Research Of Prostate Segmentation Methodologies In TRUS Images

Prostate cancer is one of the most common cancers for men. Its mortality ranks the second place in some European and American countries. In recent years, the death rate of this cancer is rising in China. In many diagnostic and treatment procedures for prostate disease, trans-rectal ultrasound is being widely used because of its real time imaging, low cost and no radiation. Such as how to locate the precise position of puncture needles when sampling these issues biopsy with these needles, how to distribute the dose of radioactive substances during the treatment process of prostate cancer and how to measure the volume of a prostate. Prostate segmentation in TRUS images is highly desired in these clinical applications. However, manual segmentation methods often used in clinic has workload, the segmentation results dependent on operator experience, poor reproducibility problems. Study on the method of automatic segmentation of prostate in TRUS images is a hot research topic in recent years. It is still a challenging and difficult task to accurately detect the prostate boundaries due to the speckle noises, low SNR in TRUS and the missing boundaries in shadow areas caused by calcification or hyper dense prostate tissues.This paper presents a novel method of utilizing image feature and priori statistic shapes for segmenting the prostate. Firstly, this approach makes use of the dense sift features which is invariant in scale and orientation of image to locate the prostate from TRUS images quickly by classifying these features. Secondly, selects the optimal model from the multiple mean shape models. During the segmentation process, missing boundaries in shadow areas are estimated by using the shape model. Finally,with the shape guidance, an optimal search is performed by the gray level gradient feature model and the local Gaussian distribution energy functional to minimize for image segmentation. The segmentation of an image is executed in a multi-resolution mode from coarse to fine for higher robustness and computational efficiency.A number of experiments are conducted to validate this method and the result compared with the boundary provided manually by experts, the average Mean Absolute Distance is 1.03±0.27 mm, the average Hausdorff Distance is 3.37±0.65 mm, the average Dice Similarity Coefficient is 91.9 ± 0.8%, the mean Sensitivity is 94.7 ± 2.1%, the average Accuracy is 99.2 ± 0.2%. The segmentation accuracy of the method used in this paper is greatly improved compared with PASM method.