Ultrasound has been increasingly used in surgical procedures of the prostate in recent years. Segmentation of prostate boundaries from ultrasound images is clinically useful in such situations as accurate volume measurement, and tumor margin estimation, and it can also provide real-time targeted image guidance during procedures such as biopsy and ablation. Automatic segmentation of the prostate, however, is a challenging task since the ultrasound images usually have high level of speckle noises due to large amount of random scatters and thus they have a very low signal-to-noise ratio. As a result, physicians have to use manual methods to draw contours of the prostate, slice by slice, in order to calculate prostate volume information. This is a tedious work and apparently it delays the whole clinical procedures. In addition, accuracy of the segmented prostate boundaries cannot be guaranteed due to significant variations among different physicians or with the same physician at different times. In this paper, we will present a top-down approach to the segmentation of prostate ultrasound images using a snake model, as compared to most existing bottom-up methods. Special measures were taken to deal with the high speckle noises and complex shapes of prostate boundaries. In general, median filtering proved to be effective in removing speckle noises. We extensively evaluated most of the existing edge detection methods and found that the logic combination of Laplacian of Gaussian (LoG) and Sobel operator provided the best performance in finding the useful image gradients. Parameters of the snake were dynamically optimized, and the shape information of the prostate was used as a strong guidance during the deformation process of the snake model. Experimental results with several ultrasound prostate images with various levels of noises were presented to demonstrate the effectiveness of the proposed approach.