Research In Medical Image Segmentation Based On Mumford-Shah Model

Image segmentation is a kind of image processing technique and process that divides the image into different regions,each of which owns its distinct feature.One of the major research fields of image segmentation is the medical image segmentation,which is the basic and necessary step for human diseases quantitative and qualitative analysis and visualization.In 1989 Munford and Shah proposed an energy functional model (Mumford-Shah model) in their paper.The Mumford-Shah model attracts a great attention since that model can achieve image segmentation and image de-noising simultaneously.By introducing the level set method into the Mumford-Shah model,Chan and Vese proposed their active contour without edges(C-V model),which uses the image global information but not the image local gradient to optimize the image segmentation globally through minimizing the Mumford-Shah energy functional. However,if the initial level set is located in the smooth region or the image is concave,it will be need more iteration time for the C-V model to reach the convergence state.Meanwhile it is necessary for the C-V model and other curve evolution algorithms to reinitialize the signed distance function(i.e.SDF),which is a time consuming procedure.Based on the problems mentioned above,this paper presents an improved Mumford-Shah model on combining image's local gradient and penalizing energy for image segmentation.The Mumford-Shah functional is modified so that the energy depends on the image's global information and the image's local gradient information.The introduction of the penalizing energy enables us to select larger time step in implementation.Experimental results show that our improved model is superior to the conventional one both in terms of speed and robustness.Another point worth mentioning is that the 2-phase MS model mentioned above cannot effectively segment medical images with multi-object such as MRI.Though multi-phase Mumford-Shah model can be employed,it is very time consuming and has bad accuracy for the overlap of multi-contour.In addition,the Mumford-Shah functional is not a convex problem.Thus the solution is sensitive to the initial contour and can be trapped by the local minimum but not the global one.Thus in this paper,we propose another improved Mumford-Shah hybrid model coupled with fuzzy C-means(i.e.FCM) clustering information.In order to achieve the goal of extracting region of interests(i.e.ROI) in medical image with more than two objects, firstly we make a coarse segmentation by using FCM clustering.Then we use the fuzzy membership degree from FCM for two purposes.One is that fuzzy membership degree is utilized to initialize contour placement.This auto-selection of initial contour greatly fastens the contour's convergent speed and enables the avoidance of trapping by local minimum.The other is that we incorporate the fuzzy membership degree into the fidelity term of the Mumford-Shah model for multi-object segmentation with higher accuracy.