Research On Vessel Segmentation Based On Active Contours

Due to Medical needs such as diagnosis of vascular diseases and assistance on interventionaltreatment, it is increasingly significant and urgent to conduct deep researches on vesselsegmentation of medical images, which has already become a research focus of scholars bothat home and abroad. This dissertation addresses the problem by means of active contours, inconsideration of the particular vascular features. A novel methodology with feature spaceconsisting of intensity and shape index is propsed in combination with classical CV modeland LBF model frames. The dissertation mainly achieves the following results:1) In terms of initial contour setting which is the very first step of active contour models, anautomatic setting method based on fuzzy C means algorithm with spaticalinformation(SFCM) is proposed in the dissertation in order to reduce manual interventionto a large extent.Compared to manual setting method and traditional fuzzy C meansmethod, the new way presents good performance in the aspect of keeping spatialcontinuity of the initial contour, preventing fracture phenomenon effectively. Furthermore,SFCM is capable of improving the convergence speed of subsequent active contouralgorithm, further strengthening resistance to noises.2) Since segmention algorithms based on pure intensity often fail to achieve idealoutcomes,a shape index formula is put forward according to differential geometry theoryand multi-scale vascular shape model proposed by Frangi.Then, the two-dimentionalfeature space can be established which include information of both image intensity andvascular shape, thus offering comprehensive characteristic basis for the followingsegmentaion by means of active contours.3) Based on active contours without edges like CV and LBF models with shape index andintensity taken into account, novel BCV and BLBF active contours,the evolution of whichare driven by energy function based on Bhattacharyya distance are put forward. Theproposed algorithms employ Bhattacharyya distance in active contours instead of the sumof difference square to measure discrepancy between a finite number of empiricalmoments of the “inside”and “outside”distributions. Experiments show BCV models showgreat advantages over classical CV in terms of segmentation accuracy, while BLBFaddresses the segementation of inhomogeneous images properly.