Algorithm Of Prostate Segmentation On MR Image

Radiation therapy is an effective means of treating prostate cancer, whose key point is to locate prostate cancer precisely, including the location and physical size of prostate cancer, therefore precise segmentation of prostate and its surrounding organs is of critical importance. The manual segmentation of prostate on the MR images is still commonly used, but due to variation of the prostate shape, fuzzy boundary with adjacent organs and changeable volume of prostate during the treatment, the manual segmentation is subjective and time-consuming although with high precision.In the manual segmentation of prostate, experts usually identify prostate boundary on the image based on their previous experiences. Multi-atlas method is similar to this manual way. Experiences of the position and shape of the prostate remain on the label image on atlas. When more atlas, much richer experiences and atlas images similar to test image are easier to be found. Combining atlas image by using difference weights, we could reconstruct the test image approximately. Labels corresponding to these atlas images are fused and transmitted to the test image to accomplish segmentation of test image. Multi-atlas segmentation can help researchers obtain more experiences on positions and shape, which has a great influence on precision of the final segmentation result. In addition, based on more atlas, we could be in a better position to adapt to segmentation target with more complicated changes with robustness. The present study, therefore, segments prostate using the way of multi-atlas.In order to obtain reference to our segmentation results, this paper measured image similarity using normalized mutual information so as to select atlas. We then used popular global weighted voting method to fuse label images, with the weight calculated by using normalized mutual information, to accomplish segmentation finally.Atlas selection means identifying those in atlas set with the highest similarity to the test image based on certain algorithm. Image, as a high-dimensional data, necessitates that the primary feature information is retained while secondary information ignored. Since the atlas selection is made based on the similarity between test image and atlas image, with the main purpose of identifying atlas image similar to test image in features of prostate shape, it is necessary to maintain invariance of the neighboring data in reducing dimension. After dimension reduction, the intrinsic distance of data could be represented by a simple Euclidean distance in low-dimensional space. Meanwhile, we hope, after projection matrix calculation, we would not bother to make recalculation for those new sample data to solve problems outside the sample. Locality Preserving Projection can meet this requirement, which is characterized with linear projection, local neighboring domain relationship maintenance, solution of outside-of-sample problems, as well as fast calculation. The paper examined the application of Locality Preserving Projection in atlas selection.Since in atlas selection, prostate doesn’t contribute a lot to similarity, it is very possible that prostate in the selected atlas image is not similar to that in the test image; It is also very possible that atlas image with similar prostate to that in test image is not selected due to interference of surrounding tissues with obvious dissimilarity. So in calculating the similarity it is necessary to limit the prostate region, which is the indicated part in label image. This paper also studied the process of using label images to influence dimension reduction and projection to increase contribution of prostate to similarity calculation, so as to achieve better precision of selecting atlas image.Label fusion refers to fuse the selected label image, whose fusion algorithm directly affects the segmentation results. Global correlation cannot display local difference information in prostate anatomy under different circumstances. Local weighted voting arithmetic and Patch-based Method can solve this problem. Local weighted method considers voxels of atlas image and test image, and calculates the similarity of the region surrounding the voxels, and then gives the resulting weight to the voxels; but this algorithm demands high pre-registration standards. Patch-based Method to a certain extent overcomes this problem. This method searches for more similar blocks within a certain range around the voxels, and Patch-based Method uses only sum of squared distance to measure similarity, feasible to cases involving a large amount of computation.This paper further expanded range of the patch search based on Patch-based Method; within the range of all selected atlas, we searched for the patch corresponding to test image patch and neighboring patch to reconstruct the test image patch and performed sparse restrictions for the atlas image patches involved in the reconstruction to obtain the optimized weight based on Sparse Representation.This paper focused on the following two aspects.1) The present study performed dimension reduction for the image data utilizing manifold learning method and selected atlas in low-dimensional data space; at the same time, we limited the area of interest based on label image, to increase contribution of prostate into image similarity measure.2) We utilized patch corresponding to test image patch and neighboring patch on all atlas images to reconstruct the test image patch based on weighted patch and sparse representation, with an aim to achieve the best results using the best and lest atlas image patches.We conducted this experiment on 50 sets of training data and 18 test images to validate the proposed method. Segmentation results using Dice coefficient and Hausdorff distance was compared with segmentation results of test image provided by experts to illustrate the effect of the proposed method. The experimental results show that atlas image selection and image fusion algorithm proposed in this paper, compared with atlas selection and fusion algorithm based on normalized mutual information, has improved the effect to certain extent.