muscle in Real-Time 3D Echocardiographic Sequences Using Image Statistics

A fully automated segmentation of the endocardial surface was developed by integrating spatio-temporal information of 3D ultrasound image sequences. 3D echocardiographic image sequences of the left ventricle of five healthy children were obtained in transthoracic short/long axis view. 2D and 3D (adaptive) filtering was used to reduce speckle noise and optimize the distinction between blood and myocardium, while preserving sharpness of edges between various structures. Four different filters (2D adaptive mean, 2D and 3D adaptive mean squares filter and 2D local entropy) were tested. The filter kernel was related to speckle size to yield statistically robust data. Filter quality was measured by comparing overlap percentages of histograms of manually segmented blood and myocardial regions. ROC curves of manually segmented blood regions were determined to compare effects of the different filters. A deformable contour algorithm was used, after automatic thresholding, to yield a closed contour of the endocardial border in each elevational plane. Each contour was optimized using contours of surrounding spatio-temporal planes as limiting condition to ensure spatio-temporal. The combination of adaptive filtering using image statistics and deformable contours has potential to segment the endocardial surface in 3D