Auto-Detection Of The Optimal Cross Sections In 3D Echocardiographic Images

Nowadays real-time three-dimensional echocardiography (RT3DE) is a powerful tool for clinical diagnosis as it is able to capture movement of the whole heart and display intra-cardiac structures by choosing an arbitrary cross section (or imaging plane). In most cases, the echocardiographer makes a diagnosis by studying several important cross sections which are defined as the optimal cross sections. However, it is a time-consuming and tedious task for echocardiographers to manually search the optimal cross sections by the existing RT3DE systems or off-line three-dimensional (3D) software. To assist doctors with congenital heart disease (CHD) diagnosis, we design an automatic detection of the optimal cross sections in three-dimensional echocardiographic (3DE) images. Automatic detection, based on computer programs, might offer better accuracy and repeatability and save time. Furthermore, it can be a prior procedure for measurement and the computer-aided diagnosis. Since there is no reference related to this study at home and abroad, our research on the auto-detection of the optimal cross sections in 3DE images is the initial work. The main contributions of this dissertation are summarized as follows:â‘ Inspired by the techniques of image retrieval and manual search of the optimal cross sections, we present a method to search the four-chamber plane (4CP) in the apical dataset according to the characteristic of the apical dataset. Firstly, a typical end-diastolic (ED) four-chamber (4C) image is chosen as a template. Secondly, in order to find the 4CP in a 3D apical dataset, an image database is built by extracting cross sections from the ED volume of the dataset. Then by image retrieval, the most similar image of the template is retrieved from the stack of images as the 4CP of the dataset. We tried several methods to retrieve the 4CP and the modified algorithm based on mutual information performed best.â‘¡To cope with the misalignment error of each datasets, a fast coarse-to-fine strategy is developed to retrieve the 4CP from the stack of images extracted from the 3DE image, suitable for the real-time applications. Firstly, the coarse retrieval is performed by comparing the cumulative histograms of the two images. Next, a rigid registration approach is adopted to find the best match of the template in each image of the reduced image set, resulting from the coarse retrieval. Then, because the rectangular region at the center of a 4C image contains distinct edges, the high frequency details of the image are selected as the features for the fine retrieval. The other optimal cross sections are obtained, according to the spatial relationships between themselves and the 4CP. Our method can automatically and rapidly detect the optimal cross sections in the apical datasets of different subjects by using only the corresponding template images in one normal ED volumetric data.â‘¢To perform real-time auto-detection, an algorithm based on 3D registration is designed to detect the optimal cross sections in 3DE images. Firstly, instead of the whole ED volume, only part of it is selected as the template, which could be more robust to the change of the size or shape of the heart and reduce the time of the registration. Secondly, as the automatic segmentation in the 3DE image is difficult to perform, mutual information based registration is adopted. Since the positions of the optimal cross sections are not fixed or highly accurate, the float template image is aligned with the reference image by 3D rigid registration. And Powell method is applied to maximize the mutual information measure. This registration-based method can detect all the optimal cross sections of a 3DE image by aligning the template image with the 3DE image. And there are not too many problems to apply this scheme for the auto-detection of the optimal cross sections in the parasternal or subcostal datasets.