Automatic detection of mitral annulus in echocardiography based on prior knowledge and local context

Due to the inherent noisy, low resolution and limited imaging range of echocardiography, it is difficult to identify mitral annulus (MA) where valves end that is crucial for further segmentation, modeling and multi-modalities registration of mitral valves. This work aims to automatically detect MA hinge points combining information of intra-cardiac local context and location relationships. The method includes the following steps: (1) segment left ventricle (LV) by prior shape and local histogram fitting based Active Contour Model (ACM); (2) design the local context features for training and classification of MA hinge points; (3) utilize additive min kernel based Support Vector Machines (SVM) classifier for fast computation to obtain MA candidates; (4) estimate MA hinge points by K-means algorithm under the location constraint of LV and MA. Our method was tested on echocardiographic four chamber image sequence of 10 pediatric patients (6 boys, 4 girls, 7.6±3.4 years). Compared with the manual annotations, the automatically detected MA results are reliable with reasonable accuracy, for lateral point (2.0±1.9, 1.8±1.2) pixels and for septal point (2.9±2.6, 1.2±1.0) pixels.