DE102008036812B4 - Method and device for segmenting an object from an image data set - Google Patents

Abstract

The present invention relates to a method and a device for segmentation of an object from an image data set. The method uses an implicit ASM for segmentation. The adaptation of the active contour (1) to the image data is achieved by a special technique in which within a narrow range around the active contour (1) probabilities are calculated with which halftone dots of the image data set lie on the edge of the object. On projection beams (12, 13) which end on the same grid point (16) of the active contour (1), only the grid point with the highest probability is then selected in each case. With the proposed method and the associated device a robust segmentation is achieved even for objects with strong curvatures.

Description

The The present invention relates to a method and a device for segmentation of an object from an image data set in which from training data first an implicit statistical shape model of the object is created and then segmentation of the object by fitting a active contour to an object contour in the image data set, where the form model limits deformation of an active contour.

in the The field of medical imaging represents the segmentation anatomical Structures such as the liver from image data sets a fundamental task Based on the segmentation result, the size or the volume content of the object can be determined. These sizes are For example, in oncology important to tumor growth and thus the Therapy success monitor. A segmentation is also for the planning of an intervention very important, for example, for a planned Organ transplant.

The Segmentation of complex structures, in the present patent application also referred to as objects, represents a challenging task represents, for the numerous algorithms are known. When using algorithms, alone on the evaluation of intensity values or derived from them Characteristics of the image data, difficulties arise Image data with low contrast or missing edges. In addition, pathological Structures and strong noise, for example due to the reduced Radiation dose in computed tomography (CT), false edges and uneven feature statistics cause algorithms that affect edge detection or homogeneity criteria for regions based.

S. J. Lim et al., "Segmentation of the Liver Using the Deformable Contour Method to CT Images ", PCM 2005, LNCS, 3767: 570-581, 2005, describe a 2D segmentation of the liver, in which initially a combination a threshold technique with morphological operations on different image resolutions is applied. The resulting initial contour is then on the Search for a smooth path in the gradient magnitudes of the image data refined.

Liu et al., "Liver Segmentation for CT Images Using GVF Snake ", Medical Physics, 32 (12): 3699-3706, 2005, use an active contour based on a gradient vector field (GVF) for segmentation of the liver. About a so-called Canny edge detector An edge image is generated which is preprocessed to be erroneously excluded concave regions prior to the calculation of the gradient vector field eliminate.

Heimann et al., "A Statistical Deformable Model for the Segmentation of Liver CT Volumes ", 3D Segmentation in the Clinic - A Grand Challenge, MICCAI 2007 Workshop, 2007, describe a 3D segmentation the liver by fitting an Active Shape Model to the image data. This is done according to the known technique by Cootes et al., "Active Shape Models - Their Training and Application ", Computer Vision and Image Understanding, 61 (1): 38-59, 1995. The explicit Active Shape Model comes with a main component analysis (PCA: Principal Component Analysis) to training contours, consisting of sets of corresponding landmark points. New Consistent contours with the shape model are then given by a linear combination the middle contour and the largest eigenmodes expressed which are obtained by the PCA. To adapt the shape model to the image data are for each landmark intensity profiles towards the surface normal the currently active contour with different shifts certainly. A so-called neighbor Neighbor classifier is then used to the point along the Normal to obtain the most likely to Contour of the liver belongs. The current contour of the Active Shape Model will then move in the direction the most likely object contour points under consideration the limitations of the Active Shape Model deformed.

One Another approach to 3D segmentation of the liver is in Kainmüller et al., "Shape Constrained Automatic Segmentation of the Liver based on a Heuristic Intensity Model ", 3D Segmentation in the Clinic - A Grand Challenge, MICCAI 2007 Workshop, 2007. This approach is also based on an active shape model of the liver. Instead of a neighbor Neighbor classifier, however, here is a heuristic intensity model used to adapt the contour to the image data.

Most 2D segmentation techniques for segmentation of the liver today are based on the so-called Active Contours Approach ("Snakes"). Here, heuristic approaches are often included to prevent the formation of impermissible contours, for example, to remove concave areas. However, this limits the applicability of these techniques. Most known 3D liver segmentation techniques today use an Active Shape Model. These are so-called explicit Active Shape Models, where the contours are represented by landmarks. Such a representation can normally only deal with a rigid topology, so that, for example, a splitting or merging of object areas is not possible. Also known free-form deformations, which are often included in the models to loosen constraints imposed by the shape models, require a special approach to avoid self-intersections. The use of explicit Active Shape Models causes problems, especially with objects, that have large curvatures. In this case, valuable image information can easily be overlooked due to the often coarse distance of the landmarks points as well as strong curvature of the contour.

R. Goldenberg et al., "Almost Geodesic Active Contours ", IEEE Transactions to Image Processing, Oct. 2001, Vol. 10, no. 10th Pp. 1467-1475, describe a numerical system for implementing the geodesic active contour model. For the segmentation is a level set method used where the for the Segmentation Interesting 0-level set function as a distance function can be viewed. To reduce the computational effort is just looking at a narrow band around the active contour.

The DE 10 2006 017 113 A1 describes a method and apparatus for detecting vessel boundaries in an image using edge detection. The edge detection is based on the change of the intensity over a certain distance, for example by observation of projection rays emanating from a seed point. The search is done at discrete points. An initial vessel boundary is determined based on the detected edges and a shape descriptor applied to the initial vessel boundary to determine a final vessel boundary.

The The object of the present invention is a method and to provide a device for segmenting objects in image data sets, a robust segmentation even of objects with strong curvatures or small thickness allow.

The The object is achieved with the method and the device according to claims 1 and 5 solved. advantageous Embodiments of the method and the device are the subject the dependent claims or can be taken from the following description.

at the proposed method for segmentation of an object an image dataset is not an explicit but an implicit one Active Shape Model used. The implicit Active Shape Model own statistical form model is first in a known manner Training data created. Then the segmentation takes place by approach or adaptation of an active contour based on the Active Shape Model to a probable object contour in the image data set. Here can it's both a 2D and a 3D image dataset, where the shape model and the active contour then fit respectively 2D or 3D training data are created. The approach of the active contour takes place in the present process in several approximation steps. In this case, a range of equal distance is repeated according to a Band around the image data set superimposed active contour set. This results in an outer boundary - outside of the area enclosed by the active contour - and one inner limitation - within of the area enclosed by the active contour. Within the defined area of equal distance around the active contour the individual grid points of the image data record signed distance values which is a measure of the Euclidean Display the distance between the respective grid point and the active contour. Get points within the range and outside the contour a positive distance value, points within the range and within the contour get a negative distance value.

In this case, the individual pixels of the image data set are understood by the grid points. In addition, for each grid point in the area, a probability is determined with which the point lies on the contour of the searched object. This can be z. B. done with the above-mentioned next neighbor classifier. In this case, an intensity profile is sampled from the image data set for each raster point, the direction of the profile being given by the gradient vector of the signed distance function. Illustratively, this gradient vector corresponds to points on the contour of the Active Shape Model of the normal to the contour. Based on the intensity profile, each grid point in the area is assigned a probability. From the outer boundary of the area, the inner boundary of the area and local minima of the distance values within the area, projection beams are considered in the direction of the contour of the Active Shape Model that end up perpendicular to the active contour. Among all raster points lying on a projection beam, the raster point which belongs on this projection beam with the highest probability of limiting the object to be segmented is determined. For each pair of projection beams that end on the same grid point of the active contour, the two grid points determined in this way are compared in their probability and the grid point with the highest Probability selected. This is done for all considered projection beams. The projection beams are chosen so that they include all halftone dots of the image data set within the same distance around the active contour. The grid points selected in this way give a probable object contour to which the contour of the Active Shape Model approximates the constraints of the underlying shape model. This procedure is repeated in each case in the next approximation step until a sufficient approximation or adaptation of the active contour to the object to be segmented is achieved, ie until the difference between the active contour and the object contour determined in the corresponding approximation step falls below a predefinable threshold value. In this case, the active contour then reflects the boundaries of the object, taking into account the Active Shape Model, so that the object is segmented. The computation of the signed distances, the probabilities and the projection paths does not have to be done in every approximation step, but only if the contour of the Active Shape Model leaves the area in which the distances etc. were calculated last time. If this is not the case, it is simply sufficient to approximate the contour of the Active Shape Model to the next already determined points with the highest probability.

at the proposed method and apparatus in which this Method is implemented, thus becomes an implicit Active Shape Model (implicit ASM) used a robust segmentation allows. By abandoning a landmark-based model circumvented the difficulties associated with such models. However, the adaptation of an implicit ASM to the image data more complicated as there are no landmarks available in the direction the most likely boundary of the object can be moved. By the proposed approach based on the same area Distance and the carried out in it Provisions become the approach or adaptation with high reliability. The proposed procedure leads especially for objects with strong curvatures to better results as a landmark-based model because there are no grid points during the exam get lost. Furthermore you can without further action also very thin Structures or objects without the risk of self-duplication be segmented.

wobei ψ - durch Mittelung aller vorzeichenbehafteten Distanzfunktionen der Trainingsdatensätze bestimmt wird und ψ_i die Eigenmoden der PCA der Trainingsdatensätze sind.Preferably, the proposed method employs an implicit ASM as described by Rousson et al., "Implicit Active Shape Models for 3D Segmentation in MR Imaging", MICCAI 2004, 3216: 209-216, 2004, wherein the implied ASM is integrated into the so-called level set methodology. To adapt such an implicit ASM, the active contour is defined as a zero level set of a level set function φ that evolves on the one hand according to the image content and on the other hand is limited in acceptable forms by another level set function ψ generated from the statistical shape model , This level set function ψ is composed of the middle level set function ψ - and a weighted sum of the N largest eigenmodes ψ _i , ie

where ψ - is determined by averaging all signed distance functions of the training data sets and ψ _{i are} the eigenmodes of the PCA of the training data sets.

The adaptation of the active contour taking into account the ASM is determined by the following energy functional indicating the quadratic difference between both level set functions:

A is an affine transformation which adapts the contour of the shape model embedded in ψ as a zero-level set to the active contour embedded in φ as the zero-level set, which is aligned on the image content. The components of the vector λ are the weights of the eigenmodes. Since only the adaptation of the two contours, ie the zero level quantities of φ and ψ, is of interest, the Dirac delta function δ is introduced into the above energy function. The integration space is referred to as Ω ⊂ R ² (or Ω ⊂ R ³ for the three-dimensional case).

This equation is iteratively minimized with respect to φ as well as A and λ. For the individual weights λ _i as well as for the components of the transformation A, which are rotation angles about the coordinate axes as well as scaling and translations, the derivative of E for the respective weight λ _i or for the respective component of A is calculated and therefore included Gradient descent performed. A rule for the optimization of E with respect to φ results according to the variational calculation from the Euler Lagrange equation to the above equation (1). The result is a restriction of the partial differential equation which controls the deformation of φ and thus the deformation of the active contour. Further details of such an approach can be found in the already mentioned publication by Rousson et al. Thus, the adaptation of the active contour, ie the level set function φ, based on both the statistical shape model and the image information. While the adaptation based on the shape model is basically known, the peculiarity of the present method lies in the proposed procedure for adaptation to the image information or image data. The determination of the halftone dots on the respective projection beams, which count with the highest probability of object limitation, can be carried out in different ways, preferably on the basis of a so-called appearance model, such as, for example, the nearest neighbor classifier.

Around the active contour to the grid points within the boundary with the highest In principle, the probability of moving probability can be the gradient of the probability map be calculated. As the catchment area of the gradient of a probability map is limited, an additional measure is preferably taken. This measure exists in the calculation of the gradient vector field (GVF) of the probability map. As a result, the gradient vectors, which are otherwise only close to Raster points with a probability greater than zero exist over the diffused throughout the same distance. In this way creates a much larger catchment area the gradient vectors to guide the active contour. The Calculation of such a gradient vector field is, for example, off Xu et al., "Snakes, Shapes and Gradient Vector Flow ", IEEE Transactions to Image Processing, 7 (3): 359-369, 1998. At this embodiment the GVF term becomes the partial differential equation for deformation introduced the active contour and competes with the term, the the restrictions represented on the basis of the shape model. For example, Paragios et al. Describe "Gradient Vector Flow Fast Geodesic Active Contours ", International Conference to Computer Vision 2001, 67-73, 2001, the introduction of a such GVF terms in the equation for the development of the active Contour.

The proposed device comprises at least one storage unit for the Training data and the image data set, a computing unit for the execution of for the execution the process required steps, a training module to create the statistical shape model based on the training data sets as well a segmentation module that describes the process steps for customization the active contour to the object to be segmented according to the proposed Performs procedure. The segmented object can then be displayed on an image display device can be represented, which can be connected to the device.

The proposed methods and the associated device are suitable especially for Image data sets medical imaging, for example, for segmentation of the liver from tomographic image data sets. such Image data sets can come from different imaging devices, for example. From a computer tomograph or a magnetic resonance tomograph, but are not on the the latter devices limited.

The Proposed methods as well as the associated device are hereafter based on an embodiment briefly explained in connection with the drawings. in this connection demonstrate:

1 a schematic representation of the basic procedure in the proposed method;

2 a schematic representation of the definition of the area of equal distance around the active contour;

3 a schematic representation for determining the probabilities using projection beams and

4 a schematic representation of the proposed device.

The proposed method and the associated device are explained in more detail below using the example of the segmentation of the liver from a CT image data set. 1 shows the process in a highly schematic representation. The explanations here refer to a 2D image data set, so that the active contour represents a closed line and the area of equal spacing represents a band on the central axis of which this closed line runs. However, it is obvious that the proposed method can easily be extended to three dimensions to perform segmentation in a 3D image data set.

at the proposed method, several image data sets of Liver provided by different patients as training records for the creation serve a statistical model of shape. A set of segmentation masks, obtained by manual segmentation of the liver in the training records represent the base of the statistical shape model. Mask points with a value not equal to 0 to the object, d. H. the liver. All other mask points with one Value equal to 0 represent the background. For the analysis of the possible Forms of the liver need first Variations between the individual segmentation masks eliminated which are characterized by similarity transformations caused, d. H. by translation, rotation and isotropic Scaling. Therefore, first one of the masks selected as a reference mask. For all other masks will be then determines appropriate transformations by which the degree of overlap between each mask pair is maximized. This can, for example, with the Technique by Tsai et al., "A Shape-Based Approach to the Medical Imagery Using Level Sets " IEEE Transactions on Medical Imaging, 22 (2): 137-154, 2003. Translation, Rotation and scaling are for each transformation is optimized with the technique of gradient descent. Because the procedure is a larger number requires iterations and is prone to local minima Translation, rotation and scaling normalized before the technique Gradient descent to refine the initially only coarse overlay is used.

in the Following this mutual alignment and appropriate scaling The individual segmentation masks become the ones through these masks represented Forms into an implicit representation transformed. This is done by generating a sign-afflicted Distance function for any mask where each grid point is positive or negative Euclidean distance associated with the limitation of the form shown becomes. Grid points within the form receive a negative Sign, grid points outside the background form of the form is a positive sign.

ausgedrückt, wobei λ dem Vektor der Modengewichte λ_i der Eigenmoden, ψ - dem Mittelwert über alle Abstandsfunktionen und ψ_i den N signifikanten Eigenmoden entsprechen, die aus der PCA erhalten werden.Subsequently, a principal component analysis (PCA) is performed with the distance maps (distance map) obtained thereby to analyze the remaining variations due to different forms of the liver. New forms are implicitly called the zero level set of the level set function

where λ corresponds to the vector of the mode weights λ _{i of} the eigenmodes, ψ - the mean over all distance functions, and ψ _{i to} the N significant eigenmodes obtained from the PCA.

For the use of this image training model created from training data, a further level set function φ is introduced, which develops as an active contour according to the image content and whose development is limited by the level set function ψ of the shape model. The restriction of the adaptation of the active contour by the shape model can be described by the energy functional already explained above, which indicates the quadratic difference between the two level set functions:

This equation is minimized in an iterative manner until a static solution results, ie the values of the mode weights λ _i and the values of the transformation parameters in A (rotation angle, scaling factor, translation values) and in particular the active contour level set function φ no longer change. The active contour then corresponds to the segmented object, which, for example, can be displayed on a monitor.

The Special feature of the proposed method is the combination an implicit ASM with a special approach to the Adaptation of the active contour to the image information. This adaptation will be explained in more detail below with reference to the present example.

In an explicit ASM so-called Appearance Models are usually used to guide the deformation or development of the active contour. This includes determining intensity profiles at each landmark in a direction normal to the surface of the object during a training phase. Assuming a normal distribution, the appearance model is obtained by calculating the mean and covariance matrix of the profiles. This principle has also been extended to random distributions by determining both edge and non-edge profiles and using a neighbor neighbor classifier. During segmentation, each landmark point is then moved in its normal direction towards the most likely boundary point of the object. Compared with simple Chen edge detectors is achieved with this known approach based on an explicit ASM greater robustness of the segmentation.

In contrast to this known technique, the present method and apparatus employ an implicit ASM into which an image term based on an appearance model is integrated. The procedure for adapting the active contour to the image data is described in the schematic diagrams of FIG 2 and 3 explained in more detail.

2 shows the active contour 1 , which is represented by the zero level amount of Φ. The image data itself, to which this contour is superimposed, can not be seen in this and the next figure. The active contour 1 is part of the ASM and represents in its initial form, for example, the average over the forms of the liver obtained from the training data. For the adaptation to the image data in the proposed method is initially a narrow band equal distance around the active contour 1 placed. The width of the tape may be, for example, 5-30 mm. This band is in the figure by the outer boundary 3 and the inner boundary 2 shown in dashed lines. The distance of the outer boundary 3 to the active contour 1 and the inner boundary 2 to the active contour 1 corresponds to the distance r, denoted by the reference numeral 4 is marked in the figure. Each halftone dot of the image data set that lies within this band becomes a positive or negative distance value, depending on the position and the distance to the active contour 1 , assigned. Thus, for example, receives point 5 a positive distance value 7 and point 6 a negative distance value 8th , The distance map obtained in this way can be created, for example, via the so-called fast marching algorithm. For all grid points within the narrow band, ie, for example, for the points 5 . 6 and 9 , intensity profiles are determined at their corresponding position within the image, whereby the intensity profiles are determined in the normal direction, which is given by ∇Φ. This normal direction is used in the 2 exemplary at the point 9 through the vector 10 indicated. A suitable appearance model calculates the probability for each grid point that it belongs to the boundary or surface of the object to be segmented. As an Appearance Model, for example, a so-called k Nearest Neighbor (k-NN) classifier can be used. Instead of a k-NN classifier, of course, other techniques, such. Mean vector & covariance matrix, which also provide a continuous probability.

Subsequently, the raster point with the highest probability on the respective projection beam is determined in each case along projection beams. The projection beams start on the edges of the narrow band (boundary lines 2 . 3 ) and at local extrema of the underlying distance map passing through the dotted line of local minima 11 in the 3 are indicated. The projection beams end vertically on the active contour 1 , For every projection beam that is inside (projection beam 12 ) or outside (projection beam 13 ) through the active contour 1 enclosed areas, a gradient slope or descent is performed on the signed set distance map representing the level set function Φ until the active contour 1 is reached. Only the halftone dot with the highest probability is determined along each projection path. This applies, for example, in the case of the projection path 12 the point 14 and at the projection path 13 the point 15 , All other grid points are no longer considered.

Finally, for all pairs of internal and external projection beams that are on the same grid point 16 on the active contour 1 only the one of the two previously determined grid points is selected, which has the higher probability of belonging to the object boundary. As a result of this algorithm, a map is obtained which contains only the highest probability halftone dots. Assuming that the Appearance Model's real boundary points have higher probabilities than non-boundary raster points, the proposed approach eliminates false boundary points. This will prevent the fitting algorithm from getting stuck in the wrong position in a local minimum.

Since the trapping range of the probability map described above is limited, the guidance of the active contour toward the edge points alone based on the gradient may not be sufficient. Therefore, in the present example, the gradient vector field (GVF) of the probability map is calculated to distribute the gradient vectors over the entire narrow band. This significantly increases the capture area. The GVF term is included in the development of the active contour, as described, for example, in the already mentioned publication by Paragios et al. is known. The GVF term competes with the limiting term previously described on the basis of the shape model derived from the above-mentioned energy functional (equation (1)). The development equation for the level set function Φ of the active contour can then be represented as follows:

Where Φ _{t represents} the time derivative of the level set function, g a stop function, "div" the divergence operator, and ν the GVF field. The first term on the right side of this equation (2), which is weighted by α, contributes to smoothing the contour taking into account the image content by means of the stop function g. The second term, weighted by β, introduces the GVF to drive the contour to the edge points. Finally, the third term, weighted by γ, represents the constraints imposed by the ASM. It yields according to the variational calculation on Φ from the Euler-Lagrange equation of the above energy function (Equation (1)). Further details on the first and second terms may be found, for example, in the above-mentioned paper by Paragios et al. and to the third term of the above publication by Rousson et al. be removed.

4 Finally, shows an example of the apparatus for performing the proposed method in a highly schematic representation. The device can, for example, by a computing station with a computing unit 17 and a data store 18 be formed, which receives the corresponding training records, the image data set as well as all accumulated during the implementation of the method data. A module 19 creates from the training data sets the statistical shape model based on the known process steps. A segmentation module 20 Finally, performs the special process steps of the proposed method, in particular the adaptation of the active contour to the image data, taking into account the ASM. The segmented object may eventually be attached to a monitor connected to the device 21 presented to a user. The segmentation can be done fully automatically after creation of the ASM without any user intervention.

The proposed method works compared to the algorithms, based on landmarks, especially in object areas with strong curvature reliable, where edge points in the known methods due to the coarse distances overlooked the landmarks can be. By shifting the landmarks in the normal direction can here in areas between the divergent displacement paths possibly important image information is not taken into account. Another one Problem occurs, for example, even if a boundary point with the highest probability lies behind another part of the surface.

These Problems are avoided with the present algorithm. at this algorithm becomes probabilities for the entire narrow band calculated so that no single edge point can be overlooked. Indeed have to false border points are removed in order to set the algorithm to prevent in local minima. This is suggested in the proposed Method achieved by all the probabilities determined here back be projected onto the current active contour and only the point with the highest Probability for every pair of projection beams for the adaptation of the active Contour taken into account becomes.

Claims (5)

Method for segmenting an object from an image data set, in which - an implicit shape model for the object to be segmented is created from training data records and - the object is created by stepwise approximation of an active contour ( 1 ) is segmented on the basis of the shape model to a probable object contour in the image data record, wherein for the approach repeated - an area of equal distance according to a band around the active data set superimposed on the image data set ( 1 ), - each grid point ( 14 . 15 ) of the image dataset within the range is assigned a signed distance value which is a measure of a distance of the raster point ( 14 . 15 ) to the active contour ( 1 ), - projection beams ( 12 . 13 ), each of which is defined by an inner boundary ( 2 ) of the area, from an outer boundary ( 3 ) of the area as well as of local extremes ( 11 ) of the distance values and perpendicular to the active contour ( 1 ), the projection beams ( 12 . 13 ) are chosen so that all grid points ( 14 . 15 ) within the range of the projection beams ( 12 . 13 ), - for each of the projection beams ( 12 . 13 ) each of the image data a local intensity profile at each on the projection beam ( 12 . 13 ) grid point ( 14 . 15 ) determines the image data set and a raster point is determined, which is most likely a point on the object contour due to the local intensity curve, - for each pair of projection beams ( 12 . 13 ) on the same grid point ( 16 ) of the active contour ( 1 ), the previously determined halftone dots are compared with respect to their probability of representing a point on the object contour, and the halftone dot with the highest probability is selected, and - the active contour ( 1 ) is approximated to the course of the selected halftone dots, taking into account specifications of the shape model. Method according to Claim 1, in which the approximation of the active contour ( 1 ) to the course of the selected halftone dots while minimizing the equation

where φ is the active contour ( 1 ) and ψ represent permissible shapes of the shape model, λ correspond to the vector of the mode weights of the eigenvectors of the shape model and δ of the Dirac delta function, and A represents an affine transformation that adapts ψ to φ. Method according to one of Claims 1 to 2, in which the determination of the halftone dot ( 14 . 15 ) on each projection beam ( 12 . 13 ), which due to the local intensity curve is most likely to represent a point on the object contour, on the basis of an Appearance Model. Method according to one of Claims 1 to 3, in which, for all halftone dots ( 14 . 15 ) Within the range, on the basis of the local intensity profiles determined for the halftone dots, probabilities are determined with which they represent a point on the object contour, and a gradient vector field is calculated from the probabilities and a spatial distribution of these probabilities given by the position of the halftone dots Approximation of the active contour ( 1 ) by the following development equation:

in the Φ _{t is} a time derivative of the active contour ( 1 g) is a stop function, "div" is the divergence operator, ν is the gradient vector field and α, β and γ are weighting factors. Device for carrying out the method according to one of Claims 1 to 4, which has at least one computing unit ( 17 ), a data store ( 18 ), a shape model module ( 19 ) and a segmentation module ( 20 ), wherein the mold model module ( 19 ) is adapted to enable the creation of an implicit statistical shape model of an object to be segmented from training data sets, and wherein the segmentation module ( 20 ) is designed so that it performs the method steps of the method according to one or more of claims 1 to 4.