Study On The Novel Methods Of Cell Image Segmentation Based On Snake Model

Aim:With the rapid development of computer techniques, computer medical images have played an increasingly important role in clinical diagnosis and treatment. Segmentation of structure from the medical images and construction of the accurate geometrical models for the structure are most challengeable tasks, among which edge detection and image segmentation are of the critical branches. Judgment and study of cell images including the shape and area changes of the same type of cells under different physiological, pathological and experimental conditions, can provide new scientific foundation for pathological analysis and disease diagnosis. But with kinds of shape changes of the cells, the large numbers, and stagger distribution, the analysis results heavily depend on the human doctors' empiricism and visional estimation, most of which are of qualitative analysis instead of quantitative analysis, lacking of quantitative description, comparison and analysis. Development of highly-accuracy and highly-efficiency computer aided auxiliary analysis system serves to aid the medical research.In the study of cell structure and shape changes, first comes the problem of cell detection and segmentation, and then the later measurement such as length, width, and area, and still later the components structure and large molecules. Cell segmentation is the basis of stereometry, as well as 3-dimensional reconstruction of cells, which can provide simulation foundation for pathology research and medicine curative effects. This sphere has great potential and the cell segmentation algorithms possess significance in both academic and applied senses.Methods:First of all, this dissertation makes a widely review and survey of the current main methods and techniques of image segmentation, both fully developed and under exploration, and analyzes the respective application spheres, advantages and disadvantages. After careful study of the aspects of aimed medical cell images, we conclude that special sphere knowledge is demanded to guide automatic segmentation to a certain extent.The segmentation strategy needs two steps: in the first step select the targeted sphere of cells; and the second step use proper segmentation methods to accurately extract cells.In segmentation methods, Snake Model with deep physical basis can describe boundaries more down-to-earth, and reflects the biological property of cells.Snake Model is a enclosed curve of object boundary: First an approximate curve corresponding to the object to be segmented is given, namely the initial model; then the curve deforms under the control of inner and image information; finally the curve stops onto the real boundary. The contour deforms under the inner forces and the external forces, expressed in the energy form:csnake ^int ^ ce8 W ^Eml is the inner energy:F Eexit ( image ) is me extema\ energy:0 (3)The calculation of model is the resolution of the minimum value ofEsnake = )l(a(s) | ^ |2 +/?(,) | ^± |2)]/2-1 y(S)I(v(s)) \2]ds (4) o ds dsAccording to Curve Theory, the first order derivative reflects the continuity of the curve, and the second order derivative reflects the tangent continuity of the curve,or the smoothness of the curve. The minimum energy spline is driven by the inner force (shape restraint) and the external force (behavior restraint) toward the object boundary.This dissertation tests the Basic Snake Model segmentation, and comments on its advantages and disadvantages, and then makes assumption of improvement.In order to bring to play its advantages (it emerges the three stages of segmentation into one and obtain a smoothed consecutive curve result), preliminary procedures are commanded to get satisfactory subjects ready for the Snake Model.This dissertation proposes two ways for improvement of Snake Model: the first one, Maintenance Method, aims to meet the need of Basic Snake Model and labors to find an initial contour of high-quality. The main steps are the folio wings:(1) Calculate each pixel's maximum absolute gradient(2) Calculate the judge threshold of maximum gradient(3) Mark the background pixels(4) The initial contours growth(5) Edges thinning, holes filled, and insular points removed(6) Obtain of sub-domain initial contours and make Snake iterationThe strong point of this method is the result is an accurate and smooth curve of the boundary. However, it depends on comprehensive image techniques and the procedures are complex.So we emphasis on the second way, the Innovation Method, with many aspects of reformation measures and new implement techniques: the revision of inner energy, the improvement of external energy, the creation of Dynamic External Force, the regulation of algorithm strategies, and the deformable procedure control measures.(1) Revision of the external energy includes:(ï¿¡> Enlarge the range of the gray value of the edge points.(2) Use Gaussius Function to extend the external energy's incidence.(3) Take new representation formula of the Image Energy as to strengthen its effect against the noises:E image (Vi> = il I V/( Vi > I I COS 9( Vi >> if \0(V t )\< 7t / 2 } (5)(2) Introduction of Dynamic External Force.Even with the above regulations, the image force is still small at places far awayfrom the boundary. Moreover, the algorithm cannot deal with the boundary of sunk area, which should be taken into consideration when devising new external forces. The gradient values are used to determine whether the points are on the boundary or not, can not distinguish noises from the real boundary points. All these factors lead to some bad-behavior results of the final contours which do not coincide with the real boundaries. Some new external force is demanded to settle down the problems.Definition of the new external force has to handle two problems: the direction of and the magnitude. We illustrate kinds of situations including protruding curves, concave curves and sharp turns. As the direction of the new external force changes according to the position of the points and the shape of the contour, we name it asDynamic External Force. The angle 6 between direction n of the Dynamic External Force on point v, and the horizontal direction is defined as :0 = arccos^L (6)vivjkand its magnitude is formalized asv = ,Vf+'^Vf, (7)With this foundation, we take many kinds of exceptional situations into consideration when applying the Dynamic External Force, and put forward the solution or meliorate strategies. For example, the magnitude of the Dynamic External Force should guarantee the points' moving rates as well as preventing them fromtranscending the boundary with a thresholding Vmia :v' ^ v? (8)( 3 ) Algorithm Strategy Regulation and Deformable Procedure Control MeasuresBasic Snake Model is calculated with greedy algorithm, which runs fast, and the selection of points irrelevant to the future choices. The total energy of Snake Model ties to the future possible new points, which does not conform to the greedy selection. So we adopt the Dynamic Programming algorithm, which generates multiplecandidates to the sub-problems, and this guarantees the final result to be the best one of the original problem. The reformed DP algorithm incarnates the simulated annealing strategy. The deformation of Snake contour is divided into two stages: in the coarse adjusting stage, Snake contour moves rapidly toward the object boundary; in the subtle regulating stage, Snake contour readjusts its smoothness and consecutiveness as to wrap the object boundary closely enough. And revise the inner energy function i as:ï¿¡. = \l(d I v, - vM |)2 + (d-1 vw - v,. |)2] (9)E2 = |[| v,₂ - 2vM + v, |2 + | vM - 2vt + v,.+1 |2 + | v,. - 2v,.+l + v,+2 |2 ] (10)In Formula (9), the restrain of the mean distance between two points can prevent them from becoming too close or too far away from each other. Another revisement for higher smoothness in Formula (10).During the procedure, contour deformation gives rise to some problems. The positions and numbers of the points should be readjusted, new points be added, and redundant or out-of-date points be deleted. All the measures are devised in this section. After experiments, CA criteria is set as the suitable terminate condition for the iterations.For multiple objects in the image, the contour needs to be split: determine the contour dividing points, then carry out the contour splitting and connecting algorithm, finally delete the invalid contours.After the iterations of each sub-contours are completed, we use the interpolation method to construct consecutive enclosed contours.The new segmentation methods of this dissertation meet the cell image segmentation evaluation criteria. One estimation parameter, CorrectRate, is definedas:CorrectPixelNum 1/ww /11XCorrectRate =-----------------------xl00% (11)TotalPixelNumAnd we make evaluation of both the algorithm and the segmentation results with the above criteria. The candidate algorithms are traditional differentiator, the Basic Snake Model, and two novel models of this dissertation.In the end we make a conclusion of the whole study and the originality of this dissertation. We predict a vigorous perspective of the future of the segmentation field.Originality:In order to overcome the shortcomings of Basic Snake Model, this dissertation devises the Renovation Method. On the base of measures to expand External Energy's range, strengthen External Energy as well as battle against noises, this dissertation proposes and applies a new Dynamic External Force. The direction of Dynamic External Force is defined and the magnitude is designed so that the contour points can be guided to move toward the real object boundary. The devise of Dynamic External Force is reasonable and has low computation cost.Moreover, this dissertation designs reformed fast Dynamic Algorithm, the strategies of the insertion, deletion and the spacing control of contour points, the splitting and connection of contours, and the detection of fake contours.Conclusions:The experiment results have proved that the two new methods of this dissert have attainted boundary contour with high accuracy. The Innovation Method has overcome the shortcomings of the traditional Snake Model. Compared with the traditional differential methods and other revised Snake Models, our methods have the advantages of self-contained theory, concise concepts, clear procedure, excellent segmentation results, and relatively small amount of calculation. The shortcoming is the complexity of design procedure.