No abstract

In this paper, we propose a novel method for the automatic detection of fetal head in 2D ultrasound images. Fetal head detection has been a challenging task, as the ultrasound images usually have poor quality, the structures contained in the images are complex, and the gray scale distribution is highly variable. Our approach is based on a deep belief network and a modified circle detection method. The whole process can be divided into two steps: first, a deep learning architecture is applied to search the whole image and determine the result patch that contains the entire fetal head; second, a modified circle detection method is used along with Hough transform to detect the position and size of the fetal head. In order to validate our method, experiments are performed on both synthetic data and clinic ultrasound data. A good performance of the proposed method is shown in the paper.

The Nuchal translucency (NT), which is the collection of fluid at the back of the fetal neck, is related to chromosomal defects and early cardiac failure in first trimester of pregnancy. In clinic, the thickness of NT is used as an important marker in prenatal screening, and is manually measured by sonographers in the mid-sagittal plane. In this paper, an automatic method based on dynamic programming is proposed to detect the thickness and area of NT in the mid-sagittal plane. Furthermore, the volume of NT in the whole three-dimensional ultrasound data is also measured. A novel cost function for dynamic programming is proposed and results in higher accuracy of NT border detection. As the nuchal translucency is the collection fluid part, higher dimensional markers of NT possess more potential to represent the amount of the fluid.

A novel approach is presented to automatically segment the left ventricle in fetal echocardiograms. The proposed approach strategically integrates sparse representation, global constraint, and local refinement algorithms into an active appearance model (AAM) framework. In the training stage, we construct an enhanced AAM texture model to deal with the speckle and texture ambiguities. In the segmentation stage, the initial pose is located by a sparse representation method. Globally constrained points and local features with high clinical relevance are effectively incorporated into the AAM framework to make the model converge toward a desired position. Our proposed approach has been compared with the traditional ASM, the traditional AAM, and the globally constrained AAM on the synthetic and clinical data. The results show that compared with experts drawn contours, the overall segmentation accuracy of overlapped area in the synthetic and clinical images are 84.12% and 84.39%, respectively, superior to those of the other three methods. The experiments demonstrate that sparse representative methods greatly facilitate the initializations and our approach is capable of detecting the fetal left ventricle effectively, offering a better segmentation results.

Nowadays, 3D ultrasound imaging has been increasingly used in clinics for fetal examination. However, it is cumbersome and time-consuming, even for an experienced clinician, to manually locate the fetal head and the mid-sagittal plane. In this paper, we introduce a totally automatic method for fetal head detection, which is based on a shape model and the marginal space learning framework. We approximate the shape of the fetal head as an oriented sphere defined by 7 parameters, turning the detection task into a process of parameter estimation. As the number of hypotheses increases exponentially with the dimensionality of the parameter space, exhaustive searching is computationally complex and time consuming. To reduce the number of the test hypotheses, we use a marginal space framework, which learns the parameters on sub-spaces in a sequential way. Haar features and steerable features are used in the learning based method. Once trained successfully, our method can be used to locate the fetal head from 3D ultrasound images, with no need for any other information.