Auxiliary Diagnosis Technology Study On Congenital Heart Disease In Echocardiography

The mitral valve(MV) is an important part of the heart. Its physiological function is controlling the one-way flow of blood from the left atrium(LA) to the left ventricle(LV). According to the statistics, about half of the congenital heart disease cases in China are resulting from the structural abnormalities and dysfunction of MV. Mitral regurgitation and mitral stenosis are common valvular heart disease. Clinically, doctor often use ultrasound to observe the morphology and trajectory of mitral leaflet(ML) to diagnose heart disease.While the MV root point can be used for the multimodality registration of heart, the leaflet is also an important diagnostic evidence. In this paper, we not only proposed an automatic recognition algorithm of MV root point based on Bag of Features model, but also proposed an automatic recognition algorithm of ML from echocardiography based on auto context model. Our algorithms successfully solve the question that the traditional segmentation algorithm can only segment the whole MV, but failed to distinguish the root point and leaflet.The difficulty that MV root point recognition faced is the feature design. Because of the position of MA, LA and LV is relatively fixed, also the fast moving of ML makes the neighborhood of root point contain rich texture, the algorithm uses local context feature and LBP feature as description. Bag of Features model firstly divides the sample image into patches and then extracts features from each patch. Secondly the algorithm uses k-means clustering algorithm to generate a dictionary of visual vocabulary, which also called codebook. Then the BOF histograms of each sample image is calculated by codebook. Finally, the algorithm trains the classifier using the BOF histograms we got in last step. We use histogram intersection kernel SVM as classifier.The segmentation of leaflet faces two major problems. As the rapid movement of leaflet is irregular, how to build an effective model is the first difficulty. On the other hand, the leaflet appears to be broken into several parts during the ventricular systole for the limit of current ultrasound imaging system. It is not easy to segment all the parts of leaflet. Our algorithm treats the leaflet as a set of points so that we can translate the problem into the identification of a point set, which solves these two questions indirectly. Auto context model extracts features from both the local image and the classification image. Through training a sequence of classifiers, auto context model combines the low level image information and high level context information effectively. We adopt LDB feature and local context feature to describe leaflet. The algorithm firstly identifies the leaflet and then uses Bag of Features method to recognize the MA root points. According to the relative position of leaflet and root points, we can locate the leaflet area more precisely. In order to remove the influence of shadow, a thinning process is needed.Our algorithm was tested on RT3DE(Real Time Three Dimension Echocardiography) datasets of ten patients. The experiment shows an accurate result. The average error of the recognition of MA root point is 1.4 ± 2.1 pixels compared with manual expert calibration. The leaflet also obtain a good result both in diastolic and systolic period.