Improved Image Segmentation Method Based On Graph Cuts And Its Applications In Knee Cartilage Segmentation From MR Images

Segmentation of MR image for knee cartilage is medically important for thediagnosis of knee osteoarthritis and other joints diseases. However, in MRI images,the knee cartilage is generally of elongated shape and has low contrast with theiradjacent tissues. So it’s difficult for some traditional image segmentation methods toobtain acceptable accurate results in this case. This thesis presents a new imagesegmentation method for MRI knee cartilage images based on the graph cutsframework. The main works of this thesis are:1) The most important issues in graph cuts framework are the constructions ofthe regional energy and the boundary energy functions. In this thesis, we proposesome image characteristics to form the feature vector for each pixel, and map theimage to a labeled one by clustering these feature vectors. We calculate the histogramdistances of each pixel to the target and background seed points, which are used toconstruct the energy functions.2) In order to integrate the spatial information into the segmentation algorithm,this thesis presents a new method for the connectivity enhancement. For each pixel,its shortest path and shortest distance to the seed nodes are calculated by using thepath weights, which is defined based on the similarity between two nodes. Theshortest path and the shortest distance information are used to construct the energyfunctions in our method. It can improve the homogeneity within the target andbackground classes, increase the difference between these two classes, and it can alsoenhance the connectivity of the target class. This thesis also made some improvementsin the human-computer interactive operation by using the shortest path information,which helps to reduce the interaction workloads.3) As the MR knee cartilage image generally has some elongated structures, thisthesis proposes to first cluster the initial regional energy into two classes. Then theregional energy is modified by the distances of each pixel to the cluster centers. Wehave tested our algorithm on some knee cartilage MR images with the manualsegmentation results provided by some experienced doctors. The experiment resultsshow that our method can effectively solve the distortion problem, which is generallyreferred as “shrinkage bias” in the graph-cut segmentation algorithm. This thesis alsodoes some detailed analysis of the proposed algorithm’s parameter settings and a setof optimal parameters for the MR knee femoral cartilage image segmentation are given.