Research And Analysis Of MR Image Clustering Based On Soft Subspace Algorithm

Magnetic Resonance Imaging(MRI)occupies an important position in modern medical research and clinical diagnosis.Because MRI uses electromagnetic signals to reconstruct human body information,its imaging principle is complex,and the image information obtained by MRI is large.The wide application of digital imaging leads to the surge of image data.Obviously,it is time-consuming and laborious for doctors to recognize mass image data only by relying on relevant doctors,and the long work of doctors also tends to cause missed diagnosis and misdiagnosis.In order to improve work efficiency,Computer-Aided Diagnosis(CAD)plays an important role.CAD can provide a second opinion for radiologists,reducing the time required for image reading.Since image segmentation is a crucial step in CAD technology,in recent years,a large number of researchers are committed to effective segmentation of MR images.However,the influence of rician noise and uneven gray scale in the process of MRI imaging brings great challenges to the effective segmentation of images.Soft subspace clustering algorithm can reflect the compact type and difference between attributes and cluster classes,and it has been well applied in MR image segmentation.However,due to the weak ability of the algorithm to balance the local search and the global search,the boundary weakening and the local optimization will occur in the segmentation of MR images.Therefore,this paper will focus on improving the soft subspace clustering algorithm to make it suitable for MR image segmentation.The main contents include the following three points:(1)In the segmentation of MR images,the traditional soft subspace clustering algorithm is susceptible to the influence of the initial clustering center and noise data,resulting in the algorithm falling into the shortage of local optimization.A soft subspace clustering algorithm(BSOSSC)optimized by brainstorm algorithm is proposed to segment MR images.Firstly,the algorithm designs an objective function which combines relaxed boundary constraint with generalized noise clustering.Secondly,in the process of clustering,the subspaces of clustering are found by using the membership degree calculation method designed.Then,the given index is used to adapt the clustering task of the subspace.Finally,the brain storm algorithm can effectively balance the advantages of local search and global search,optimize the soft subspace,and make up for the shortcomings of the existing algorithm that is easy to fall into the local optimization.Experiments in the Berkeley natural image data set,the simulated brain data set and clinical medical MR images provided by McGill university show that the proposed algorithm provides better segmentation results than the traditional algorithm,which verifies the robustness of the proposed algorithm.(2)In order to improve the adaptability of soft subspace clustering algorithm in segmentation of MR images,a soft subspace clustering algorithm(AMBSOSSC)based on the optimization of adaptive brainstorm algorithm was proposed.The algorithm firstly calculates the subgradient and dynamic step size of the brainstorm optimization algorithm.Then,the adaptive selection strategy is applied,and the membership calculation method in BSOSSC algorithm is adopted.Finally,in the clustering process,the adaptive brainstorm algorithm is used to optimize,so as to improve the self-adaptability of MR image segmentation algorithm.The comparison experiments on the natural image data set,the simulated brain data set and the clinical medical MR images show that the proposed algorithm has better robustness and higher execution efficiency,which verifies the advantages of the proposed algorithm.(3)Aiming at the two problems still existing in the improved soft subspace clustering algorithm:① High time complexity due to the calculation of the repeated distance between the clustering center and the local adjacent window pixels;②Adjacent Windows usually destroy the real local spatial structure of the image,resulting in poor segmentation of the MR image.The algorithm first defines a multiscale morphological gradient reconstruction operation to obtain a super-pixel image with accurate contour.Secondly,the obtained superpixel image simplifies the number of pixels,thus effectively simplifying the complexity of the original image.Finally,the super-pixel image is used for soft subspace MR image clustering to obtain the final segmentation results.Experiments on the Berkeley natural image data set,the simulated brain data set and clinical medical MR images show that the proposed algorithm provides better segmentation results and takes less time than the traditional algorithm.