| Liver is one of the most important organs in human body, it has a series of important functions such as adjust nutrient substance, keep body warm and expel toxin. But liver disease has high morbidity rate, and causes thousands of patients die every year. Computed Tomography is a common way to exam liver disease, with the fast development of computer technology, computer-aided diagnosis become the most important way to prevent liver disease and lower death rate of liver cancer. Fast and accurate liver segmentation from CT images is one of the most crucial step of computer-aided diagnosis.In this paper, we design an automatic parallel liver segmentation method based on GPU for abdominal CT sequences. First, SOM-based vector quantization(VQ) method is used to segment CT images in gray-domain. We enhance the edges first, and divide image into edge pattern and non-edge pattern vectors. Then use non-edge pattern vectors as training samples to train SOM network, after convergence of SOM, the coefficient of network is used as codebook to cluster non-edge pattern vectors. Pixels in edge pattern vectors are classified based on the results of non-edge pattern vectors. After VQ, all the vectors are divided into several vector patterns, each pattern corresponding to a gray category, then the image is divided into several categories in gray-domain. Since liver and some other tissues in abdomen have similar gray values, we cannot segment liver only depends on gray information. But organs with same vector pattern don’t connect with each other in spatial domain. So in order to obtain liver organ, we implement spatial-domain segmentation algorithm on the segmented images to get area and boundary information of each connected domain, and use these information to distinguish liver from other organs. For improving the computation efficiency, we design several parallel sub-algorithms for each segmentation step that mentioned above, including parallel enhancing edges algorithm, parallel classification of sub-blocks into edge and non-edge pattern vectors, parallel training of SOM network, parallel clustering of non-edge pattern vectors and parallel connectivity discriminating algorithm.Ten CT sequences are used for testing the proposed method and all the parallel algorithms are implemented on GPU using the parallel programming language CUDA. The experiment results show that our method can get a good segmentation accuracy, and especially perform well on computation efficiency, only take 0.45 minutes to segment one CT sequence. The overall speedup ratio of the parallel algorithm is 14.67 compare to serial algorithm. |
