Title of Invention	A METHOD FOR SEGMENTING ANATOMICAL STRUCTURES FROM 3D IMAGE DATA
Abstract	This invention relates to a method for segmenting anatomical structures from 3D image data, the method comprising ; setting a starting point in the 3D image data; identifying at least one of at least one known anatomically significant point (6,9) and at least one known anatomically significant surface (11) in the 3D image data; segmenting, proceeding from the starting point, the structure pixel by pixel with a multiplicity of segmentation steps such that an instantaneous distance is determined automatically relative to at least one of the anatomically significant point (6,9) and to the anatomically significant surface (11) in each segmentation step; and establishing at least one of segmentation parameters and a selection of adjacent pixels for continuing the segmentation as a function of the determined distance, taking account of a model topology.

Title of Invention

A METHOD FOR SEGMENTING ANATOMICAL STRUCTURES FROM 3D IMAGE DATA

Abstract

This invention relates to a method for segmenting anatomical structures from 3D image data, the method comprising ; setting a starting point in the 3D image data; identifying at least one of at least one known anatomically significant point (6,9) and at least one known anatomically significant surface (11) in the 3D image data; segmenting, proceeding from the starting point, the structure pixel by pixel with a multiplicity of segmentation steps such that an instantaneous distance is determined automatically relative to at least one of the anatomically significant point (6,9) and to the anatomically significant surface (11) in each segmentation step; and establishing at least one of segmentation parameters and a selection of adjacent pixels for continuing the segmentation as a function of the determined distance, taking account of a model topology.

Full Text	Description Method for segmenting anatomical structures from 3D image data by using topological information The present invention relates to a method for segmenting anatomical structures, in particular of the coronary vessel tree, from 3D image data such as are generated, for example, in CT angiography (CTA). The present method is applied chiefly in the field of computer tomography when recording vessel structures. A great advantage of CT angiography by comparison with other imaging techniques such as magnetic resonance (MR) tomography, PET (Positron Emission Tomography), SPECT (Single Photon Emission Computed Tomography) or the 3D ultrasound technique consists in that, for example, the entire vessel tree of the heart can be recorded with a single CT scan by adding a contrast agent. The 3D image data thereby obtained can be visualized using different techniques. For a quantitative evaluation, in particular a measurement of stenoses or of plaque deposits, it is necessary for the corresponding regions of the vessel structure to be segmented from the 3D image data. This segmentation is performed in a follow-up process on an image computer. The currently most frequently used and commercially available technique of segmentation is the technique of so-called region growing. In this technique, all the respectively adjacent pixels (voxels) are analyzed starting from seed points, which can be prescribed by the user, in the 3D image data, and identified as part of the vessel structure upon fulfillment of specific conditions. As one condition for the membership of the vessel structure, it is possible, for example, to check whether the voxel falls into a prescribed HU range (HU: Hounsfield Units). It is also possible to prescribe for the density gradients between adjacent voxels a highest value above which the adjacent voxel is no longer regarded as part of the vessel structure. The voxels respectively newly identified as part of the vessel structure are used, in turn, as starting points for the next step of analysis or segmentation. In this way, the already identified structure grows three-dimensionally until the complete, prescribable region of the vessel structure is segmented. An example of the use of such a technique for segmenting vessel structures can be taken from the publication by T. Boskamp et al. 'New Vessel Analysis Tool for Morphometric Quantification and Visualization of Vessels in CT and MR Imaging Data Sets', Radiographics 2004, 24, 287 - 297. The known region growing technique operates in may instances satisfactorily, but does not reach all the vessels that are visible to a viewer in a display of the 3D image recording. Whereas, given a suitable 3D visualization technique, the human eye can also still detect the smallest vessels as part of the structure, the segmentation algorithm detects only specific homogeneous, coherent parts in the volume examined. Furthermore, it is also possible for the segmented structure to be washed out in adjacent image areas when they have similar HU values and lie very near to the vessel structures. The cited publication published in Radiographic journal, 2004, 24 (page 287- 297), and entitled 'New vessel analysis tool for Morphometric Qualification and visualization of vessels in CT and MR image data sets', by T. BOSKAMP et al, describes an image processing algorithms and a prototypical research software tool for visualization and quantitative analysis of vessels in data sets from computed tomography and magnetic resonance imaging. In diagnostic angiographic imaging, the traditional projection method of digital subtraction angiography has recently been more and more superseded by the tomographic acquisition methods computed tomographic (CT) angiography and magnetic resonance (MR) angiography. Owing to the three-dimensional (3D) nature of CT angiography and MR angiography data, these methods allow sophisticated visualization of vascular systems, such as volume rendering or maximum intensity projection from any angle, providing information of high relevance for diagnostic as well as demonstration purposes. In addition, CT angiography and MR angiography studies are able to provide quantitative morphometric information, such as vessel cross-sectional areas, volume measures of certain vessel segments, and curvature and tortuosity measures, with high diagnostic value for a number of vascular diseases. To obtain such information from the original CT angiography and MR angiography sections, it is necessary to perform postprocessing steps to produce a description of the vessel path and multiplanar reformation (MPR) images orthogonal to this path. Computer systems providing this kind of functionality should require only a minimum of user interaction while still being able to successfully process images of small or strongly curved vessels or vessels with low or varying contrast to surrounding structures. In addition, it is desirable to obtain an automatic measurement of cross-sectional quantities along the vessel path, together with some indication of how reliable and accurate these results are. Other important features are length measurements along the vessel path and volumetry of user-defined vessel segments. Although a number of system packages providing this kind of functionality are available from manufacturers of radiologic workstations, none of these systems fully satisfy all of the listed requirements (at least from the authors' point of view). Above all, a high accuracy in the determination of the vessel centerlines is a prerequisite for precise cross-sectional rendering and measurements. An inaccurate vessel centerline may result in cross-sectional sections that are tilted against the true vessel path direction and thus in artificially enlarged cross- sectional areas and vessel diameters. In difficult cases where the initially computed centerline is not satisfactory, the ability to improve the result by few additional user interactions is needed. The cited reference provides a computer application prototype addressing these issues. The requirements for the cited invention are (a) high accuracy in the determination of the centerlines of the vessels of interest, (b) automatic measurement of cross-sectional areas and diameters along the vessel path, (c) feedback on the reliability of these results, and (d) a user interface supporting a smooth work flow and integration into the clinical routine. In this article, we describe the algorithmic components used in this development and the prototypical research application VascuVision that makes use of these algorithms and present some examples of clinical applications. The known methods for the determination of vessel centerlines can be roughly grouped into two categories: (a) methods that compute an optimal path connecting two given points by minimizing some cost functional, taking into account external, image-related factors (eg, gray value, local contrast) and internal, path-related factors (length, curvature) (1-4); and (b) methods that perform vessel segmentation first followed by a subsequent skeletonization step that yields the vessel centerlines (5-41). The first type of methods (a term functional approach) offers some advantages in that it avoids the time- consuming and possibly error-prone segmentation step and is typically faster and requires less user interaction. On the other hand, the consideration of path length and curvature may result in a minimizing path that deviates from the true centerline, especially at bifurcations or at locations where the vessel is strongly bent. In addition, the internal algorithmic components used by these methods are not easily perceived by the user, and it is thus difficult to extend such a method to integrate additional user hints in order to improve a dissatisfactory initial vessel path. The second type of method, called a geometric approach, makes use of an intermediate processing result, the vessel segmentation, which can be verified by the user and for which it is easy to integrate interactive extensions that allow improvement of the segmentation in case the automatic algorithm fails. On the basis of the segmentation result, an accurate skeletonization can usually be computed without any interaction, but again it is possible to incorporate user hints, such as explicitly specified skeleton endpoints. The vessel analysis method proposed in the cited reference belongs to the second category and consists of the following algorithmic steps: (a) vessel segmentation based on a region growing algorithm, (b) interactive "premasking" to optionally exclude interfering structures close to the vessels of interest, (c) distance transform-based skeletonization, (d) MPR orthogonal to the vessel path, (e) lumen boundary identification on the orthogonal cross-section images, and (f) morphometric measurements. The first step, region growing-based vessel segmentation, is started by manually placing one or more seed points in the vessel(s) of interest. From these seed points, more and more neighboring voxels are included in the segmentation, provided that certain inclusion criteria are met. The most important criterion is the lower gray value threshold: Neighboring voxels are included only if their gray values do not fall below this threshold. A second criterion provides an adaptive upper threshold, thus avoiding the inclusion of brighter structures (eg, bones in CT images). The upper threshold is combined with a size limit in order to still include small, bright spots, such as small calcifications or small regions with a high concentration of contrast agent. (The location of these small, bright spots is later used in the refined lumen boundary determination step.) A third criterion is the gray level difference of neighboring voxels (gradient threshold). For the successful application of this, it is crucial to select an appropriate lower threshold. Since it is not possible to specify a good threshold value in advance (neither for the algorithm nor for the user), an implementation of the region growing algorithm performs the segmentation for a whole range of thresholds in a single pass by successively decreasing the current global threshold and restarting the growing process, with the present segmentation used as seed voxels. The result of this multithreshold segmentation is presented to the user as a threshold-volume diagram, which indicates the size of the segmentation result corresponding to each threshold in a given interval. The threshold-volume curve shows how the volume of the segmented object in milliliters (vertical axis) increases as the threshold value in Hounsfield units or gray values (MR imaging) (horizontal axis) is reduced. The inclusion of neighboring structures due to leakage of the region growing algorithm is expressed as a sudden step in the threshold curve. Consequently, possible candidates for a reasonable threshold value are easily identified as locations of larger steps in the threshold curve. This diagram, combined with a fast, interactive 3D display of the segmentation result for any selected threshold, thus allows one to find a good threshold with only a few user interactions. In many cases, the first automatic guess provided by the algorithm on the basis of an analysis of the threshold curve already yields an acceptable threshold value. In some cases, a global lower threshold is not sufficient for appropriate differentiation between the vessels of interest and neighboring structures, such as bones or closely passing veins. In these cases, an optional premasking step can be used to interactively exclude such structures. The premasking is based on the watershed transform, which subdivides the image volume into several bright regions separated by darker areas. Once this image transformation has been computed, include and exclude markers can be used to interactively define regions containing the vessels of interest and regions containing adjacent structures. In the implementation, by the cited reference, the watershed transform is applied in three dimensions to the original image, yielding boundary surfaces that lie somewhere in between the bright image regions. Consequently, the result of this step is not segmentation of a vessel but definition of a mask containing the vessel and some surrounding dark regions and excluding adjacent structures of similar brightness. In a subsequent segmentation step that uses the region growing algorithm described earlier, the segmentation can be restricted to this mask, thus avoiding the inclusion of neighboring structures. This use of the watershed transform is in contrast to the gradient watershed segmentation methods, where the transformation is applied to the gradient image, yielding a separation between bright and dark regions. Once the vessels of interest have been segmented, the vessel skeleton is computed by successive erosion of border voxels until a set of connected voxels with a thickness of 1 voxel is obtained. The algorithm used in the cited reference our method is based on a discretized version of the distance transform, is able to correctly handle nonisotropic images, and offers special features to suppress spurious side branches resulting from boundary irregularities. The distance transform assigns to each voxel of the segmented object the minimal distance between this voxel and voxels outside the segmented object. By weighting the discretized distances with the voxel resolution, a correct distance is computed in the case of anisotropic voxels. For the side branch suppression, the algorithm computes the gradient of the distance transform, which is close to 1 at points where the boundary is smooth and close to 0 at endpoints of major skeleton lines. Thus, the magnitude of the distance transform gradient serves as an indicator for the significance of skeleton segments, and thresholding of this gradient can be used to suppress more or less insignificant side branches. It can be shown that the theoretical accuracy of this skeletonization algorithm is 1 voxel distance; in practice, the accuracy of the centerline determination is mainly dependent on the accuracy of the segmentation. The high accuracy of the skeletonization algorithm has been corroborated by phantom studies, in which the maximum reproduction errors of the true centerlines have been measured for stenosis, aneurysm, and bent vessel phantoms with different diameters and orientations. In a total of 85 measurements, the reproduction errors were found to be in the range of up to 1 voxel distance in 42 cases (49%), up to 2 voxel distances in 76 cases (89%), and above 2 voxel distances in nine cases (11%). The vessel centerlines obtained from the skeletonization step can be used as a starting point for the analysis of arbitrary paths in the vessel tree. In an interactive 3D display, the user is able to define an analysis path by selecting two points anywhere on the skeleton. This triggers the computation of the cross- sectional MPR images orthogonal to the connecting path, which are presented to the user and which form the basis of the subsequent morphometric analysis. In addition, a stretched rendered image of the vessel is computed, which appears as if the vessel had been pulled straight and sliced lengthwise. This rendered image can be helpful in obtaining an overview of the vessel, identifying suspicious points, and selecting a current path position corresponding to the cross-sectional MPR image displayed. To perform cross-sectional measurements, it is necessary to identify the vessel lumen boundary on the cross-sectional MPR sections. Although this could be done by simply projecting the result of the primary vessel segmentation onto the cross-sectional plane, it is often advantageous to perform a refined and more accurate vessel segmentation on a section-by-section basis. This refinement allows an adaptation to varying contrast between the vessel and its surroundings and improved differentiation between the vessel lumen and calcifications at the vessel wall that may have been included in the primary segmentation. The cited reference uses two-dimensional (2D) gradient watershed segmentation for the lumen boundary determination. (Given a symmetric point spread function of the imaging modality, this segmentation method is equivalent to the full width half maximum criterion applied to the gray value profile.) This method differs from the watershed transform used in the premasking step in that it operates in two dimensions and is applied to the gradient image of the cross-sectional MPR sections, in which the lumen boundary appears as a bright line separating the darker interior and exterior regions of the vessel. Include and exclude masks defining the interior and exterior of the vessel within a single cross-sectional plane are derived automatically from the primary vessel segmentation. In addition, calcification spots are identified based on (a) a high relative gray value in the interior of the spot, (b) a high gradient value, and (c) the location of small bright spots detected during the primary region growing segmentation step. The final step, the morphometric analysis of the vessel along the selected analysis path, is based on the quantification of the cross-sectional areas and diameters for each cross-sectional section. The cross-sectional area of a single section is obtained by counting the voxels inside the previously determined vessel lumen boundary. The computation of the cross-sectional diameters is performed as follows: For each in-plane direction, the maximum distance between two boundary points parallel to that direction is determined. The minimum, average, and maximum vessel diameters are then computed as the minimum, average, and maximum values, respectively, of all of these distances. Since the vessel cross-sectional area is closely related to the hemodynamic properties of the vessel at that point, it is more significant than the vessel diameter measures from a physiologic point of view. On the other hand, diameter measures are widely used since these are the only values that are available from traditional methods such as digital subtraction angiography. For this reason, we also express the cross-sectional area measurement is expressed as a "circular equivalent diameter" by calculating the diameter of a circle with the same area as the vessel cross section at a given point. This quantity carries the same information as the cross-sectional area while still representing a diameter and thus being more similar to the traditional diameter measurements. Software Design and User Interface The vessel analysis methods described hereinabove have been implemented and made available for clinical evaluation in the course of the ongoing Virtual Institute for Computer Assistance in Clinical Radiology (VICORA) project (20,21). VICORA was started in November 2000 as a joint project between six major German radiologic centers, two industrial partners (Siemens Medical Solutions, Erlangen, Germany, and MeVis Technology, Bremen, Germany), and the MeVis research institute. VICORA focuses on the development and clinical evaluation of algorithms and prototypical applications in the area of vessel analysis, dynamic MR imaging, and preoperative planning. The VascuVision prototype is based on the medical image processing and prototyping platform ILAB 4, a Windows (Microsoft, Redmond, Wash)-based software application for the development of image processing and visualization algorithms, networks, and application prototypes, with special support for medical imaging applications and integration into clinical environments. The VascuVision software encapsulates the individual algorithmic parts of the vessel analysis as a collection of five processing steps, which are performed one after the other during the examination of a case. The five processing steps are (a) data import, (b) region of interest (ROI) selection, (c) segmentation, (d) premasking, and (e) analysis. To support the individual user's work flow and the requirements imposed by the case, the sequence in which these steps are processed is not strictly fixed. For example, it is not necessary to select an ROI— if this step is omitted an internal ROI is generated automatically from the bounding box of the segmentation step. The premasking step may be applied if suggested by the result of a first segmentation step, but the user may also choose to perform the premasking step before starting the first segmentation, if the necessity for premasking can be anticipated. The design of the user interface of the application has been carried out and evaluated in close cooperation with clinical partners to ensure a high degree of usability. During the development process, we have made extensive use of the rapid prototyping capabilities of the ILAB 4 platform, which allow one to add, remove, or modify user interface elements and corresponding functionality very easily. The main application window is subdivided into the image display area, a general control panel containing user interface elements mainly for image display and rendering options, and a control panel containing elements specific to the currently active processing step. The image display area is further subdivided into several viewing windows, including a 3D viewer and three orthogonal 2D viewers showing the original image data in axial, sagittal, and coronal sections. During the analysis step, these orthogonal views are replaced by a cross- sectional MPR view at the current path location, a stretched MPR view, and a diagram showing the results of the automatic cross-sectional area or diameter measurement along the analysis path. The results of intermediate algorithmic steps, such as primary segmentation, skeletonization, or vessel lumen boundary identification, can be displayed as overlays to the original image data in two or three dimensions, thus providing important feedback on their reliability and accuracy. To verify cross-sectional diameter and path length measurements, a manual measurement tool has been implemented, which can be used in the original orthogonal views as well as the cross-sectional MPR views. The presentation of the cross-sectional and stretched vessel MPR views deserves special attention because these views differ from the familiar radiologic views parallel to the patient's coordinate axes. In VascuVision, the perception of these MPR views is supported by visual indicators in both the 2D and 3D views, providing feedback on the location and spatial orientation of the corresponding image planes. The clinical application of VascuVision, is focused on CT angiography studies of the carotid, coronary, and renal arteries, abdominal and thoracic aorta, and peripheral arteries, as well as MR angiographic studies of the intracranial arteries. In most of the cases examined so far, segmentation and analysis of the vessels of interest by using the VascuVision software have been feasible. In some cases, difficulties have been imposed by large aneurysms with an extremely inhomogeneous distribution of contrast agent or aortic dissections (eg, in thoracic aorta studies) or by completely occluded arteries (eg, in the lower extremities). Visual presentation of the vessels of interest and surrounding structures by using the combination of 2D and 3D renderings has turned out to be useful not only for diagnostic purposes but also for demonstration of the diagnosis to a surgeon, for example. For a more in-depth evaluation, ongoing clinical studies have been started as part of the VICORA project. These studies focus on the processing time needed to analyze a case by using the VascuVision application and on the measurement accuracy, which is evaluated with phantom studies and by comparing the automatic cross-sectional measurements with results of digital subtraction angiography. In parallel, extensions of the segmentation algorithms are being developed to improve results in cases of strongly varying contrast between the vessel and its surroundings and to simplify interactions in the application of the premasking step. The VascuVision application prototype described in this article represents an efficient software tool for the analysis and visualization of vessels on the basis of CT angiographic or MR angiographic data. The option of successively improving the vessel segmentation by additional user interactions extends the applicability of the software to difficult cases, and visual indicators of the reliability of the automatically derived results contribute to the diagnostic value of this tool. The object of the present invention consists in specifying a method for segmenting anatomical structures, in particular of the coronary vessel tree, from 3D image data which permits a more reliable segmentation of the structures. The object is achieved with the aid of the method according to the features of the invention. Advantageous refinements of the method can be gathered from the following description and the exemplary embodiments. For the purpose of segmenting anatomical structures from 3D image data, in particular from CTA image data, in the case of the present method, the first step is to set a starting point, preferably in a region starting from which the structure to be segmented extends, and identify at least one known anatomically- significant point and/or at least one known anatomically- significant surface in the 3D image data. Subsequently, proceeding from the starting point the structure is segmented pixel by pixel with the aid of a multiplicity of segmentation steps in such a way that an instantaneous distance is determined automatically relative to the anatomically significant point and/or to the anatomically significant surface in each segmentation step, and segmentation parameters and/or a selection of adjacent pixels for continuing the segmentation are/is established as a function of the distance by taking account of a known model topology. In the case of the present method, therefore, use is additionally made for the segmentation of known topological information, also denoted in the present patent application as model topology, which enables a more reliable segmentation of the structure. By using the respective knowledge of the instantaneous segmentation position, relative to the previously identified significant points or surfaces, and the knowledge of the fundamental topology in the region of the structure, it is possible firstly to exclude erroneous segmentation in regions where no parts of the structure can be present any more because of the topological knowledge. On the other hand, by automatically changing the segmentation parameters as a function of the segmentation position, the segmentation can also still locate vessels in regions where vessels must still be present because of the topological knowledge, a normal segmentation would, however, be truncated because of a local undershooting or overshooting of the threshold values set. The segmentation parameters that are established for each segmentation step as the structure is traversed comprise in the field of the CTA, for example, HU threshold values for the voxels corresponding to the pixels or threshold values for density gradients between adjacent voxels. The solid angle at which the further segmentation is performed is restricted by establishing a selection of adjacent pixels for a continuation of segmentation as a function of the instantaneous distance. The segmentation can be carried out in this way in a fashion similar to the known region growing although here it is not normally all the pixels adjacent in the volume that are subjected to an analysis, but only pixels in the respectively established solid angle. For this reason, the starting point is preferably set in a region starting from which the structure to be segmented extends such that a specific segmentation direction is already thereby prescribed. The instantaneous distance from the previously identified significant points or surfaces is preferably determined with the aid of at least one distance array that is set up before the start of segmentation and proceeds from the respectively identified point or the respectively identified surface. In the case of a number of identified points and/or surfaces, it is also possible to calculate a number of distance arrays that then respectively specify the distance from the point on which they are based or of the surface on which they are based. These one or more distance arrays are assigned to the pixels or voxels in the 3D image data set such that the distance from the respective anatomically significant point or the respective anatomically significant surface is immediately known for each individual voxel during the segmentation. This distance is then used in establishing the instantaneous segmentation parameters by taking account of the known model topology. When carrying out the method, taking account of the known model topology in this way in order to determine the instantaneous segmentation parameters can be performed by recourse to a provided table in which for a known model topology respectively prescribed segmentation parameters are assigned to the different distances relative to the anatomically significant point and/or the anatomically significant surface, and/or one or more prescribed segmentation directions are assigned in order to continue of the segmentation. Thus, depending on distance from the anatomically significant point, it is possible to establish the solid angle at which the segmentation is continued. Furthermore, depending on this distance it is also possible to vary the range of the HU values within which the voxels belonging to the structure must lie. The same holds for any other segmentation parameters that can be used to segment the structure. In a development of the present method, when the segmentation is truncated because of a lack of adjacent voxels that satisfy the prescribed conditions a search is automatically carried out for continuing the structure in the closer surroundings if a truncation of the structure because of the topological information is unlikely given the instantaneous distance from the one or more anatomically significant points and/or surfaces. If continuation voxels are found that match the truncated structure with regard to the further course, the gap lying therebetween can be closed by interpolation, and the segmentation can be continued with the aid of the continuation voxels. This search for continuation structures in the event of a truncation of the segmentation in a branch of the structure can be performed either already directly after the respective truncation or else not until after the segmentation has come to a standstill in all branches of the structure. In the latter case, such truncation points are firstly stored in order then to be able to search for a continuation of the structure in the end phase of the method at these points. Even though the main field of application of the present method is the segmentation of anatomical vessel structures in 3D image data in CTA imaging, the method can also be used for segmenting other anatomical structures, including in 3D image data of other imaging techniques such as, for example, MR, PET, SPECT, or 3D ultrasound as long as suitable segmentation parameters are available for segmentation. The present method is explained below once more in more detail with the ai^d of an exemplary embodiment in conjunction with the /drawings, diff which: figure 1 shows an example of a first step in carrying out the present method; figure 2 shows an example of a second step in carrying out the present method; figure 3 shows an example of a third step in carrying out the present method; figure 4 shows an example of a fourth step in carrying out the present method; and figure 5 shows an example of a fifth step in carrying out the present method. There is a description in the present example of various steps for segmenting the coronary vessel tree from 3D image data that have been recorded using a CTA technique. Here, the figures illustrate different views of the heart and of the heart chambers and vessels contained therein which are merely indicated schematically because the images themselves lack the capacity for illustration. Figure 1 shows in this case a view of the entire heart 1 in which a section through the aorta 2 is to be recognized in the top area. In the first step, a starting point is set interactively on the display screen by clicking on the aorta 2, for example using a mouse pointer 3. Proceeding from this starting point, the two branch points 4 at which the aorta 2 branches into the coronary vessel structure 5 are detected automatically. These branch points 4 are marked. As is to be seen from the part figures 2A, 2B and 2C, in the next step the heart chambers 11, that is to say the left-hand and right-hand atrium and also the ventricles, are segmented and anatomical landmarks, in the present example the apex 6 of the left-hand ventricle 8, are detected and marked. Different stages of this segmentation are to be seen in this case in the part figures 2A, 2B. Figure 2C shows the segmented left-hand ventricle 8 on whose surface the apex 6 is marked as anatomical landmark. This marking is performed automatically by an image processing algorithm that recognizes the apex of the left-hand ventricle. A distance array around the heart is calculated in the next step. In this case, the surface points of the segmented heart chambers 11, which are enriched with the aid of contrast agent, are used as starting points. Figure 3A shows, to this end, an illustration in which the myocardium 7 is to be recognized on the outside and the left-hand ventricle 8 is to be recognized on the inside. Figure 3B again shows another view of the left- hand ventricle 8. The vessel structure 5 is indicated in these illustrations. The actual segmentation of the vessel structure 5 is performed after these preparatory steps. The segmentation begins at the branch points 4 that are detected in the first step and which are also indicated in the illustration of figure 4. The segmentation is carried out in accordance with the present method with the aid of an adaptive, topological segmentation technique. Figure 4 shows in this regard a part of the aorta 2, together with the coronary vessel structure 5 branching away therefrom. The heart 1 is indicated by the dashed frame. Furthermore, the position of the left-hand ventricle 8 and of two anatomical landmarks, the apex 6 and the geometric centroid 9 of the heart, are also marked in this illustration. In the case of the present method, the segmentation is performed in an ordered way in which the complex structure 5 is analyzed step by step as a function of the instantaneous position relative to the previously marked landmarks 6, 9. The current shape, size and position of the vessel section inside the heart is detected in this way at any time by the segmentation algorithm, since the distance from the surface of the heart chambers, in particular from the apex 6 or from the centroid 9 of the heart is known at any time from the distance array. The threshold value for the segmentation, for example, can be adapted in each case by means of this knowledge to the current segmentation position. More strongly distally positioned vessels frequently require a lower threshold value in order to distinguish them from the surrounding structures. The distance from the apex 6, which is known for every segmentation step, is used in the event of a truncation of the segmentation in a vessel branch to decide whether a search will be made in the closer surroundings for a continuation of the structure. If the distance from the apex 6 is greater than a prescribable value for which vessels normally lose their good contrast with the surroundings, this truncation position is then firstly stored for a later continuation step. The segmentation itself can, for example, be performed by using a distance transformation algorithm in which spheres are enlarged in the respective vessel section until they cut the vessel. Segmentation is performed in this case by taking account of the topological information relating to the structure of a heart, that is to say by taking account of a topological heart model not in all spatial directions as for the conventional region growing technique, but in a directed way. This segmentation with the aid of spheres that are enlarging is indicated by the rings in figure 4. The respective information relating to position or distance is determined here in a fashion related to an anatomical landmark or else to a number of anatomical landmarks, as in the present example. The segmentation device can be prescribed by this knowledge of the position such that erroneous segmentation is avoided in regions where vessels can no longer be present because of the topology. In the present example, after the vessel structure 5 has been traversed in accordance with the steps previously explained, a search is carried out at the previously stored truncation positions for continued structures at which the segmentation was truncated for lack of adjacent pixels satisfying the segmentation condition. These positions are stored when a truncation of the vessel structure is unlikely because of the topological information at this point. An automatic search in the closer surroundings is then carried out at these points for voxels that could constitute a continuation of the vessel structure. Use is made in this case of the fact that the coronary tree can be represented by a vector set having an associated item of diameter information, that is to say by a set of juxtaposed cylinders. The tangential vector is calculated along this ordered vector set. This calculation can be used to estimate the region in which the continuation of the vessel would have to lie. The image processing algorithm attempts to identify possible vessel regions in this calculated search region, for example by analyzing the local Hessian matrix, or by calculating the eigenvectors of voxel clusters that lie inside the valid HU range for the vessel structure. If such voxels are found, they are connected to the already segmented structures by interpolation. The tangential vectors 12 are indicated in figure 5, as is also the search region 13 prescribed by these vectors. A gap 10 in the illustration of the vessel structure 5 that is filled up by this last step is also respectively to be seen in this case in the figure. Structurally specific or topological information is taken into account with the aid of the present method when segmenting the structure. Given the prescription of a smoothness condition with regard to the course of the vessels, it is possible to achieve the situation where only plausible voxels are calculated in relation to the vessel structure in each segmentation step. Taking account of the distance from previously identified anatomical landmarks, in particular from the surface of the heart chambers, ensures that only vessels outside the already segmented heart chambers are identified. This prevents the segmented structure from straying into other surrounding structures such as, for example, the heart chambers. The proposed method can be used, for example, to segment the coronary vessel tree much more accurately than is possible with the aid of the previously known methods of the prior art. Not only the vessel tree itself, but the entire anatomy of the heart are analyzed in the segmentation. The additional search step for filling up interrupted vessel structures renders it possible to detect and segment even the smallest vessel structures. WE CLAIM 1. A method for segmenting anatomical structures from 3D image data, the method comprising: setting a starting point in the 3D image data; identifying at least one of at least one known anatomically significant point and at least one known anatomically significant surface in the 3D image data; segmenting, proceeding from the starting point, the anatomical structure pixel by pixel with a multiplicity of segmentation steps such that an instantaneous distance is determined automatically relative to at ieast one of the at least one known anatomically significant point and to the at least one known anatomically significant surface in each segmentation step; establishing at least one of segmentation parameters and a selection of adjacent pixels for continuing the segmentation as a function of the determined distance, taking account of a model topology; and displaying a result of the segmentation process, wherein when a truncation point during the segmentation process is reached at a position at which the anatomical structure should not be truncated, based on the model topology, a search algorithm searches for pixels and continues the segmentation of the anatomical structure in an image region determined by an extrapolation of the already segmented anatomical structure, and wherein a gap in the segmented anatomical structure is filled up by means of interpolation. 2. The method as claimed in claim 1, wherein before the start of segmenting the anatomical structure, at least one distance array is set up proceeding from the at least one known anatomically significant point, or the at least one anatomically significant surface and is assigned to the pixels, at least one of the instantaneous distance relative to the at least one known anatomically significant point and the at least one known anatomically significant surface being determined directly from the distance array for each pixel during the segmentation process. 3. The method as claimed in claim 1, wherein the segmentation parameters are established for each segmentation step by recourse to a table in which respectively prescribed segmentation parameters are assigned to the different distances relative to at least one of the at least one known anatomically significant point and the at least one known anatomically significant surface. 4. The method as claimed in claim 1, wherein the selection of adjacent pixels for continuing the segmentation process is established for each segmentation step by recourse to a table in which in each case at least one prescribed segmentation directions are assigned to the different distances relative to at least one of the at least one known anatomically significant point and the at least one known anatomically significant surface in order to continue the segmentation process. 5. The method as claimed in claim 1, wherein tangential vectors of the already segmented structure are calculated in the region of the truncation point for the extrapolation. 6. The method as claimed in claim 5, wherein in order to detect pixels for continuing the structure a smoothness condition must be satisfied between the tangential vectors of the already segmented structure and tangential vectors of the continued structure in the region of the truncation point. 7. The method as claimed in claim 1, wherein when segmenting a coronary vessel tree from 3D image data of the heart the starting point is set in the aorta, and at least one of the apex and the geometrical centroid of the heart is identified in the 3D data as the at least one known anatomically significant point. 8. The method as claimed in claim 1, wherein when segmenting a coronary vessel tree from 3D image data of the heart the starting point is set in the aorta and a surface of the heart chambers that has been segmented in advance is identified in the 3D image data as the at least one known anatomically significant surface.

Full Text

Description

Method for segmenting anatomical structures from 3D image data
by using topological information
The present invention relates to a method for segmenting
anatomical structures, in particular of the coronary vessel
tree, from 3D image data such as are generated, for example, in
CT angiography (CTA).
The present method is applied chiefly in the field of computer
tomography when recording vessel structures. A great advantage
of CT angiography by comparison with other imaging techniques
such as magnetic resonance (MR) tomography, PET (Positron
Emission Tomography), SPECT (Single Photon Emission Computed
Tomography) or the 3D ultrasound technique consists in that,
for example, the entire vessel tree of the heart can be
recorded with a single CT scan by adding a contrast agent. The
3D image data thereby obtained can be visualized using
different techniques.
For a quantitative evaluation, in particular a measurement of
stenoses or of plaque deposits, it is necessary for the
corresponding regions of the vessel structure to be segmented
from the 3D image data. This segmentation is performed in a
follow-up process on an image computer. The currently most
frequently used and commercially available technique of
segmentation is the technique of so-called region growing. In
this technique, all the respectively adjacent pixels (voxels)
are analyzed starting from seed points, which can be prescribed
by the user, in the 3D image data, and identified as part of the
vessel structure upon fulfillment of specific conditions. As one
condition for the membership of the vessel structure, it is
possible, for example, to check whether the voxel falls into a
prescribed HU range (HU: Hounsfield Units). It is also possible
to prescribe for the density gradients between adjacent voxels a
highest value above which the adjacent voxel is no longer

regarded as part of the vessel structure. The voxels respectively newly identified
as part of the vessel structure are used, in turn, as starting points for the next
step of analysis or segmentation. In this way, the already identified structure
grows three-dimensionally until the complete, prescribable region of the vessel
structure is segmented. An example of the use of such a technique for
segmenting vessel structures can be taken from the publication by T. Boskamp
et al. 'New Vessel Analysis Tool for Morphometric Quantification and Visualization
of Vessels in CT and MR Imaging Data Sets', Radiographics 2004, 24, 287 - 297.
The known region growing technique operates in may instances satisfactorily,
but does not reach all the vessels that are visible to a viewer in a display of the
3D image recording. Whereas, given a suitable 3D visualization technique, the
human eye can also still detect the smallest vessels as part of the structure, the
segmentation algorithm detects only specific homogeneous, coherent parts in the
volume examined. Furthermore, it is also possible for the segmented structure to
be washed out in adjacent image areas when they have similar HU values and lie
very near to the vessel structures.
The cited publication published in Radiographic journal, 2004, 24 (page 287-
297), and entitled 'New vessel analysis tool for Morphometric Qualification and
visualization of vessels in CT and MR image data sets', by T. BOSKAMP et al,
describes an image processing algorithms and a prototypical research software
tool for visualization and quantitative analysis of vessels in data sets from
computed tomography and magnetic resonance imaging.

In diagnostic angiographic imaging, the traditional projection method of digital
subtraction angiography has recently been more and more superseded by the
tomographic acquisition methods computed tomographic (CT) angiography and
magnetic resonance (MR) angiography. Owing to the three-dimensional (3D)
nature of CT angiography and MR angiography data, these methods allow
sophisticated visualization of vascular systems, such as volume rendering or
maximum intensity projection from any angle, providing information of high
relevance for diagnostic as well as demonstration purposes. In addition, CT
angiography and MR angiography studies are able to provide quantitative
morphometric information, such as vessel cross-sectional areas, volume
measures of certain vessel segments, and curvature and tortuosity measures,
with high diagnostic value for a number of vascular diseases.
To obtain such information from the original CT angiography and MR
angiography sections, it is necessary to perform postprocessing steps to produce
a description of the vessel path and multiplanar reformation (MPR) images
orthogonal to this path. Computer systems providing this kind of functionality
should require only a minimum of user interaction while still being able to
successfully process images of small or strongly curved vessels or vessels with
low or varying contrast to surrounding structures. In addition, it is desirable to
obtain an automatic measurement of cross-sectional quantities along the vessel
path, together with some indication of how reliable and accurate these results
are. Other important features are length measurements along the vessel path
and volumetry of user-defined vessel segments.

Although a number of system packages providing this kind of functionality are
available from manufacturers of radiologic workstations, none of these systems
fully satisfy all of the listed requirements (at least from the authors' point of
view). Above all, a high accuracy in the determination of the vessel centerlines is
a prerequisite for precise cross-sectional rendering and measurements. An
inaccurate vessel centerline may result in cross-sectional sections that are tilted
against the true vessel path direction and thus in artificially enlarged cross-
sectional areas and vessel diameters. In difficult cases where the initially
computed centerline is not satisfactory, the ability to improve the result by few
additional user interactions is needed.
The cited reference provides a computer application prototype addressing these
issues. The requirements for the cited invention are (a) high accuracy in the
determination of the centerlines of the vessels of interest, (b) automatic
measurement of cross-sectional areas and diameters along the vessel path, (c)
feedback on the reliability of these results, and (d) a user interface supporting a
smooth work flow and integration into the clinical routine. In this article, we
describe the algorithmic components used in this development and the
prototypical research application VascuVision that makes use of these algorithms
and present some examples of clinical applications.
The known methods for the determination of vessel centerlines can be roughly
grouped into two categories: (a) methods that compute an optimal path
connecting two given points by minimizing some cost functional, taking into

account external, image-related factors (eg, gray value, local contrast) and
internal, path-related factors (length, curvature) (1-4); and (b) methods that
perform vessel segmentation first followed by a subsequent skeletonization step
that yields the vessel centerlines (5-41). The first type of methods (a term
functional approach) offers some advantages in that it avoids the time-
consuming and possibly error-prone segmentation step and is typically faster and
requires less user interaction. On the other hand, the consideration of path
length and curvature may result in a minimizing path that deviates from the true
centerline, especially at bifurcations or at locations where the vessel is strongly
bent. In addition, the internal algorithmic components used by these methods
are not easily perceived by the user, and it is thus difficult to extend such a
method to integrate additional user hints in order to improve a dissatisfactory
initial vessel path.
The second type of method, called a geometric approach, makes use of an
intermediate processing result, the vessel segmentation, which can be verified by
the user and for which it is easy to integrate interactive extensions that allow
improvement of the segmentation in case the automatic algorithm fails. On the
basis of the segmentation result, an accurate skeletonization can usually be
computed without any interaction, but again it is possible to incorporate user
hints, such as explicitly specified skeleton endpoints.
The vessel analysis method proposed in the cited reference belongs to the
second category and consists of the following algorithmic steps: (a) vessel

segmentation based on a region growing algorithm, (b) interactive "premasking"
to optionally exclude interfering structures close to the vessels of interest, (c)
distance transform-based skeletonization, (d) MPR orthogonal to the vessel path,
(e) lumen boundary identification on the orthogonal cross-section images, and
(f) morphometric measurements.
The first step, region growing-based vessel segmentation, is started by manually
placing one or more seed points in the vessel(s) of interest. From these seed
points, more and more neighboring voxels are included in the segmentation,
provided that certain inclusion criteria are met. The most important criterion is
the lower gray value threshold: Neighboring voxels are included only if their gray
values do not fall below this threshold. A second criterion provides an adaptive
upper threshold, thus avoiding the inclusion of brighter structures (eg, bones in
CT images). The upper threshold is combined with a size limit in order to still
include small, bright spots, such as small calcifications or small regions with a
high concentration of contrast agent. (The location of these small, bright spots is
later used in the refined lumen boundary determination step.) A third criterion is
the gray level difference of neighboring voxels (gradient threshold).
For the successful application of this, it is crucial to select an appropriate lower
threshold. Since it is not possible to specify a good threshold value in advance
(neither for the algorithm nor for the user), an implementation of the region
growing algorithm performs the segmentation for a whole range of thresholds in
a single pass by successively decreasing the current global threshold and

restarting the growing process, with the present segmentation used as seed
voxels. The result of this multithreshold segmentation is presented to the user as
a threshold-volume diagram, which indicates the size of the segmentation result
corresponding to each threshold in a given interval. The threshold-volume curve
shows how the volume of the segmented object in milliliters (vertical axis)
increases as the threshold value in Hounsfield units or gray values (MR imaging)
(horizontal axis) is reduced. The inclusion of neighboring structures due to
leakage of the region growing algorithm is expressed as a sudden step in the
threshold curve. Consequently, possible candidates for a reasonable threshold
value are easily identified as locations of larger steps in the threshold curve. This
diagram, combined with a fast, interactive 3D display of the segmentation result
for any selected threshold, thus allows one to find a good threshold with only a
few user interactions. In many cases, the first automatic guess provided by the
algorithm on the basis of an analysis of the threshold curve already yields an
acceptable threshold value.
In some cases, a global lower threshold is not sufficient for appropriate
differentiation between the vessels of interest and neighboring structures, such
as bones or closely passing veins. In these cases, an optional premasking step
can be used to interactively exclude such structures. The premasking is based on
the watershed transform, which subdivides the image volume into several bright
regions separated by darker areas. Once this image transformation has been
computed, include and exclude markers can be used to interactively define
regions containing the vessels of interest and regions containing adjacent
structures.

In the implementation, by the cited reference, the watershed transform is
applied in three dimensions to the original image, yielding boundary surfaces
that lie somewhere in between the bright image regions. Consequently, the
result of this step is not segmentation of a vessel but definition of a mask
containing the vessel and some surrounding dark regions and excluding adjacent
structures of similar brightness. In a subsequent segmentation step that uses the
region growing algorithm described earlier, the segmentation can be restricted to
this mask, thus avoiding the inclusion of neighboring structures. This use of the
watershed transform is in contrast to the gradient watershed segmentation
methods, where the transformation is applied to the gradient image, yielding a
separation between bright and dark regions.
Once the vessels of interest have been segmented, the vessel skeleton is
computed by successive erosion of border voxels until a set of connected voxels
with a thickness of 1 voxel is obtained. The algorithm used in the cited reference
our method is based on a discretized version of the distance transform, is able to
correctly handle nonisotropic images, and offers special features to suppress
spurious side branches resulting from boundary irregularities. The distance
transform assigns to each voxel of the segmented object the minimal distance
between this voxel and voxels outside the segmented object. By weighting the
discretized distances with the voxel resolution, a correct distance is computed in
the case of anisotropic voxels. For the side branch suppression, the algorithm
computes the gradient of the distance transform, which is close to 1 at points
where the boundary is smooth and close to 0 at endpoints of major skeleton
lines. Thus, the magnitude of the distance transform gradient serves as an

indicator for the significance of skeleton segments, and thresholding of this
gradient can be used to suppress more or less insignificant side branches.
It can be shown that the theoretical accuracy of this skeletonization algorithm is
1 voxel distance; in practice, the accuracy of the centerline determination is
mainly dependent on the accuracy of the segmentation. The high accuracy of the
skeletonization algorithm has been corroborated by phantom studies, in which
the maximum reproduction errors of the true centerlines have been measured for
stenosis, aneurysm, and bent vessel phantoms with different diameters and
orientations. In a total of 85 measurements, the reproduction errors were found
to be in the range of up to 1 voxel distance in 42 cases (49%), up to 2 voxel
distances in 76 cases (89%), and above 2 voxel distances in nine cases (11%).
The vessel centerlines obtained from the skeletonization step can be used as a
starting point for the analysis of arbitrary paths in the vessel tree. In an
interactive 3D display, the user is able to define an analysis path by selecting two
points anywhere on the skeleton. This triggers the computation of the cross-
sectional MPR images orthogonal to the connecting path, which are presented to
the user and which form the basis of the subsequent morphometric analysis. In
addition, a stretched rendered image of the vessel is computed, which appears
as if the vessel had been pulled straight and sliced lengthwise. This rendered
image can be helpful in obtaining an overview of the vessel, identifying
suspicious points, and selecting a current path position corresponding to the
cross-sectional MPR image displayed.

To perform cross-sectional measurements, it is necessary to identify the vessel
lumen boundary on the cross-sectional MPR sections. Although this could be
done by simply projecting the result of the primary vessel segmentation onto the
cross-sectional plane, it is often advantageous to perform a refined and more
accurate vessel segmentation on a section-by-section basis. This refinement
allows an adaptation to varying contrast between the vessel and its surroundings
and improved differentiation between the vessel lumen and calcifications at the
vessel wall that may have been included in the primary segmentation.
The cited reference uses two-dimensional (2D) gradient watershed segmentation
for the lumen boundary determination. (Given a symmetric point spread function
of the imaging modality, this segmentation method is equivalent to the full width
half maximum criterion applied to the gray value profile.) This method differs
from the watershed transform used in the premasking step in that it operates in
two dimensions and is applied to the gradient image of the cross-sectional MPR
sections, in which the lumen boundary appears as a bright line separating the
darker interior and exterior regions of the vessel. Include and exclude masks
defining the interior and exterior of the vessel within a single cross-sectional
plane are derived automatically from the primary vessel segmentation. In
addition, calcification spots are identified based on (a) a high relative gray value
in the interior of the spot, (b) a high gradient value, and (c) the location of small
bright spots detected during the primary region growing segmentation step.

The final step, the morphometric analysis of the vessel along the selected
analysis path, is based on the quantification of the cross-sectional areas and
diameters for each cross-sectional section. The cross-sectional area of a single
section is obtained by counting the voxels inside the previously determined
vessel lumen boundary. The computation of the cross-sectional diameters is
performed as follows: For each in-plane direction, the maximum distance
between two boundary points parallel to that direction is determined. The
minimum, average, and maximum vessel diameters are then computed as the
minimum, average, and maximum values, respectively, of all of these distances.
Since the vessel cross-sectional area is closely related to the hemodynamic
properties of the vessel at that point, it is more significant than the vessel
diameter measures from a physiologic point of view. On the other hand,
diameter measures are widely used since these are the only values that are
available from traditional methods such as digital subtraction angiography. For
this reason, we also express the cross-sectional area measurement is expressed
as a "circular equivalent diameter" by calculating the diameter of a circle with the
same area as the vessel cross section at a given point. This quantity carries the
same information as the cross-sectional area while still representing a diameter
and thus being more similar to the traditional diameter measurements.

Software Design and User Interface
The vessel analysis methods described hereinabove have been implemented and
made available for clinical evaluation in the course of the ongoing Virtual
Institute for Computer Assistance in Clinical Radiology (VICORA) project (20,21).
VICORA was started in November 2000 as a joint project between six major
German radiologic centers, two industrial partners (Siemens Medical Solutions,
Erlangen, Germany, and MeVis Technology, Bremen, Germany), and the MeVis
research institute. VICORA focuses on the development and clinical evaluation of
algorithms and prototypical applications in the area of vessel analysis, dynamic
MR imaging, and preoperative planning. The VascuVision prototype is based on
the medical image processing and prototyping platform ILAB 4, a Windows
(Microsoft, Redmond, Wash)-based software application for the development of
image processing and visualization algorithms, networks, and application
prototypes, with special support for medical imaging applications and integration
into clinical environments.
The VascuVision software encapsulates the individual algorithmic parts of the
vessel analysis as a collection of five processing steps, which are performed one
after the other during the examination of a case. The five processing steps are
(a) data import, (b) region of interest (ROI) selection, (c) segmentation, (d)
premasking, and (e) analysis. To support the individual user's work flow and the
requirements imposed by the case, the sequence in which these steps are
processed is not strictly fixed. For example, it is not necessary to select an ROI—

if this step is omitted an internal ROI is generated automatically from the
bounding box of the segmentation step. The premasking step may be applied if
suggested by the result of a first segmentation step, but the user may also
choose to perform the premasking step before starting the first segmentation, if
the necessity for premasking can be anticipated.
The design of the user interface of the application has been carried out and
evaluated in close cooperation with clinical partners to ensure a high degree of
usability. During the development process, we have made extensive use of the
rapid prototyping capabilities of the ILAB 4 platform, which allow one to add,
remove, or modify user interface elements and corresponding functionality very
easily. The main application window is subdivided into the image display area, a
general control panel containing user interface elements mainly for image display
and rendering options, and a control panel containing elements specific to the
currently active processing step. The image display area is further subdivided
into several viewing windows, including a 3D viewer and three orthogonal 2D
viewers showing the original image data in axial, sagittal, and coronal sections.
During the analysis step, these orthogonal views are replaced by a cross-
sectional MPR view at the current path location, a stretched MPR view, and a
diagram showing the results of the automatic cross-sectional area or diameter
measurement along the analysis path. The results of intermediate algorithmic
steps, such as primary segmentation, skeletonization, or vessel lumen boundary
identification, can be displayed as overlays to the original image data in two or
three dimensions, thus providing important feedback on their reliability and
accuracy. To verify cross-sectional diameter and path length

measurements, a manual measurement tool has been implemented, which can
be used in the original orthogonal views as well as the cross-sectional MPR
views.
The presentation of the cross-sectional and stretched vessel MPR views deserves
special attention because these views differ from the familiar radiologic views
parallel to the patient's coordinate axes. In VascuVision, the perception of these
MPR views is supported by visual indicators in both the 2D and 3D views,
providing feedback on the location and spatial orientation of the corresponding
image planes.
The clinical application of VascuVision, is focused on CT angiography studies of
the carotid, coronary, and renal arteries, abdominal and thoracic aorta, and
peripheral arteries, as well as MR angiographic studies of the intracranial
arteries. In most of the cases examined so far, segmentation and analysis of the
vessels of interest by using the VascuVision software have been feasible. In
some cases, difficulties have been imposed by large aneurysms with an
extremely inhomogeneous distribution of contrast agent or aortic dissections (eg,
in thoracic aorta studies) or by completely occluded arteries (eg, in the lower
extremities). Visual presentation of the vessels of interest and surrounding
structures by using the combination of 2D and 3D renderings has turned out to
be useful not only for diagnostic purposes but also for demonstration of the
diagnosis to a surgeon, for example.

For a more in-depth evaluation, ongoing clinical studies have been started as
part of the VICORA project. These studies focus on the processing time needed
to analyze a case by using the VascuVision application and on the measurement
accuracy, which is evaluated with phantom studies and by comparing the
automatic cross-sectional measurements with results of digital subtraction
angiography. In parallel, extensions of the segmentation algorithms are being
developed to improve results in cases of strongly varying contrast between the
vessel and its surroundings and to simplify interactions in the application of the
premasking step.
The VascuVision application prototype described in this article represents an
efficient software tool for the analysis and visualization of vessels on the basis of
CT angiographic or MR angiographic data. The option of successively improving
the vessel segmentation by additional user interactions extends the applicability
of the software to difficult cases, and visual indicators of the reliability of the
automatically derived results contribute to the diagnostic value of this tool.
The object of the present invention consists in specifying a method for
segmenting anatomical structures, in particular of the coronary vessel tree, from
3D image data which permits a more reliable segmentation of the structures.
The object is achieved with the aid of the method according to the features of
the invention. Advantageous refinements of the method can be gathered from
the following description and the exemplary embodiments.

For the purpose of segmenting anatomical structures from 3D image data, in
particular from CTA image data, in the case of the present method, the first step
is to set a starting point, preferably in a region starting from which the structure
to be

segmented extends, and identify at least one known anatomically-
significant point and/or at least one known anatomically-
significant surface in the 3D image data. Subsequently,
proceeding from the starting point the structure is segmented
pixel by pixel with the aid of a multiplicity of segmentation
steps in such a way that an instantaneous distance is
determined automatically relative to the anatomically
significant point and/or to the anatomically significant
surface in each segmentation step, and segmentation parameters
and/or a selection of adjacent pixels for continuing the
segmentation are/is established as a function of the distance
by taking account of a known model topology.
In the case of the present method, therefore, use is
additionally made for the segmentation of known topological
information, also denoted in the present patent application as
model topology, which enables a more reliable segmentation of
the structure. By using the respective knowledge of the
instantaneous segmentation position, relative to the previously
identified significant points or surfaces, and the knowledge of
the fundamental topology in the region of the structure, it is
possible firstly to exclude erroneous segmentation in regions
where no parts of the structure can be present any more because
of the topological knowledge. On the other hand, by
automatically changing the segmentation parameters as a
function of the segmentation position, the segmentation can
also still locate vessels in regions where vessels must still
be present because of the topological knowledge, a normal
segmentation would, however, be truncated because of a local
undershooting or overshooting of the threshold values set.
The segmentation parameters that are established for each
segmentation step as the structure is traversed comprise in the
field of the CTA, for example, HU threshold values for the
voxels corresponding to the pixels or threshold values for
density gradients between adjacent voxels. The solid angle at
which the further segmentation is performed is restricted by

establishing a selection of adjacent pixels for a continuation
of segmentation as a function of the instantaneous distance.
The segmentation can be carried out in this way in a fashion
similar to the known region growing although here it is not
normally all the pixels adjacent in the volume that are
subjected to an analysis, but only pixels in the respectively
established solid angle. For this reason, the starting point is
preferably set in a region starting from which the structure to
be segmented extends such that a specific segmentation
direction is already thereby prescribed.
The instantaneous distance from the previously identified
significant points or surfaces is preferably determined with
the aid of at least one distance array that is set up before
the start of segmentation and proceeds from the respectively
identified point or the respectively identified surface. In the
case of a number of identified points and/or surfaces, it is
also possible to calculate a number of distance arrays that
then respectively specify the distance from the point on which
they are based or of the surface on which they are based. These
one or more distance arrays are assigned to the pixels or
voxels in the 3D image data set such that the distance from the
respective anatomically significant point or the respective
anatomically significant surface is immediately known for each
individual voxel during the segmentation. This distance is then
used in establishing the instantaneous segmentation parameters
by taking account of the known model topology.
When carrying out the method, taking account of the known model
topology in this way in order to determine the instantaneous
segmentation parameters can be performed by recourse to a
provided table in which for a known model topology respectively
prescribed segmentation parameters are assigned to the
different distances relative to the anatomically significant
point and/or the anatomically significant surface, and/or one
or more prescribed segmentation directions are assigned in
order to continue of the segmentation. Thus, depending on

distance from the anatomically significant point, it is
possible to establish the solid angle at which the segmentation
is continued. Furthermore, depending on this distance it is
also possible to vary the range of the HU values within which
the voxels belonging to the structure must lie. The same holds
for any other segmentation parameters that can be used to
segment the structure.
In a development of the present method, when the segmentation
is truncated because of a lack of adjacent voxels that satisfy
the prescribed conditions a search is automatically carried out
for continuing the structure in the closer surroundings if a
truncation of the structure because of the topological
information is unlikely given the instantaneous distance from
the one or more anatomically significant points and/or
surfaces. If continuation voxels are found that match the
truncated structure with regard to the further course, the gap
lying therebetween can be closed by interpolation, and the
segmentation can be continued with the aid of the continuation
voxels. This search for continuation structures in the event of
a truncation of the segmentation in a branch of the structure
can be performed either already directly after the respective
truncation or else not until after the segmentation has come to
a standstill in all branches of the structure. In the latter
case, such truncation points are firstly stored in order then
to be able to search for a continuation of the structure in the
end phase of the method at these points.
Even though the main field of application of the present method
is the segmentation of anatomical vessel structures in 3D image
data in CTA imaging, the method can also be used for segmenting
other anatomical structures, including in 3D image data of
other imaging techniques such as, for example, MR, PET, SPECT,
or 3D ultrasound as long as suitable segmentation parameters
are available for segmentation.

The present method is explained below once more in more detail
with the ai^d of an exemplary embodiment in conjunction with the
/drawings, diff which:
figure 1 shows an example of a first step in carrying out the
present method;
figure 2 shows an example of a second step in carrying out the
present method;
figure 3 shows an example of a third step in carrying out the
present method;
figure 4 shows an example of a fourth step in carrying out the
present method; and
figure 5 shows an example of a fifth step in carrying out the
present method.
There is a description in the present example of various steps
for segmenting the coronary vessel tree from 3D image data that
have been recorded using a CTA technique. Here, the figures
illustrate different views of the heart and of the heart
chambers and vessels contained therein which are merely
indicated schematically because the images themselves lack the
capacity for illustration. Figure 1 shows in this case a view
of the entire heart 1 in which a section through the aorta 2 is
to be recognized in the top area. In the first step, a starting
point is set interactively on the display screen by clicking on
the aorta 2, for example using a mouse pointer 3. Proceeding
from this starting point, the two branch points 4 at which the
aorta 2 branches into the coronary vessel structure 5 are
detected automatically. These branch points 4 are marked.
As is to be seen from the part figures 2A, 2B and 2C, in the
next step the heart chambers 11, that is to say the left-hand
and right-hand atrium and also the ventricles, are segmented

and anatomical landmarks, in the present example the apex 6 of
the left-hand ventricle 8, are detected and marked. Different
stages of this segmentation are to be seen in this case in the
part figures 2A, 2B. Figure 2C shows the segmented left-hand
ventricle 8 on whose surface the apex 6 is marked as anatomical
landmark. This marking is performed automatically by an image
processing algorithm that recognizes the apex of the left-hand
ventricle.
A distance array around the heart is calculated in the next
step. In this case, the surface points of the segmented heart
chambers 11, which are enriched with the aid of contrast agent,
are used as starting points. Figure 3A shows, to this end, an
illustration in which the myocardium 7 is to be recognized on
the outside and the left-hand ventricle 8 is to be recognized
on the inside. Figure 3B again shows another view of the left-
hand ventricle 8. The vessel structure 5 is indicated in these
illustrations.
The actual segmentation of the vessel structure 5 is performed
after these preparatory steps. The segmentation begins at the
branch points 4 that are detected in the first step and which
are also indicated in the illustration of figure 4. The
segmentation is carried out in accordance with the present
method with the aid of an adaptive, topological segmentation
technique. Figure 4 shows in this regard a part of the aorta 2,
together with the coronary vessel structure 5 branching away
therefrom. The heart 1 is indicated by the dashed frame.
Furthermore, the position of the left-hand ventricle 8 and of
two anatomical landmarks, the apex 6 and the geometric centroid
9 of the heart, are also marked in this illustration.
In the case of the present method, the segmentation is
performed in an ordered way in which the complex structure 5 is
analyzed step by step as a function of the instantaneous
position relative to the previously marked landmarks 6, 9. The
current shape, size and position of the vessel section inside

the heart is detected in this way at any time by the
segmentation algorithm, since the distance from the surface of
the heart chambers, in particular from the apex 6 or from the
centroid 9 of the heart is known at any time from the distance
array. The threshold value for the segmentation, for example,
can be adapted in each case by means of this knowledge to the
current segmentation position. More strongly distally
positioned vessels frequently require a lower threshold value
in order to distinguish them from the surrounding structures.
The distance from the apex 6, which is known for every
segmentation step, is used in the event of a truncation of the
segmentation in a vessel branch to decide whether a search will
be made in the closer surroundings for a continuation of the
structure. If the distance from the apex 6 is greater than a
prescribable value for which vessels normally lose their good
contrast with the surroundings, this truncation position is
then firstly stored for a later continuation step.
The segmentation itself can, for example, be performed by using
a distance transformation algorithm in which spheres are
enlarged in the respective vessel section until they cut the
vessel. Segmentation is performed in this case by taking
account of the topological information relating to the
structure of a heart, that is to say by taking account of a
topological heart model not in all spatial directions as for
the conventional region growing technique, but in a directed
way. This segmentation with the aid of spheres that are
enlarging is indicated by the rings in figure 4. The respective
information relating to position or distance is determined here
in a fashion related to an anatomical landmark or else to a
number of anatomical landmarks, as in the present example. The
segmentation device can be prescribed by this knowledge of the
position such that erroneous segmentation is avoided in regions
where vessels can no longer be present because of the topology.
In the present example, after the vessel structure 5 has been
traversed in accordance with the steps previously explained, a

search is carried out at the previously stored truncation
positions for continued structures at which the segmentation
was truncated for lack of adjacent pixels satisfying the
segmentation condition. These positions are stored when a
truncation of the vessel structure is unlikely because of the
topological information at this point. An automatic search in
the closer surroundings is then carried out at these points for
voxels that could constitute a continuation of the vessel
structure. Use is made in this case of the fact that the
coronary tree can be represented by a vector set having an
associated item of diameter information, that is to say by a
set of juxtaposed cylinders. The tangential vector is
calculated along this ordered vector set. This calculation can
be used to estimate the region in which the continuation of the
vessel would have to lie. The image processing algorithm
attempts to identify possible vessel regions in this calculated
search region, for example by analyzing the local Hessian
matrix, or by calculating the eigenvectors of voxel clusters
that lie inside the valid HU range for the vessel structure. If
such voxels are found, they are connected to the already
segmented structures by interpolation. The tangential vectors
12 are indicated in figure 5, as is also the search region 13
prescribed by these vectors. A gap 10 in the illustration of
the vessel structure 5 that is filled up by this last step is
also respectively to be seen in this case in the figure.
Structurally specific or topological information is taken into
account with the aid of the present method when segmenting the
structure. Given the prescription of a smoothness condition
with regard to the course of the vessels, it is possible to
achieve the situation where only plausible voxels are
calculated in relation to the vessel structure in each
segmentation step. Taking account of the distance from
previously identified anatomical landmarks, in particular from
the surface of the heart chambers, ensures that only vessels
outside the already segmented heart chambers are identified.
This prevents the segmented structure from straying into other

surrounding structures such as, for example, the heart
chambers.
The proposed method can be used, for example, to segment the
coronary vessel tree much more accurately than is possible with
the aid of the previously known methods of the prior art. Not
only the vessel tree itself, but the entire anatomy of the
heart are analyzed in the segmentation. The additional search
step for filling up interrupted vessel structures renders it
possible to detect and segment even the smallest vessel
structures.

WE CLAIM
1. A method for segmenting anatomical structures from 3D image data, the
method comprising:
setting a starting point in the 3D image data;
identifying at least one of at least one known anatomically significant
point and at least one known anatomically significant surface in the 3D
image data;
segmenting, proceeding from the starting point, the anatomical structure
pixel by pixel with a multiplicity of segmentation steps such that an
instantaneous distance is determined automatically relative to at ieast one
of the at least one known anatomically significant point and to the at least
one known anatomically significant surface in each segmentation step;
establishing at least one of segmentation parameters and a selection of
adjacent pixels for continuing the segmentation as a function of the
determined distance, taking account of a model topology; and
displaying a result of the segmentation process,
wherein when a truncation point during the segmentation process
is reached at a position at which the anatomical structure should
not be truncated, based on the model topology, a search algorithm

searches for pixels and continues the segmentation of the
anatomical structure in an image region determined by an
extrapolation of the already segmented anatomical structure, and
wherein a gap in the segmented anatomical structure is filled up by
means of interpolation.
2. The method as claimed in claim 1, wherein before the start of segmenting
the anatomical structure, at least one distance array is set up proceeding
from the at least one known anatomically significant point, or the at least
one anatomically significant surface and is assigned to the pixels, at least
one of the instantaneous distance relative to the at least one known
anatomically significant point and the at least one known anatomically
significant surface being determined directly from the distance array for
each pixel during the segmentation process.
3. The method as claimed in claim 1, wherein the segmentation parameters
are established for each segmentation step by recourse to a table in which
respectively prescribed segmentation parameters are assigned to the
different distances relative to at least one of the at least one known
anatomically significant point and the at least one known anatomically
significant surface.
4. The method as claimed in claim 1, wherein the selection of adjacent pixels

for continuing the segmentation process is established for each
segmentation step by recourse to a table in which in each case at least
one prescribed segmentation directions are assigned to the different
distances relative to at least one of the at least one known anatomically
significant point and the at least one known anatomically significant
surface in order to continue the segmentation process.
5. The method as claimed in claim 1, wherein tangential vectors of the
already segmented structure are calculated in the region of the truncation
point for the extrapolation.
6. The method as claimed in claim 5, wherein in order to detect pixels for
continuing the structure a smoothness condition must be satisfied
between the tangential vectors of the already segmented structure and
tangential vectors of the continued structure in the region of the
truncation point.
7. The method as claimed in claim 1, wherein when segmenting a coronary
vessel tree from 3D image data of the heart the starting point is set in the
aorta, and at least one of the apex and the geometrical centroid of the
heart is identified in the 3D data as the at least one known anatomically
significant point.
8. The method as claimed in claim 1, wherein when segmenting a coronary

vessel tree from 3D image data of the heart the starting point is set in the
aorta and a surface of the heart chambers that has been segmented in
advance is identified in the 3D image data as the at least one known
anatomically significant surface.

Documents:

00814-kol-2005-abstract.pdf

00814-kol-2005-claims.pdf

00814-kol-2005-description complete.pdf

00814-kol-2005-drawings.pdf

00814-kol-2005-form-1.pdf

00814-kol-2005-form-2.pdf

00814-kol-2005-form-3.pdf

00814-kol-2005-form-5.pdf

814-KOL-2005-(04-11-2011)-FORM 27.pdf

814-KOL-2005-ABSTRACT 1.1.pdf

814-KOL-2005-CANCELLED PAGES.pdf

814-KOL-2005-CLAIMS 1.1.pdf

814-kol-2005-correspondence.pdf

814-KOL-2005-DESCRIPTION (COMPLETE) 1.1.pdf

814-KOL-2005-DRAWINGS 1.1.pdf

814-kol-2005-examination report.pdf

814-KOL-2005-FORM 1.1.1.pdf

814-kol-2005-form 13.1.pdf

814-KOL-2005-FORM 13.pdf

814-kol-2005-form 18.pdf

814-KOL-2005-FORM 2.1.1.pdf

814-kol-2005-form 3.pdf

814-kol-2005-form 5.pdf

814-KOL-2005-FORM-27.pdf

814-kol-2005-gpa.pdf

814-kol-2005-granted-abstract.pdf

814-kol-2005-granted-claims.pdf

814-kol-2005-granted-description (complete).pdf

814-kol-2005-granted-drawings.pdf

814-kol-2005-granted-form 1.pdf

814-kol-2005-granted-form 2.pdf

814-kol-2005-granted-specification.pdf

814-KOL-2005-OTHERS.pdf

814-kol-2005-others1.1.pdf

814-KOL-2005-PETITION UNDER RULE 137.pdf

814-KOL-2005-PRIORITY DOCUMENT.pdf

814-KOL-2005-REPLY TO EXAMINATION REPORT.pdf

814-kol-2005-reply to examination report1.1.pdf

814-KOL-2005-SCHEDULE.pdf

abstract-00814-kol-2005.jpg

« Previous Patent

Next Patent »

Patent Number

248933

Indian Patent Application Number

814/KOL/2005

PG Journal Number

37/2011

Publication Date

16-Sep-2011

Grant Date

12-Sep-2011

Date of Filing

06-Sep-2005

Name of Patentee

SIEMENS AKTIENGESELLSCHAFT

Applicant Address

WITTELSBACHERPLATZ, 80333 MUNCHEN, GERMANY

Inventors:

#	Inventor's Name	Inventor's Address
1	DANIEL RINCK	RONTGENSTR. 17, 91080 UTTENREUTH, GERMANY
2	MICHAEL SCHEUERING	SCHUHSTR. 52, 91052 ERLANGEN, GERMANY

PCT International Classification Number

N/A

PCT International Application Number

N/A

PCT International Filing date

PCT Conventions:

#	PCT Application Number	Date of Convention	Priority Country
1	102004043694.0	2004-09-09	Germany