Segmentation And Feature Analysis Of The Fetal Cerebellum Of Three-dimensional Ultrasound

The ultrasound imaging technique is the most frequently used examination method in prenatal diagnosis, due to its merits of noninvasiveness, real-time imaging, low cost and convenient utilization. In the routine prenatal ultrasound examinations, the fetal cerebellum must be examined. The transverse cerebellar diameter and volume are important indices in the assessment of fetal growth and health. While the cerebellar shape and the completeness is the standard reference for prenatal diagnosis of congenital malformation. However, the low signal-to-noise feature of ultrasound images may bring unfavorable effects to the quantitative analysis and diagnosis. In clinical applications, the ultrasound diagnosis confronts several shortcomings, such as that the diagnosis accuracy highly depends on the sonographer’s experience and the feature analysis of three-dimensional data is time consuming. Therefore it is significant to develop automatic feature exaction and analysis techniques for the ultrasound image processing.In this dissertation, we propose a novel and effective computer-aided method to segment the fetal cerebellum on three-dimensional ultrasound volumes and analyze the cerebellar features, without any manual intervention. The extraction of salient features in ultrasound images, the localization of the objective structure, the extraction of the phase symmetry and the segmentation techniques are employed in our research. The studies have been carried out in following four aspects.First, we extract the salient features in ultrasound images. In the central slice of a three-dimensional ultrasound volume, the fetal brain midline is the most obvious feature, due to its high intensities and linear shape. The next is the fetal skull, due to its large area and elliptical shape. Based on these features, this dissertation applies the weighted Hough transform algorithm to detect the fetal brain midline. The Gaussian weighted mask makes the change of the image intensity distribution and improves the accuracy of the straight line detection. Moreover, a constrained randomized Hough transform algorithm is proposed to extract the fetal skull. The constraint points used in ellipse fitting can accelerate the calculation process.Second, based on facts that the fetal cerebellum has a small area, an irregular shape and less distinct contours, an indirect localization method that combines salient feature detections with exhaustive search is proposed. This localization method embeds the salient image features and the priori-knowledge of the cerebellum into a probabilistic classification tree. The possible location of the cerebellum is then restricted in a relatively small area. Afterwards, two kinds of circular filters are applied to perform an exhaustive search and achieve the accurate localization of the fetal cerebellum.Third, to overcome the drawback of the three-dimensional active surface model segmentation algorithm that the segment result is sensitive to the initial deformable model, the localization result is extended to the three-dimensional space and used as the initial model. To overcome the leakage problem of this segmentation algorithm, the directional phase symmetry is proposed to construct the energy function. The directional phase symmetry divides the region-of-interest of the cerebellum into several sub-regions, and adaptively selects the most matched filter direction for each sub-region. This technique obtains more continuous and clearer cerebellar edges that will bring benefits to the segment algorithm, and thus improves the segmentation accuracy.Last, based on the segmented three-dimensional model of the fetal cerebellum, we measure three parameters:the cerebellar volume, the transverse cerebellar diameter and the cerebellar centroid distance. Moreover, we analyze the relation between these parameters and the gestational age, and calculate the corresponding regression functions respectively.