Data- or patient-specific vascular segmentation

Abstract

The vessels of a biological object are to be segmented more reliably. To this end, a method for training a machine learning algorithm to segment such vessels is proposed. First, a 3D reconstruction (8) of the vessels is provided. A starting vessel region (7) is identified in the 3D reconstruction (8). Sub-regions representing a starting vessel segment are extracted from the starting vessel region (7). The algorithm is trained using these extracted sub-regions. Subsequently, the trained algorithm is applied to a first neighboring region (10) immediately adjacent to the starting vessel region (7). This identifies a first vessel segment in the first neighboring region (10). Finally, the algorithm is retrained using the first vessel segment identified in the first neighboring region (10).

Description

[0001] The present invention relates to a method for training a machine learning algorithm for segmenting blood vessels. For this purpose, a 3D reconstruction of a biological object including the blood vessels is provided. Furthermore, the present invention relates to a method for segmenting blood vessels, a medical imaging device, and a computer program.

[0002] Performing 3D-DSA (three-dimensional digital subtraction angiography) is a common method for assessing cerebral vessels without bone interference in the visualization. 3D-DSA relies on an initial mask scan without contrast agent injection. However, from a clinical perspective and in terms of usability, this is not optimal for the following reasons. Firstly, movements between scans can generate unwanted artifacts. Secondly, the contrast agent injection must be precisely timed with the mask scan. Furthermore, the entire scan, including the mask scan and the initial scan, takes twice as long. Additionally, an extra dose of X-ray radiation is required for the mask scan. This is not necessarily problematic, however, as the subtraction artifacts can be reduced due to the comparatively small number of projections.Without a mask image, the number of projections for the fill image would probably have to be increased.

[0003] To distinguish bones or vessels from the rest of the image, a supervised trained AI (artificial intelligence) can be used. This requires labeled training data. Using bare 3D DSA as data is unsuitable due to undersampling artifacts. This problem could be circumvented by increasing the number of projections and the dose.

[0004] In the article "Multiscale vessel enhancement filtering" by Alejandro F. Frangi, Wiro J. Niessen, Koen L. Vincken, and Max A. Viergever, published in the Medical Image Computing and Computer-Assisted Intervention-MICCAI'98: First International Conference, Cambridge, MA, USA, October 11-13, 1998, Proceedings 1, pages 130-137, Springer Berlin Heidelberg; https: / / link.springer.com / chapter / 10.1007 / bfb0056195, a model-based method for quantitative 3D MRA (magnetic resonance angiography) is proposed. Linear vessel segments are modeled with a central vessel axis curve coupled to a vessel wall surface. This results in a so-called "vesselness filter," which, however, often classifies bone as a vessel and thereby overlooks less contrast-rich distal arterial vessels and veins. A further development of the method is described in: AF Frangi, WJ Niessen, RM Hoogeveen, T. van Walsum and MAViergever, “Model-based quantitation of 3-D magnetic resonance angiographic images,” in IEEE Transactions on Medical Imaging, vol. 18, no. 10, pp. 946-956, Oct. 1999, doi: 10.1109 / 42.811279. https: / / ieeexplore.ieee.org / document / 811279.

[0005] The object of the present invention is to enable the simpler yet reliable segmentation of vessels of a biological object.

[0006] According to the invention, this problem is solved by the subject matter of the independent claims. Advantageous embodiments of the invention are set out in the dependent claims.

[0007] According to the invention, a method for training a machine learning algorithm for segmenting blood vessels is provided. The machine learning algorithm can be, for example, an artificial neural network, a support vector machine, or the like. The aim is to segment blood vessels. This means that the vessels should be detected as belonging together and distinguished from foreign elements (e.g., bones). The blood vessels are associated with a biological object, such as the brain, liver, kidney, etc. Preferably, the biological object is a part of the human body. However, it can also be a part of an animal body or a part of a plant.

[0008] In the method according to the invention, a 3D reconstruction of the biological object including its vessels is first provided. This 3D reconstruction forms the basis for the segmentation of the vessels. The aim is thus to perform spatial vessel segmentation based on a 3D vessel image.

[0009] First, a source vessel region is identified in the 3D reconstruction. This source vessel region serves as the starting point for vessel segmentation. It should be an area of ​​the 3D reconstruction that is certain to contain one or more vessels. For example, the trunk of a vascular tree is suitable as a source vessel region. The source vessel region is identified manually, automatically, or semi-automatically.

[0010] In a further step, sub-areas representing a segment of the original vessel are extracted from the source vessel region. In other words, so-called "patches" are extracted from the source vessel region, each representing a vessel, a vessel segment, or a part thereof. This extraction can be performed manually, automatically, or semi-automatically.

[0011] In a subsequent step, the algorithm is trained using the extracted sub-sections. The algorithm thus learns the extracted vessel segments in the original vessel region using machine learning.

[0012] The trained algorithm is then applied to a first neighboring region immediately adjacent to the original vessel region, thereby identifying a first vessel segment in this region. The trained algorithm is thus used to recognize or segment vessel segments in a neighboring region directly adjacent to the original vessel region. The trained algorithm can also be applied to other first neighboring regions, if necessary, that are also in close proximity to the original vessel region. This allows the trained algorithm to be applied to multiple neighboring regions of the original vessel region. In the first neighboring region, one or more first vessel segments are identified using the algorithm.Similarly, in the other first neighboring regions immediately surrounding the initial vascular region, the respective first vascular segments can also be determined or identified using the trained algorithm.

[0013] Finally, the algorithm is retrained using the first vessel segment identified in the first neighboring region. Typically, a first vessel segment identified in the first neighboring region is thinner and usually has less contrast than a corresponding vessel segment in the original vessel region. Therefore, during retraining, the algorithm is made more sensitive, for example, to smaller vessels. In particular, the algorithm is also specifically tailored to the biological object or the patient during retraining. This can thus be described as dataset-specific or patient-specific training.

[0014] Advantageously, this makes it possible to reliably segment vessels without creating a mask image. Particularly advantageous is the ability to train the machine learning algorithm based on the 3D data or in a patient-specific manner.

[0015] In one embodiment, the provision of the 3D reconstruction is preceded by the acquisition of projection images and a 3D reconstruction based on them. Thus, projection images are first acquired. This acquisition preferably takes place after the administration of the contrast agent. A 3D reconstruction is then derived from the projection images and made available for further processing. This allows for the provision of a very up-to-date 3D reconstruction.

[0016] In another embodiment, the first vessel segment is determined by identifying corresponding voxels. The first vessel segment, and optionally one or more further vessel segments, are identified based on the corresponding affected voxels. In particular, a set of voxels corresponding to the first vessel segment is identified. This allows vessel segments to be determined with voxel-level precision.

[0017] According to another embodiment, the initial vascular region or the first vessel segment is identified using a vessel filter. Such a vessel filter is also referred to as a "vesselness filter." It is based on a heuristic approach and highlights tube-like structures. This allows for the reliable detection of vascular structures.

[0018] In another embodiment, the identification of the initial vascular region in the 3D reconstruction is achieved by determining an area around a center of high or low intensity within the reconstruction. For example, a center of high intensity is identified in the 3D reconstruction, indicating a high density of vessels or a large vessel filled with contrast medium. This center of intensity thus represents a region that most likely contains one or more vessels. Therefore, this vascular region can be used as the initial vascular region from which the algorithm can be trained. The center of intensity thus reliably identifies an initial vascular region for training the algorithm.

[0019] In another embodiment, the initial vessel region has multiple faces, each with a first neighboring region. The trained algorithm is then applied to all first neighboring regions, and the corresponding results are used for retraining. Thus, the initial vessel region in the 3D reconstruction represents a starting volume with multiple faces. For example, a cuboid as the initial vessel region has six faces, while a triangular pyramid as the initial vessel region has four. Other shapes can also be used as the initial vessel region. Each of the faces of the initial vessel region has a first neighboring region. Optionally, each face has its own first neighboring region. The algorithm trained using the initial vessel region is then applied to all first neighboring regions.The results of the segmentations of neighboring regions are then used to retrain the algorithm. The algorithm is therefore trained from the center outwards, because the result is most reliable in the center, while the segmentation results are correspondingly less reliable in the more distant neighboring regions, where the vessels are smaller and the contrast agent concentrations are lower.

[0020] In a further development, a second neighboring region directly borders each of the lateral faces of the first neighboring regions. As a rule, every second neighboring region is located further away from the center of the originating vessel region than every first neighboring region. The second neighboring regions thus surround the first neighboring regions, and these in turn surround the originating vessel region, similar to the structure of an onion. In a special embodiment, second neighboring regions directly border all lateral faces of the first neighboring regions that point away from the center.

[0021] According to a further embodiment, the algorithm, trained as described above, is applied to one (or more) of the second neighboring regions that are immediately adjacent to the first neighboring region and whose center of gravity is located further from the center of the original vessel region than the center of gravity of the first neighboring region. This identifies a second vessel segment (or segments) in the second neighboring region, and the algorithm is then retrained using this second vessel segment. Thus, vessel segments are identified in the respective neighboring regions and used for retraining. The application of the algorithm and the retraining process, as already performed with respect to the first neighboring region, are therefore repeated with respect to the second neighboring region. This further refines the algorithm.

[0022] In one embodiment, the initial vessel region is a cuboid or cube, and the first neighboring region directly adjoins a face of the cuboid. Thus, the first neighboring region borders directly on a corresponding face of the initial vessel region. The entire 3D reconstruction can be readily divided into its respective regions using cubes.

[0023] In one embodiment, the second neighboring region touches only a single edge of the cuboid of the starting vessel region. In the case of the cuboid subdivision of the 3D reconstruction, this means that the second neighboring region not only directly adjoins a side of the first neighboring region, but also touches an edge of the cuboid of the starting vessel region. This allows the volume around the starting vessel region to be completely filled by the respective neighboring regions.

[0024] According to another embodiment, the second neighboring region is in contact with one side of the first neighboring region and with another first neighboring region that also directly borders the starting vessel region. This means that the second neighboring region borders at least two sides of each of the two first neighboring regions. In this way, the volume of the 3D structure can be completely divided into the respective regions.

[0025] It was described above that the algorithm is retrained with respect to second neighboring regions to refine it. This retraining can be extended from the initial vessel region, through the first and second neighboring regions, outwards to third neighboring regions, and so on. Accordingly, in one embodiment, the algorithm trained as described above can be applied to one (or more) third neighboring regions that are immediately adjacent to one of the second neighboring regions and whose center of gravity is located further from the center of the initial vessel region than the center of gravity of the second neighboring region. This identifies a third vessel segment (or segments) in the third neighboring region, and the algorithm is then retrained using this third vessel segment.In this way, the algorithm is continuously refined towards the outside and can thus reliably segment even smaller vessels that are less contrasting.

[0026] According to another embodiment, one of the neighboring regions, which is farther from the starting vessel region than another of the neighboring regions, is smaller than the other neighboring region. This means that the neighboring regions can become progressively smaller towards the outside (starting from the starting vessel region). This is particularly advantageous when the vessels become progressively smaller towards the outside and the segmentation needs to become increasingly finer.

[0027] In another embodiment, the method is combined with a graph-based approach to identify the respective vessel segments. The graph-based approach is based on graph theory, according to which any structures are represented by nodes and edges. Such a method can be helpful here, since the vessel segments must merge into one another at the regional boundaries.

[0028] In another embodiment, it is provided that, during each application or retraining of the algorithm, at least a portion of a neighboring region that was already used in a previous retraining is also utilized. This means that regions overlapping with parts of the training cohort are used during the application or retraining of the algorithm. In this way, improved continuity can be achieved.

[0029] According to another embodiment, the 3D reconstruction is provided with time resolution, and the extraction of sub-areas and / or the determination of the first vessel segment is performed (also) depending on intensity fluctuations of corresponding voxels. In this case, a time-resolved 3D reconstruction is performed, which can also be referred to as 4D reconstruction. This method utilizes the fact that vessels change their volume or diameter during the natural pulse of, for example, a patient, causing the contrast-filled vessel to exhibit corresponding pulsatility. In other words, the intensity of the voxels of the vessel segments changes with the pulse of the vessels. This temporal information allows for differentiation between vessels and bones, as bones do not exhibit such pulsatility.Accordingly, the pulsatility of the 4D reconstruction can be used to validate the vascular classification. Consequently, the extracted segments or identified vascular sections can be classified more reliably with this information.

[0030] According to another embodiment, each vessel segment is identified based on intensity fluctuations of corresponding voxels. This means that in all neighboring regions defined for training the algorithm, pulsatility is checked and plausibility is verified accordingly.

[0031] According to the invention, a method for segmenting vessels by applying an algorithm trained as described above to the aforementioned or another 3D reconstruction is also provided. This means, in particular, that the algorithm trained in several iterations can be applied to an entire 3D reconstruction.

[0032] Furthermore, according to the invention, a medical imaging device with a computing unit is also provided, which is configured to perform the aforementioned segmentation method. In particular, the computing unit is capable of automatically performing several or all of the steps of the aforementioned training method.

[0033] Furthermore, according to the present invention, a computer program is also provided which includes commands that, when executed by a medical imaging device of the type mentioned, cause it to perform a method described above. Optionally, a computer program product containing such commands can also be provided.

[0034] The algorithms mentioned above for segmenting the vessels can be based on an artificial neural network, a support vector machine, or another machine learning method.

[0035] The present invention will now be explained in more detail with reference to the accompanying drawings, which show: Fig. 1 a schematic view of an imaging device; Fig. 2. a 3D reconstruction with vascular regions; and Fig. 3 a schematic flowchart of an embodiment of a method according to the invention.

[0036] The exemplary embodiments described in more detail below represent preferred embodiments of the present invention.

[0037] In Fig. Figure 1 schematically illustrates an exemplary embodiment of an imaging system 1 according to the invention. The imaging system 1 comprises at least one computing unit 2, a contrast agent injector 5, and an imaging modality which, in the present example without limitation of generality, is configured as an X-ray-based imaging modality, for example for perfusion imaging or CT angiography or DAS (digital subtraction angiography), and comprises an X-ray source 4 and an X-ray detector 3.

[0038] The at least one computing unit 2 is configured to carry out a method according to the invention for segmenting, in particular, vessels in 3D or 4D reconstructions.

[0039] Certain organs of living beings exhibit tree-like vascular structures. Starting from a larger vessel, more and more sub-vessels branch off towards the outside. The greater the distance from the main vessel, the finer the individual vessels become. This is also the case, for example, in the human brain. Fig. Figure 2 shows a corresponding vascular structure of a human brain.

[0040] Generally, such vascular structures also occur in other organs that have a convex shape. Examples of such convex objects include the liver, other internal organs, and individual limbs.

[0041] In the following embodiment, a first step S1 of the method involves the acquisition of projection images. In particular, a set of contrast-enhanced projection images is acquired, for example, using the imaging device of Fig. 1 won. However, the projection images can also be provided in other ways.

[0042] In a second step S2, a 3D reconstruction is performed from the projection images. Fig. Figure 2 shows such a 3D reconstruction of the blood vessels of a brain. If necessary, the projection images are also temporally resolved in step S1, and a 4D reconstruction, i.e., a temporally variable 3D reconstruction, is performed in step S2. Thus, at least one 3D reconstruction of a biological object (here: brain) with blood vessels can be provided. However, the provision of the 3D or 4D reconstruction can also be achieved via data carriers, networks, and the like.

[0043] In a third step, S3, a starting vascular region 7 is identified in the 3D reconstruction 8. The starting vascular region 7 is identified, for example, by a "certain" vessel 9. This certain vessel 9 is, for instance, the largest vessel or the largest vascular cluster in the 3D reconstruction 8. This large vessel or vascular cluster can be used as the starting point for training a segmentation algorithm. The voxels within this region can be reliably assigned to the class "vessels." Therefore, these voxels very likely belong to the contrasted vascular system.

[0044] Identifying the source vessel region 7 can be achieved using various methods. For example, a strong vesselness filter can be used, which only reacts to large, high-contrast vessels. Alternatively or additionally, to identify the source vessel region 7, a region around a foci of high intensity in the 3D reconstruction 8 can be sought. Depending on the imaging, this could, of course, also be a foci of low intensity. An area of ​​interest is defined around this foci, which can be considered the source vessel region 7.

[0045] In a fourth step, S4, sub-areas representing an initial vessel segment are extracted from the original vessel region. These sub-areas can also be referred to as "patches." These sub-areas represent regions in the 3D reconstruction 8 that embody vessels or parts thereof.

[0046] Optionally, a plausibility check can be performed at this point, for example in step S5. The patches extracted in step S4 can be validated, for instance, by checking whether they pulsate in accordance with a patient's heartbeat. If they do not, it is highly likely that the patch is not a blood vessel but a bone. Conversely, if the relevant voxels show pulsating intensity, it is almost certainly a blood vessel.

[0047] In a sixth step (S6), the machine learning algorithm can be trained. This algorithm can be based, for example, on an artificial neural network or a support vector machine. The algorithm is trained using the extracted subsets. In particular, this allows for the development of a semantic segmentation algorithm based on the identified safe container 9.

[0048] In a subsequent seventh step S7, the trained algorithm is applied to one or more neighboring regions 10. For example, the initial vessel region 7, which encloses the safe vessel 9, is cuboid or cube-shaped. For the following considerations, the cube shape of the initial vessel region 7 and the neighboring regions is always assumed. A neighboring region 10 can adjoin the initial vessel region 7 in each spatial direction. In the specific example of Fig. 2. The starting vessel region 7 is designed as a cube with six sides. Accordingly, a first neighboring region 10 could be directly connected to each side. In the example of Fig. 2. A first neighboring region 10 is missing on the underside of the cube of the initial vessel region 7, since the 3D reconstruction 8 ends at the lower end of the initial vessel region 7. Corresponding first neighboring regions 10 can also be located on the sides parallel to the image plane, i.e., above and below the image plane.

[0049] The previously trained algorithm can now be applied to a first neighboring region 10 (preferably all of them) that is immediately adjacent to the initial vessel region 7, thereby identifying a first vessel segment in the first neighboring region 10. This means that the segmentation algorithm identifies a vessel or vessel segment in the first neighboring region 10. The vessels in the first neighboring regions 10 are generally still relatively large, so they are detected with a high degree of certainty, even though the algorithm was only trained on the largest vessel or vessel cluster, i.e., the safe vessel 9.

[0050] In a further step, S8, a decision can be made as to whether the training is complete or not. If not "n", the procedure jumps back to step S6, and retraining takes place with the first vessel segments segmented in step S7 in the first neighboring region(s).

[0051] The steps of training S6 and applying S7 can be repeated multiple times, with each iteration focusing on more distant neighboring regions.

[0052] The algorithm trained on the first vessel segment or segments can thus be applied to second neighboring regions 11 located further outwards in a radial direction from the initial vessel region 7. For example, if every second neighboring region 11 is also cube-shaped, it can directly adjoin the sides of the first neighboring regions 10, as shown in Fig. Figure 2 shows that a second neighboring region 11 can be attached to each face of the first neighboring regions 10 that faces outwards with respect to the starting vessel region 7. In this arrangement of the cubes, some second neighboring regions 11 would then each touch an edge of the cube of the starting vessel region 7, but not its faces.

[0053] The vessels in the second neighboring regions 11 are usually smaller on average than those in the first neighboring regions 10. However, since the segmentation algorithm is already trained with the vessels of the first neighboring regions 10, and can therefore also recognize vessels with less contrast, it will also reliably segment the even smaller vessels in the second neighboring regions 11.

[0054] These training and application steps can be repeated, for example, until the entire 3D reconstruction is segmented. For instance, third neighboring regions (in Fig. (2 not shown) connect to the second neighboring regions 11 on the outside. However, the sequence of repetitions can also be terminated according to another criterion, for example a fixed number.

[0055] Once the algorithm training is complete (“j” in step S8), the fully trained algorithm can be applied to the entire 3D reconstruction 8 or to another 3D or 4D reconstruction, as described in step S9. All regions are segmented using the algorithm, which is also trained to detect the finest vessels. This ensures that even the finest vessels in the initial vessel region 7 and in the first neighboring regions 10, etc., are detected.

[0056] The training can also be terminated when a specific percentage of voxels are marked as vessels or until a predefined condition regarding a geometric distance is met.

[0057] Segmentation can lead to individual vessels not being connected to each other. To avoid this, the connectivity of the detected vessels can be improved by combining the segmentation based on the algorithm trained according to the invention with a graph-based method. In this way, a vascular tree with connected vessels can be more likely to be segmented.

[0058] To further increase the accuracy of the segmentation, the initial estimate of the safe vessel 9 can be compared with an anatomical atlas. In another advantageous embodiment, the neighboring regions to be classified are reduced in size in each iteration of the training. Specifically, the second neighboring region 11 could be smaller than the first neighboring region 10, and the third neighboring region could be smaller than the second. The further outward the neighboring regions are located, the smaller their volume would thus be. Reducing the size of the regions outwardly corresponds to the fact that vessels also generally become smaller outwardly. Furthermore, particularly in the brain, the outer regions approach the skull bone, so the smaller regions at the periphery allow for better differentiation from the bone.

[0059] In another embodiment, the neighboring regions to be classified can be superimposed with portions of the training cohort in each iteration. This means that in each iteration, portions of the newly classified voxels and the voxels used for training are mixed. Optionally, the newly classified voxels can be weighted more heavily. This allows for more reliable segmentation of the different vessel thicknesses.

[0060] The method according to the invention is not limited to vascular segmentation in the brain. Rather, the method can also be applied to other body regions, e.g., the abdominal cavity (liver, organs in the pelvis).

[0061] Particularly advantageous is the creation of a patient- and data-specific semantic segmentation algorithm that can be iteratively trained on new subsets of the recorded data and applied immediately.

[0062] In the preceding description, regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are to be included. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited non-patent literature

[0000] Alejandro F Frangi, Wiro J Niessen, Koen L Vincken, Max A Viergever: “Multiscale vessel enhancement filtering”; Medical Image Computing and Computer-Assisted Intervention-MICCAI'98: First International Conference Cambridge, MA, USA, October 11-13, 1998 Proceedings 1, pages 130-137; Springer Berlin Heidelberg

[0004] https: / / link.springer.com / chapter / 10.1007 / bfb0056195

[0004] AF Frangi, WJ Niessen, RM Hoogeveen, T. van Walsum, and MA Viergever, “Model-based quantitation of 3-D magnetic resonance angiographic images,” in IEEE Transactions on Medical Imaging, vol. 18, no. 10, pp. 101-1 946-956, Oct. doi: 10.1109 / 42.811279

[0004] https: / / ieeexplore.ieee.org / document / 811279

[0004]

Claims

[1] Method for training a machine learning algorithm to segment vessels by - Providing (S2) a 3D reconstruction (8) of a biological object with the vessels; characterized by - Identifying (S3) an originating vascular region (7) in the 3D reconstruction (8), - Extracting sub-areas representing a source vessel segment from the source vessel region (7), - Training (S6) the algorithm with the extracted sub-areas, - Applying (S7) the algorithm trained in this way to a first neighboring region (10) that is immediately adjacent to the starting vessel region (7), thereby identifying a first vessel segment in the first neighboring region (10), - Retraining the algorithm with the first vessel segment determined in the first neighboring region (10). [2] Method according to claim 1, wherein the provision (S2) of the 3D reconstruction (8) is preceded by the acquisition (S1) of projection images and a 3D reconstruction based thereon. [3] Method according to claim 1 or 2, wherein the first vessel segment is determined by identifying corresponding voxels. [4] Method according to any of the preceding claims, wherein the identification of the starting vessel region (7) or the first vessel segment is carried out using a vessel filter. [5] Method according to any of the preceding claims, wherein the identification of the starting vessel region (7) in the 3D reconstruction (8) is carried out by determining a region around a center of high or low intensity in the 3D reconstruction (8). [6] Method according to one of the preceding claims, wherein the initial vessel region (7) has several side surfaces, each of which is bordered by a first neighboring region (10), the trained algorithm is applied to all first neighboring regions (10), and the corresponding results are used for retraining. [7] Method according to claim 6, wherein several side surfaces of the first neighboring regions (10) are directly adjacent to each of a second neighboring region (11). [8] Method according to claim 7, wherein the algorithm trained according to claim 1 is applied to one of the second neighboring regions (11) which is immediately adjacent to the first neighboring region (10) and whose center of gravity is further away from the center of the initial vessel region (7) than a center of gravity of the first neighboring region (10), thereby identifying a second vessel segment in the second neighboring region (11), and the algorithm is retrained with the second vessel segment identified in the second neighboring region (11). [9] Method according to any of the preceding claims, wherein the starting vessel region (7) is a cuboid and the first neighboring region (10) is directly connected to a side face of the cuboid. [10] Method according to claims 8 and 9, wherein the second neighboring region (11) touches only a single edge of the cuboid of the starting vessel region (7). [11] Method according to claim 10, wherein the second neighboring region is in contact with one side of the first neighboring region and with a further first neighboring region, which is also directly adjacent to the starting vessel region. [12] Method according to any one of claims 8 to 11, wherein the algorithm trained according to claim 8 is applied to a third neighboring region which is immediately adjacent to one of the second neighboring regions (11) and whose center of gravity is further away from the center of the initial vessel region (7) than a center of gravity of the second neighboring region (11), thereby identifying a third vessel segment in the third neighboring region, and the algorithm is retrained with the third vessel segment identified in the third neighboring region. [13] Method according to claim 12, wherein one of the neighboring regions which is farther from the starting vessel region (7) than another of the neighboring regions is smaller than the other neighboring region. [14] Method according to any of the preceding claims, which is combined with a graph-based method to identify respective vessel segments. [15] Method according to one of the preceding claims, wherein in each application (S7) or retraining of the algorithm at least a part of a neighboring region is also used which was already used in a previous retraining. [16] Method according to one of the preceding claims, wherein the 3D reconstruction is provided in a time-resolved manner, and the extraction of the sub-areas and / or the determination of the first vessel segment is carried out depending on intensity fluctuations of corresponding voxels. [17] Method according to claim 8 or 12, wherein the determination of each vessel segment is carried out depending on intensity fluctuations of corresponding voxels. [18] Method for segmenting vessels by applying an algorithm trained according to one of the preceding claims to the or another 3D reconstruction (8). [19] Medical imaging device (1) with a computing device (2) configured to perform the method according to claim 18. [20] Computer program comprising commands which, when executed by a medical imaging device (1) according to claim 19, cause it to execute a method according to any one of claims 1 to 18.

Images