A Research On Kernel Clustering Algorithms For MRI Image Segmentation

The application of medical images in clinical medicine is becoming more and more widespread, that makes the image segmentation which is the fundamental problem of medical image processing and analysis in the increasingly critical. Medical image segmentation refers to the process of partitioning observed image data to a serial of non-overlapping regions, which have similar features, such as grayscale, color, texture, local statistical features and spectrum, etc. Segmented images can be widely applied in various applications, such as tissue volume quantitative analysis, computer-aided diagnosis, computer guided surgery, lesion tissue location and partial volume effect correction, etc.Due to the complexity and varieties of anatomy structures as well as the imperfections of imaging scanner and imaging techniques, obtained medical images will inevitably be affected by lots of corruption factors including additive noises, partial volume effect and intensity bias field. There is still not a perfect auto segmentation algorithm so far, despite the fact that there exist extensive medical image segmenting methods. Various algorithms are only for certain specific conditions or situations to obtain better segmentation.This dissertation has studied the application of kernel based clustering algorithms for segmenting two-dimensional MRI (Magnetic Resonance Imaging) medical images, adopted correction method for intensity bias field of MRI data to the traditional FCM (Fuzzy C-means Clustering) algorithm and KFCM-II (Kernel-based FCM II) algorithm, and carried out experiments to obtain segmentation results for performance analysis. The experiment results show that, KFCM-II algorithm is not better than FCM algorithm when applied to segment MRI images with low level degraded conditions; As to segmentation tasks with high level degraded conditions, algorithms with correction of intensity bias field should be adopted; If overall balance of accuracy and stability for the segmentation results under different level degrade conditions is the goal, KFCM-II algorithm with correction of intensity bias field should be chosen.The dissertation first introduces the background of medical image segmentation, MRI imaging mechanism, the segmentation target, and the assessment rules for segmentation results. Then, expounds the theory of fuzzy set and fuzzy clustering, goes into details for the classical FCM algorithm, with a analysis for its performance and shortcomings. The thought of kernel methods is set forth with its advantages clarified, focusing on the KFCM-I and KFCM-II algorithm. And then adopting the correction method of intensity bias field to improve the segmentation algorithm performance is proposed. The KFCM-III and its related concept of typical clustering dataset and data classification are introduced. Finally, using the programming skill of VC + + calling Matlab calculation engine to implement the algorithms and carried out segmentation experiments.