Method for processing three-dimensional medical input image data and for parameterizing a detection algorithm, data processing device and computer program

Abstract

Computer-implemented method for processing three-dimensional medical input image data (33), comprising the steps: - Receiving the input image data (33) describing the image values ​​(34) for a plurality of input image voxels (35), - Determining a respective vessel-likeness measure (36) for at least parts of the input image voxels (35), wherein the vessel-likeness measure (36) describes a determined degree of similarity of a structure depicted in the respective input image voxel (35) to an image of a vessel, and - on the one hand, applying a filter algorithm (38) to the input image data (33) to provide filtered image data (39), wherein the parameter value (41) of at least one filter parameter (42) of the filter algorithm (38) is specified depending on at least one of the vascularity measures (36), or - on the other hand, scaling the image values ​​(34) of the input image voxels (35) to provide scaled image data (37), wherein a respective scaling factor (40) used to scale the respective input image voxel (35) is specified depending on at least one of the vascularity measures (36).

Description

[0001] The invention relates to a computer-implemented method for processing three-dimensional medical input image data. The invention also relates to a processing device and a computer program.

[0002] In medical imaging, the achievable image quality may be limited by the characteristics of the imaging equipment used. To reduce patient exposure, it may also be desirable, for example, to use very low X-ray doses and / or to avoid contrast agents in imaging wherever possible. While in some imaging procedures, such as many magnetic resonance imaging protocols, contrast improvement or noise reduction can be achieved through longer measurement times, long measurement times reduce the acceptance of an examination, increase its costs, and can also be detrimental to the achievable image quality due to, for example, motion artifacts. In many applications, it is therefore essential to sufficiently improve the image quality through suitable post-processing or filtering of the image data to enable robust evaluation.

[0003] Such postprocessing is particularly relevant for low-field magnetic resonance measurements, especially when measurements are to be performed with a main magnetic field of less than one Tesla. While low-field measurements can be performed with significantly less expensive measuring equipment than other magnetic resonance measurements and can thus significantly improve the local availability of magnetic resonance measurements, lower signal-to-noise ratios are achieved in these systems.

[0004] Thus, for example, robust magnetic resonance angiography in the low-field range is typically only possible using contrast medium.

[0005] However, since low-field magnetic resonance angiography would be particularly advantageous, for example, for carrying out certain preventive examinations in rural areas, it would be advantageous to be able to use contrast-free angiographic measurement methods, such as ASL (arterial spin labeling) or time-of-flight methods, even in the low-field range.

[0006] However, since conventional methods for contrast enhancement and noise reduction and / or background suppression reach their limits in this application area, the development of further approaches to improve image quality is appropriate.

[0007] One possible approach to this is the use of noise reduction algorithms trained by machine learning, particularly deep learning. However, since the deployment of such an algorithm can be complex, especially when used in medical imaging, the risk of image distortion caused by such an algorithm must be minimized, it is advisable to develop other approaches for post-processing medical image data or to further optimize machine learning-based approaches.

[0008] An anisotropic diffusion filter is known from the paper by Canero, C. and Radeva, P.: “Vesselness enhancement diffusion”, Pattern recognition letters, Vol. 24, 2003, No. 16, pp. 3141-3151, ISSN 0167-8655.

[0009] The invention is therefore based on the object of specifying an improved approach for improving the quality of medical image data, in particular image data from magnetic resonance angiography, especially at low field strengths, and / or for background suppression, in particular in angiography.

[0010] The task is solved by a computer-implemented method for processing three-dimensional medical input image data, which comprises the following steps: - Obtaining the input image data describing image values ​​for a plurality of input image voxels, - determining a respective vascularity measure for at least parts of the input image voxels, wherein the vascularity measure describes a degree of similarity of a structure depicted in the respective input image voxel to a vessel, and - applying a filter algorithm to the input image data to provide filtered image data, wherein the parameter value of at least one filter parameter of the filter algorithm is specified as a function of at least one of the vascularity measures.

[0011] Optionally, the image values ​​of the input image voxels can be scaled to provide scaled image data, wherein a respective scaling factor used to scale the respective input image voxel is specified as a function of at least one of the vascularity measures.

[0012] The relevant structures imaged in the context of certain medical imaging processes, in particular in the context of angiography, are at least approximately vascular and thus in particular tubular or cylindrical in shape and are thus imaged or localized by the vascularity measures used according to the invention, which can in particular form a three-dimensional vascularity map.

[0013] By scaling the image values ​​of different input image voxels differently depending on the respective local vascularity measure, input image voxels with a vascularity measure value above a certain threshold can, for example, be scaled with a higher scaling factor than input image voxels whose vascularity measure falls below the threshold. Alternatively, the image value of each input image voxel can be scaled proportionally to the vascularity measure value determined for this input image voxel. By scaling depending on the local vascularity measure, a background can be suppressed, for example, and an imaged vascular tree can be highlighted.

[0014] By parameterizing the filter algorithm depending on the vascularity measures, different filtering strengths can be performed, for example, within the vascular region and outside the vascular region, i.e., for input image voxels with a vascularity measure value above or below a threshold. The parameter value of the filter parameter, or at least one of the filter parameters, can thus specify the filtering strength.

[0015] For example, a narrower filter kernel, such as a Gaussian curve with lower variance, can be used in the vascular region than outside the vascular region. The filter parameter or a specific filter parameter can thus specify a width of the filter kernel, such as the variance of a Gaussian kernel, and / or a shape of the filter kernel. Filtering or blurring the background region more strongly than in a vascular tree facilitates the analysis of the vascular tree, as background noise is suppressed without overly blurring the regions of the vascular tree itself.In addition to an improvement in quality, background suppression can be achieved by strong filtering or suitable choice of filter parameters in the background area in addition to or alternatively to the above-mentioned scaling, whereby background structures, especially larger background structures that a viewer can orient himself on, can still remain visible.

[0016] For example, an anisotropic diffusion filter or similar can be used and parameterized differently locally depending on the degree of vascularity to achieve the described effect. Particularly if the filter algorithm is an anisotopic filter algorithm, the filter parameter or at least one of the filter parameters can also depend on the direction in which the respective detected vascular structure extends, as will be explained in more detail later. This can, for example, result in stronger filtering in the longitudinal direction of the vascular structure than perpendicular to it.

[0017] In addition to or as an alternative to scaling the image values ​​of the input image voxels of the input image data depending on the vascularity measures, a corresponding scaling can also be performed for the filtered image data or the scaled image data can be filtered, as will be explained in more detail later.

[0018] To determine suitable vascularity measures, the algorithm referred to as a "line filter" from the publication Sato, Y. et al. (1997), 3D multi-scale line filter for segmentation and visualization of curvilinear structures in medical images, in Troccaz, J., Grimson, E., Mösges, R. (eds) CVRMed-MRCAS'97. CVRMed MRCAS 1997 1997, Lecture Notes in Computer Science, vol 1205, Springer, Berlin, Heidelberg. https: / / doi.org / 10.1007 / BFb0029240, can be applied to the input image data or to an intermediate result determined, for example, by interpolation of the input image data.

[0019] To calculate the vascularity measure, a value matrix associated with the respective input image voxel can first be determined. This value matrix can be determined, in particular, by convolving the Hessian matrix of a Gaussian kernel G or another filter kernel with the input image data. The entries of the value matrix can, in particular, be lab(x⇀,σ)={δδaδδbG(x⇀,σ)}∗I(x⇀) where I is the image value of the input image voxel at the position vector x⇀ indicates the specified position. I ab is the respective matrix entry for this input image voxel, where a,b ∈ {x,y,z} indicates the derivative in the respective spatial direction and the different matrix entries I ab different combinations or sequences of derivatives are assigned.

[0020] The variance σ of the Gaussian distribution G selects the diameter of the tubular or cylindrical structures for which the determination of vascularity should be selective. σ can be selected by the user, for example, dynamically using a slider or similar during image optimization, or determined automatically, for example, using the multiscale approach mentioned below.

[0021] From the value matrix, three eigenvalues ​​λ1, λ2, λ3 can then be calculated for the respective input image voxel. For linear structures, the eigenvalue for the eigenvector in the longitudinal direction of the vessel-like structure is approximately zero. The eigenvalues ​​for the eigenvectors perpendicular to this direction are similar to each other and much smaller than zero for cylindrical vessel-like structures.

[0022] Cluster-like or planar structures lead to significantly different eigenvalue combinations, so that the vascularity measure can be determined based on the eigenvalues. Various specific calculations are known for this purpose. The previously cited publication by Y. Sato et al., for example, suggests the following vascularity measure λ 123 before: λ123={|λ3|(λ2λ3)γ23(1+λ1|λ2|)γ12,if λ3<λ2<λ1≤0|λ3|(λ2λ3)γ23(1−αλ1|λ2|)γ12,if λ3<λ2<0<λ1<|λ2|α0, in other cases

[0023] By choosing the quantities α, γ 12 and γ 23 The selectivity of the vascularity measure can be adjusted. α should be between 0 and 1 and γ 12 and γ 23 should be chosen to be greater than or equal to zero. Higher values ​​of γ 23 False detections of surface-like structures as line-like structures can be suppressed and by higher values ​​of α and γ 12the value of the vascularity measure is reduced for noise-like structures or one-dimensional structures. Good results are typically achieved with α=0.25 and γ 12 =γ 23 =1 is achieved, whereby, depending on the specific application, a variation of this value can lead to a further improvement in the detection of vascular-like structures.

[0024] Further possible implementation details for performing the described calculation, approaches for the approximate calculation of the matrix entries to reduce the required computing time and an extension of the described approach for the multi-scalar detection of vascular-like structures by determining the vascularity measure for different values ​​of the variance σ of the Gaussian distribution are already discussed, for example, in the above-cited publication by Y. Sato et al. and will therefore not be repeated here.

[0025] Obtaining input image data can comprise acquiring the input image data, for example performing a magnetic resonance angiography or another measurement protocol for medical imaging. However, it is also possible for the acquisition of the input image data to precede the method and not be part of the method. For example, the input image data can be retrieved from a server or read from an internal memory of a processing device performing the method. The input image data can, in particular, be measurement data or three-dimensional image data reconstructed from this measurement data. However, it is also possible for the image data to already be the result of preprocessing, for example upstream filtering or scaling.

[0026] The filter algorithm can determine the image value of a respective filtered voxel of the filtered image data depending on the image values ​​of several of the input image voxels, wherein different parameter values ​​for the filter parameter or at least one of the filter parameters are used to determine the image values ​​of at least two of the filtered voxels. Thus, the method according to the invention preferably employs locally different parameterization. This allows the filtering to be adapted to the locally detected vascular structure, thus further improving the image quality.

[0027] The differences in the local parameterization of the filtering can depend exclusively on the local vascularity measure or the vascularity map. However, the directions of detected vascular structures can also be considered in addition or alternatively, for example, to select the direction in which a particularly strong or particularly weak filtering occurs with an anisotropic filter, as will be explained in more detail below.

[0028] A subset of the input image voxels can be selected, each of which is part of a vascular structure according to its vascularity measure, wherein the parameter value of the filter parameter or at least one of the filter parameters for the respective filtered voxel can be selected depending on whether an input image voxel assigned to the respective filtered voxel is an element of the subset and / or how far the assigned input image voxel is from the nearest input image voxel that is an element of the subset. The input image voxel assigned to the respective filtered voxel can, in particular, be that input image voxel that is located at the same position in the input image data as the assigned filtered voxel is located in the filtered image data.In other words, input image voxels and filtered voxels located at the same position or at the same coordinates are assigned to each other.

[0029] By selecting locally different filter parameters as described above, strong filtering can be used, for example, in areas relatively far from vascular structures, thereby robustly suppressing background noise. In the area of ​​vascular structures or with increasing proximity to them, less strong filtering can be used, for example, by adjusting the width of a filter kernel or the diffusion constants of a discussion filter, to avoid or minimize blurring of vascular boundaries or similar unwanted filtering results.

[0030] According to the invention, an anisotropic filter algorithm is used whose filter parameter is an anisotropy parameter or whose filter parameters comprise at least one anisotropy parameter, wherein the respective anisotropy parameter controls a degree and / or a direction of the anisotropy of the filter algorithm, so that for at least some of the possible parameter values ​​of the anisotropy parameter, the filter algorithm has a different filter behavior for two spatial directions.

[0031] To achieve a high-quality representation of the vascular structure, strong filtering perpendicular to the vessel wall or longitudinal direction should be avoided, at least for filtered voxels located close to the transition area between the background and the vascular structures. This can be achieved using the described procedure by specifying the anisotropy of the filter algorithm by selecting appropriate parameter values ​​so that less strong filtering perpendicular to the longitudinal direction occurs for corresponding filtered voxels. This can be achieved, for example, by selecting a narrower filter kernel or by appropriately selecting the diffusion constants of a diffusion filter.

[0032] As already explained above, in order to calculate the respective vascularity measure, a value matrix assigned to the respective input image voxel can be determined as an intermediate step, whereby the respective vascularity measure is determined from eigenvalues ​​of the respective value matrix.

[0033] As shown above, the value matrix can be determined by convolving the Hessian matrix of a Gaussian kernel or the Hessian matrix of another filter kernel with the input image data.

[0034] To determine the image values ​​of all filtered voxels of the filtered image data or a subset of the filtered voxels, a parameter value of the anisotropy parameter or at least one of the anisotropy parameters can be used, which is determined as a function of at least one eigenvector of the value matrix assigned to a respective selected input image voxel. As already mentioned above, only an approximate calculation of the value matrix entries can be performed.

[0035] Since the convolution and derivative operations are interchangeable, and the second derivative of the image values ​​of the input image data along a vascular structure typically barely deviates from 0, while the second derivative perpendicular to the longitudinal direction of the vascular structure deviates significantly from 0, the determined eigenvectors specify the longitudinal direction of the vascular structure and / or the directions perpendicular to it if the selected input image voxel lies on a vascular structure. Thus, for example, stronger filtering can occur along the vascular structure than perpendicular to its wall, whereby the boundaries of the vascular structure remain clearly recognizable despite good noise suppression.

[0036] According to the invention, one of the input image voxels is permanently assigned to each of the filtered voxels, wherein the input image voxel selected for determining the anisotropy parameter of the respective filtered voxel is selected from the or a subset of the input image voxels selected on the basis of the vascularity measures, depending on the distance of the respective input image voxel of the subset to the input image voxel assigned to the respective filtered voxel.

[0037] Using the described procedure, the longitudinal direction of a vessel-like structure that is imaged by the associated input image voxel and thus also by the filtered voxel, or that is located near the associated input image voxel, can be taken into account, in particular, as explained above, to achieve stronger filtering in the longitudinal direction of this structure than perpendicular to the longitudinal direction. This can, in particular, prevent severe smearing or blurring of the transition between vessel and background, even with relatively strong overall filtering for noise reduction.

[0038] The subset of input image voxels can, in particular, comprise exclusively those input image voxels for which the value of the vascularity measure exceeds a threshold value or which are classified as depicting a vascular structure. From this subset, in particular, the input image voxel whose distance to the input image voxel assigned to the respective filtered voxel is minimal can be selected. If the input image voxel assigned to the respective filtered voxel is itself an element of the subset, this assigned input image voxel is selected as the selected input image voxel.

[0039] In an advantageous embodiment, the subgroup of filtered voxels for which the parameter value of the anisotropy parameter is determined as a function of the respective eigenvector comprises exclusively filtered voxels whose respective associated input image voxel is either itself an element of the selected subset of the input image voxels or has a distance to at least one input image voxel of this subset that is smaller than or corresponds to a predetermined distance limit value.

[0040] For filtered voxels whose associated input image voxels are sufficiently far away from any detected vascular structure, isotropic filtering can be performed, or the at least one anisotropy parameter can have a fixed value or, for example, vary depending on the distance to the nearest vascular structure without taking its longitudinal direction into account. Through the procedure explained, the direction of the extension of the vascular structure is only taken into account during filtering if it is actually relevant, for example, if, due to a relatively short distance of the filtered voxel from such a structure, image values ​​from within the structure are taken into account during filtering or would be taken into account if the at least one anisotropy parameter were not adjusted.This makes it possible, for example, to carry out strong filtering in all directions at a sufficient distance from the vascular structure and thus to suppress disturbances in the background of the image much more effectively, whereby in the area of ​​or close to vessels the anisotropy of the filtering can be adapted to the course of the vessel.

[0041] The three-dimensional medical input image data can be based on magnetic resonance angiography without contrast agent administration and / or with a main magnetic field strength of less than 1.5 T or less than 1 T. Magnetic resonance angiography without contrast agent administration or with low field strengths typically achieves significantly lower signal-to-noise ratios than would be the case with contrast agent or with higher fields. Therefore, high-quality filtering, such as that achieved by the method described, is particularly relevant in these application areas.

[0042] After applying the filter algorithm, the image values ​​of the filtered voxels of the filtered image data can be scaled to provide scaled image data, with a respective scaling factor used to scale the respective filtered voxel being specified as a function of at least one of the vascularity measures. Apart from using the filtered image data instead of the original input image data as input data, the scaling can correspond to the scaling of the image values ​​of the input image voxels explained above and serve the same purpose, i.e., in particular, background suppression. The features explained above for scaling the input image data can also be transferred to scaling the filtered image data with the advantages mentioned there.

[0043] Alternatively, the filter algorithm can be applied to the scaled image data to provide filtered image data, wherein the parameter value of at least one filter parameter of the filter algorithm is specified depending on at least one of the vascularity measures. Apart from using the scaled image data as input data, the filtering of the scaled image data can correspond to the filtering of the input image data explained above. The features explained above for filtering the input image data can also be applied to the filtering of the scaled image data, with the advantages mentioned therein.

[0044] The filtering and scaling explained above, depending on the at least one vascularity measure, can thus also be combined or carried out in any order in order to combine the advantages explained for the two approaches and, for example, to achieve particularly robust background suppression.

[0045] A detection algorithm trained by machine learning can be used as the filter algorithm, whereby both the image values ​​of the input image voxels and the filter parameters are used as input variables of the trained detection algorithm.

[0046] In general, a trained function replicates cognitive functions that people associate with other human brains. Through training based on training data (machine learning), the trained function is able to adapt to new circumstances and detect and extrapolate patterns.

[0047] This allows corresponding algorithms, for example, to be able to take image content into account when denoising, provided they have been trained with suitable training data. Training can be achieved, for example, by including images of the same examination object, such as a phantom or a patient, with different main magnetic fields in the respective training data set. This allows the algorithm to be trained using known supervised learning to reduce higher levels of noise in the measurement data set acquired with a lower field in such a way that the resulting result is as close as possible to the measurement data set acquired with a higher field.Alternatively or additionally, high-quality measurement data can be provided with additional noise to obtain modified measurement data sets, and the algorithm can be trained in such a way that a deviation of the modified measurement data from the high-quality measurement data used as reference data is minimized.

[0048] Approaches to training filter algorithms through machine learning are known per se. These approaches can be extended by having the algorithm to be trained, and thus also the trained algorithm, additionally process the filter parameters determined as explained above. These parameters can be, in particular, the vascularity measures of the input image voxels and / or the value matrices underlying them and / or parameters determined from the aforementioned parameters.

[0049] The algorithm can thus be provided with additional information both during training and during its application after training. This information includes the filter parameters that describe the vascular structure contained in the input image data or the part of the respective training data set used as input image data, in particular the presence of a vascular structure in a specific input image voxel and / or the direction of a vascular structure in the respective voxel. This additional knowledge can significantly speed up and simplify training of the filter algorithm and / or further improve the achieved image quality.

[0050] Generally speaking, parameters of a trained function can be adjusted through training. In particular, supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, and / or active learning can be used. Representation learning (also known as feature learning) can also be used. The parameters of the trained function can be adjusted iteratively through multiple training steps.

[0051] For example, a trained function may include a neural network, a support vector machine (SVM), a decision tree, and / or a Bayesian network, and / or the trained function may be based on k-means clustering, Q-learning, genetic algorithms, and / or mapping rules. In particular, a neural network may be a deep neural network, a convolutional neural network (CNN), or a deep CNN. Furthermore, the neural network may be an adversarial network, a deep adversarial network, and / or a generative adversarial network (GAN).

[0052] As already explained, the vascularity measures of the input image voxels and / or a three-dimensional vascularity map derived from them can be used as filter parameters or as part of the filter parameters for the trained detection algorithm. The filter parameters can thus be used to inform the trained detection algorithm which structures—namely, the vascular structures known from the vascularity measure—should remain unaffected during filtering. Machine learning can thus be used to specifically train the filter algorithm to minimize the disturbance to vascular structures defined by the filter parameters.

[0053] Based on the vascularity measures, unsupervised learning can be implemented, for example, in addition to or as an alternative to supervised learning. For example, during training, a downstream algorithm can evaluate the extent and / or contrast with which vascular structures described by the vascularity measures or the three-dimensional vascularity map are depicted in the filtered image data and / or the extent to which noise or high-frequency signal components outside the vascular structure are suppressed.

[0054] A further possibility is a computer-implemented method for parameterizing a determination algorithm for providing a filter algorithm through machine learning, particularly for use in the above-explained computer-implemented method according to the invention, wherein, for three-dimensional medical input image data used in the training, vascularity measures for at least parts of the input image voxels of the input image data and / or a three-dimensional vascularity map associated with the input image data are used. The determination of the vascularity measures and / or the vascularity map can be carried out as explained above, whereby this determination can be part of the training method or can precede it.

[0055] The computer-implemented method for parameterizing a determination algorithm for providing a filter algorithm, in particular for use in the above-explained computer-implemented method according to the invention, by machine learning, may comprise the following steps: - Specification of several training data sets, each • three-dimensional medical input image data, • describe vessel-likeness measures for at least parts of the input image voxels of the input image data and / or a three-dimensional vessel-likeness map associated with the input image data, wherein the vessel-likeness measure and / or the vessel-likeness map describes a degree of similarity of a structure imaged in the respective input image voxel to an image of a vessel, and • Target result image data, - Minimizing a cost function which depends on a measure of the respective deviation between a result image data set and the target result image data for the plurality of training data sets, by varying several algorithm parameters of the determination algorithm to determine an optimal parameter set for the algorithm parameters, wherein the respective result image data set results from the application of the determination algorithm to respective input data which comprise the three-dimensional medical input image data and, on the one hand, the vascularity measures and / or, on the other hand, the vascularity map of the respective training data set.

[0056] The determination of the vascularity measures or the vascularity map can be performed in a prior step. However, it is also possible to initially provide preliminary training datasets containing the input image data and the target image data. After this, the vascularity measures or the vascularity map are first determined for the respective preliminary training dataset as part of the training process, as already explained above, in order to specify the training datasets.

[0057] The invention further relates to a processing device configured to carry out the inventive computer-implemented method for processing three-dimensional medical input image data and / or to carry out the inventive computer-implemented method for parameterizing a determination algorithm to provide a filter algorithm. The processing device can be configured, for example, as suitably programmed data processing devices, or the aforementioned functionality can alternatively be implemented at least partially in a hard-wired manner. The processing device can be integrated into a medical imaging device, in particular a magnetic resonance imaging scanner, or configured separately therefrom. It can be implemented, for example, as a workstation computer, server, or cloud solution.

[0058] In addition, the invention relates to a computer program with instructions which are designed to carry out the computer-implemented method according to the invention for processing three-dimensional medical input image data and / or the computer-implemented method according to the invention for parameterizing a determination algorithm for providing a filter algorithm when executed on a data processing device.

[0059] Further advantages and details of the invention will become apparent from the following exemplary embodiments and the accompanying drawings. These schematically show: Fig. 1 shows an embodiment of a processing device according to the invention, by means of which an embodiment of the method according to the invention for processing three-dimensional medical input image data is carried out, Fig. 2 a flow chart of the Fig. 1 procedure carried out, Fig. 3 a flowchart of an embodiment of the method for parameterizing a determination algorithm for providing a filter algorithm, and Fig. 4&5 Exemplary designs of detection algorithms that are used by the Fig. 3 are parameterizable and / or can be used as a filter algorithm in the method according to Fig. 2 are usable.

[0060] Fig. Figure 1 shows a processing device 61 used to process three-dimensional medical input image data. In the example, the input image data is acquired directly by a medical imaging device 58, which in the example is a magnetic resonance imaging scanner, and relates to the patient 59. Alternatively, the processed input image data could originate, for example, from a database or other sources.

[0061] The input image data can, for example, be measurement data from a magnetic resonance angiography scan, which was performed, in particular, at low field strengths of less than 1.5 T and without contrast agent. Therefore, the input image data in the example is highly noisy, and the image data is to be filtered by the processing device or the method implemented by it for processing the input image data. The filtered image data can then be displayed on a display device 65, and the display can be parameterized, for example, using the input device 66.

[0062] In the example, the processing device 61 is implemented by a programmable data processing device 60 comprising a processor 62 and a memory 63. The memory 63 stores the computer program 64, whose instructions implement the steps of the method for processing the input image data explained below.

[0063] Fig. 2 shows a flowchart of the method implemented by the processing device 61 for processing the input image data 33. Fig. Figure 2 shows in detail an advantageous embodiment of the process, which is why an overview of the central process steps will first be given for a better understanding.

[0064] First, in step S1, the input image data 33 is obtained. These describe image values ​​34 for a plurality of input image voxels 35. Step S1 can comprise the acquisition of the input image data 33 by the imaging device 58 or simply the reception of the input image data 33 or the reading of the input image data 33 from the memory 63 if these are already stored locally.

[0065] Through steps S2, S3, and S5, a respective vessel-likeness measure 36 is determined for each of the input image voxels 35, which describes a determined degree of similarity of a structure imaged in the respective input image voxel 35 to an image of a vessel. An example of this determination will be explained in more detail later.

[0066] In steps S8 to S10, parameter values ​​41 of filter parameters 42 of a filter algorithm 38 are then specified locally for the respective voxel as a function of the vascularity measures 36, which will also be explained in more detail later.

[0067] In step S11, a filter algorithm 38 is then applied to the input image data 33 to provide filtered image data 39, wherein the filter parameters 42 are specified by the previously determined parameter values ​​41.

[0068] The filter algorithm 38 determines the image value 43 of a respective filtered voxel 44 of the filtered image data 39 depending on the image values ​​34 of several of the input image voxels 35, for example, by using a filter kernel or a diffusion filter. To determine the image values ​​43 of at least two of the filtered voxels 44, different parameter values ​​41 are used for at least one of the filter parameters 42, resulting in the locally different parameterization explained above. This can be implemented, for example, by using locally different filter kernels or a location-dependent convolution kernel, or by using, for example, a diffusion filter with locally different parameterization.

[0069] Alternatively or in addition to the filtering explained, the image values ​​34 of the input image voxels 35 can be scaled in steps S12 and S13 to provide scaled image data 37. For this purpose, in step S12, an associated scaling factor 40 can first be determined for each of the input image voxels 35. The scaling factor 40 can be selected such that it is considerably larger for input image voxels whose vascularity measure 36 indicates membership in the vascular system than for input image voxels 35 for which this is not the case. For example, the scaling factor can be proportional to the vascularity measure 36 and / or selected to be very small if the vascularity measure falls below a threshold value.

[0070] In step S13, the image value 34 of each input image voxel 35 can then be multiplied by the scaling factor 40 determined for this input image voxel 35 to generate scaled image data 37. This highlights the vascular system depicted in the input image data and suppresses a background. The scaled image data 37 determined in this way can also be used in step S11 instead of the input image data 33 as input data for the filter algorithm 38. In an alternative embodiment, a corresponding scaling would also be possible, either additionally or alternatively, for the already filtered voxels 44.

[0071] In the following, the concrete implementation of the steps discussed only cursorily above for the Fig. 2 the exemplary process design is explained in more detail.

[0072] Here, after receiving the input image data 33 in step S1, in step S2 an associated value matrix 47 is first determined for each of the input image voxels 35 of the input image data 33 by calculating a convolution of the Hessian matrix of a Gaussian kernel with the input image data 33. A concrete calculation of the matrix entries I ab has already been presented in the general part of the description.

[0073] In step S3, the eigenvalues ​​48 of the value matrix 47, which are referred to as λ1, λ2, and λ3 in the formulas given in the general section, are then determined, and in step S4, the eigenvectors 50 associated with the eigenvalues ​​48 are determined. Algorithms for determining eigenvalues ​​and eigenvectors of matrices are part of the standard mathematical repertoire and will therefore not be presented here.

[0074] In step S5, the eigenvalues ​​48 are then used to calculate the vessel-likeness measures 36, which are defined above as λ 123be calculated as explained above.

[0075] In step S6, in the example, a subset 45 of the input image voxels 35 is first selected, each of which is part of a vascular structure according to its vascularity measure 36. For example, all input image voxels 35 whose vascularity measure 36 exceeds a predetermined threshold can be included in this subset 45.

[0076] In step S7, for each filtered voxel 44 for which an image value 43 is to be determined in step S11, it is then checked whether its distance or the distance 67 of the input image voxel 35 assigned to it, which is located at the same position as the assigned filtered voxel 44, from the nearest input image voxel that is part of the subset 45 exceeds a predetermined distance limit value 68.

[0077] If this is the case, it can be assumed that a relatively strong filtering in the area of ​​this voxel does not impair the representation of imaged vascular structures, since the voxel is sufficiently far away from voxels of subset 45 and thus from detected vascular-like structures.

[0078] In step S8, for example, fixed parameter values ​​41 for the filter parameters 42 can be used for such voxels, which can, for example, lead to a relatively strong, at least essentially isotropic filtering.

[0079] However, for the subgroup 49 of filtered voxels whose distance 67 is less than or equal to the distance threshold 68, such strong filtering could potentially lead to blurring of vessel boundaries and / or other impairments of the vessel representation, which should be avoided. Therefore, a different filtering parameterization should be used for this subgroup.

[0080] In a simplified variant of the method presented, isotopic filtering could also be carried out for the filtered voxels of the subgroup, whereby a weaker filtering could be achieved by a different choice of the parameter values ​​41 near the vascular structure, particularly depending on the distance 67, for example by lower diffusion constants or a narrower filter kernel.

[0081] In the example shown, however, an anisotropic filter algorithm 38 is used, whose filter parameters 42 include anisotropy parameters 46 that control a degree and a direction of the anisotropy 42 of the filter algorithm 38. As already discussed in the general part, by appropriately selecting the parameter values ​​41 of the anisotropy parameters 46 in step S9, it can be achieved that a weaker filtering occurs for a direction perpendicular to the longitudinal direction of a vessel-like structure than in the longitudinal direction.

[0082] To select a suitable parameterization, in the example, the nearest input image voxel 35 is first selected for the respective filtered voxel 44 of subgroup 49. This voxel belongs to subset 45 and thus appears to depict a vascular structure. If the input image voxel 35 assigned to the filtered voxel 44 is itself a member of the subset, this voxel is selected.

[0083] The eigenvectors 50 of the value matrix of the respective selected input image voxel 35 determined in step S4 define a local longitudinal direction into which the vascular structure apparently extends. Here, the eigenvector 50 associated with the smallest eigenvalue corresponds to the longitudinal direction of the nearest section in the vascular structure, so that relatively strong filtering is possible in the direction specified by this eigenvector, for example, by using a relatively wide filter kernel.

[0084] In the simplest case, by appropriately selecting an anisotropy parameter, the filter strengths can be reduced in both directions perpendicular to the longitudinal direction of the nearest vessel-like structure. The degree of reduction can depend on the distance 67.

[0085] However, the filtering result can potentially be further improved if strong filtering is also carried out perpendicular to the plane defined by the longitudinal direction of the vessel-like structure and the connecting line between the respective filtered voxel or the input image voxel assigned to it and the input image voxel selected for this filtered voxel, since for this direction, for example, no overlap of a filter kernel with the nearest vessel-like structure is to be expected.

[0086] After the explained selection of the parameter values ​​41 of the anisotropy parameters 46 in step S9, parameter values ​​41 for the remaining filter parameters 42 that do not influence the anisotropy can be selected in step S10. This selection can be made, for example, depending on the distance 67 in order to use a weaker filtering overall, regardless of the explained anisotropy near or on vascular structures, than would be achieved by the parameter values ​​41 specified in step S8 for voxels far from the vascular structure.

[0087] In step S11, as already explained, the filter algorithm 38 parameterized by the parameter values ​​41 can then be applied to the input image data 33 or the scaled image data 37 in order to generate the filtered image data 39.

[0088] Through the described procedure, the local parameterization of the filtering is adapted in order to achieve weaker filtering near or on vascular structures that are detected based on the vascularity measure 48, at least in the direction perpendicular to the longitudinal direction of the nearest vascular structure, than that used in areas that are further away from the vascular structure. Thus, in pure background areas that are far away from vascular structures, strong filtering can be used, which, without this attenuation of the filtering depending on the vascularity measure or the consideration of the longitudinal direction of the nearest vascular structure, would significantly blur the boundaries of the vascular structure and would thus be unsuitable for a high-quality representation of vascular structures.Thus, a significantly improved noise suppression, especially in the background area, can be achieved, at least essentially without impairing the sharpness of the representation of the vascular structure.

[0089] It is known per se that, for example, for filters based on a filter kernel and for diffusion filters, the filter behavior can be adapted by specifying parameter values ​​41 for filter parameters 42, especially for anisotropy parameters 46. For example, the extent of a filter kernel can be scaled or diffusion constants can be adjusted. These filter algorithms can thus be easily used in the method according to the invention.

[0090] It is also possible that a detection algorithm 51 trained by machine learning is used as the filter algorithm 38, for which an exemplary training in Fig. 3 is shown schematically. In the example, it is assumed that both the image values ​​34 of the input image voxels 35 and the filter parameters 42 are used as input variables 52 of the trained determination algorithm 51. In the example according to Fig. 2, the vascularity measures 48 of the input image voxels 35 are used directly as filter parameters 42, which can be determined on the basis of the image values ​​34, for example, according to the steps S1 to S3 explained above.

[0091] In step S14, several training data sets 53 are initially specified, on the basis of which training is to take place. The training data sets 53 each comprise three-dimensional medical input image data 33 and target result image data 54. The target result data 54 each specify a desired result, to which a result image data set 56 resulting from the application of the trained determination algorithm 51 and thus the filter algorithm 38 to the input data 52 of the respective training data set should correspond as closely as possible.

[0092] As already explained in the general section, pairs of input image data 33 and target image data 54 can be determined, for example, by using different field strengths of the main field of a magnetic resonance imaging scanner to acquire the input image data 33 and the target image data 54 and / or by artificially degrading the image quality in the input image data 33 compared to the target image data 54, for example, by additive noise. The filter parameters 42 can, as explained, be determined based on the respective input image data 33 and, in the example, correspond to the vascularity measures 36.

[0093] In step S 15, the determination algorithm 51, initially provisionally parameterized by the algorithm parameters 57, is applied to the input data 52, i.e. in the example to the input image data 33 and their vascularity measures 36, in order to determine a result image data set 56.

[0094] In step S16, a cost function 55 is then evaluated, which depends on a measure of the respective deviation of the result image data set 56 from the target result image data 54 for several training data sets 53. For example, the individual deviation measures can be summed.

[0095] By varying the algorithm parameters 57, the cost function 55 is then iteratively minimized to determine an optimal parameter set for the algorithm parameters 57. Known approaches for error feedback can be used in particular. Overall, the example thus implements supervised learning based on the training data sets 53.

[0096] The determination algorithm 51 parameterizing according to the optimal parameter set can then be used as filter algorithm 38.

[0097] Possible structures of detection algorithms that can be used here are described below with additional reference to the Fig. 4 and Fig. 5. For the sake of clarity, neural networks with a relatively small number of nodes are discussed as examples. In real-world applications, higher numbers of nodes and thus parameters are typically used. Other neural network structures or, fundamentally, other machine learning algorithms can also be used, examples of which have already been mentioned in the general section.

[0098] Fig. 4 shows an embodiment of an artificial neural network 1. English terms for the artificial neural network 1 are “artificial neural network”, “neural network”, “artificial neural net” or “neural net”.

[0099] The artificial neural network 1 comprises nodes 6 to 18 and edges 19 to 21, where each edge 19 to 21 is a directed connection from a first node 6 to 18 to a second node 6 to 18. In general, the first node 6 to 18 and the second node 6 to 18 are different nodes 6 to 18, but it is also conceivable that the first node 6 to 18 and the second node 6 to 18 are identical. For example, in Fig. 4 the edge 19 is a directed connection from the node 6 to the node 9 and the edge 21 is a directed connection from the node 16 to the node 18. An edge 19 to 21 from a first node 6 to 18 to a second node 6 to 18 is called an ingoing edge for the second node 6 to 18 and an outgoing edge for the first node 6 to 18.

[0100] In this embodiment, nodes 6 to 18 of artificial neural network 1 can be arranged in layers 2 to 5, wherein the layers can have an intrinsic order introduced by edges 19 to 21 between nodes 6 to 18. In particular, edges 19 to 21 can only be provided between adjacent layers of nodes 6 to 18. In the illustrated embodiment, there is an input layer 2 that only has nodes 6, 7, and 8, each without an incoming edge. Output layer 5 comprises only nodes 17, 18, each without an outgoing edge, with hidden layers 3 and 4 furthermore lying between input layer 2 and output layer 5. In the general case, the number of hidden layers 3, 4 can be chosen arbitrarily.The number of nodes 6, 7, 8 of the input layer 2 usually corresponds to the number of input values ​​to the neural network 1, and the number of nodes 17, 18 in the output layer 5 usually corresponds to the number of output values ​​of the neural network 1.

[0101] In particular, a (real) number can be assigned to nodes 6 to 18 of the neural network 1. Here, x denotes (n) i the value of the i-th node 6 to 18 of the n-th layer 2 to 5. The values ​​of nodes 6, 7, 8 of input layer 2 are equivalent to the input values ​​of neural network 1, while the values ​​of nodes 17, 18 of output layer 5 are equivalent to the output values ​​of neural network 1. In addition, each edge 19, 20, 21 can be assigned a weight in the form of a real number. In particular, the weight is a real number in the interval [-1, 1] or in the interval [0, 1,]. Here, w denotes (m,n)i,j is the weight of the edge between the i-th nodes 6 to 18 of the m-th layer 2 to 5 and the j-th nodes 6 to 18 of the n-th layer 2 to 5. Furthermore, the abbreviation wi,j(n) for the weight wi,j(n,n+1) defined.

[0102] To calculate output values ​​of neural network 1, the input values ​​are propagated through neural network 1. In particular, the values ​​of nodes 6 to 18 of the (n+1)-th layer 2 to 5 can be calculated based on the values ​​of nodes 6 to 18 of the n-th layer 2 to 5 by xj(n+1)=f(∑ixi(n)⋅wi,j(n)).

[0103] Here, f is a transfer function, which can also be referred to as an activation function. Common transfer functions include step functions, sigmoid functions (e.g., the logistic function, the generalized logistic function, the hyperbolic tangent, the arctangent, the error function, the smoothstep function), or rectifier functions. The transfer function is primarily used for normalization purposes.

[0104] Specifically, the values ​​are propagated layer by layer through neural network 1, with values ​​of input layer 2 being given by the input data of neural network 1. Values ​​of the first hidden layer 3 can be calculated based on the values ​​of input layer 2 of neural network 1, values ​​of the second hidden layer 4 can be calculated based on the values ​​in the first hidden layer 3, and so on.

[0105] To the values wi,j(n) for edges 19 to 21, neural network 1 must be trained using training data. In particular, training data includes training input data and training output data, which are referred to below as t i For a training step, neural network 1 is applied to the training input data to determine calculated output data. Specifically, the training output data and the calculated output data comprise a number of values, where the number is determined as the number of nodes 17, 18 of output layer 5.

[0106] In particular, a comparison between the calculated output data and the training output data is used to recursively adjust the weights within the neural network 1 (back propagation algorithm). In particular, the weights can be adjusted according to w′i,j(n)=wi,j(n)−γ⋅δj(n)⋅xi(n) where γ is a learning rate and the numbers δj(n) can be calculated recursively as δj(n)=(∑kδk(n+1)⋅wj,k(n+1))⋅f′(∑ixi(n)⋅wi,j(n)) based on δj(n+1), if the (n+1)-th layer is not the output layer 5, and δj(n)=(xk(n+1)−tj(n+1))⋅f'(∑ixi(n)⋅wi,j(n)) if the (n+1)-th layer is the output layer 5, where f' is the first derivative of the activation function and yj(n+1) is the comparison training value for the j-th node 17, 18 of the output layer 5.

[0107] In the following, with regard to Fig. Figure 5 also provides an example of a convolutional neural network (CNN). Note that the term "layer" is used in a slightly different way than for classical neural networks. For a classical neural network, the term "layer" refers only to the set of nodes that form a layer, i.e., a specific generation of nodes. For a convolutional neural network, the term "layer" is often used to refer to an object that actively modifies data—in other words, a set of nodes of the same generation and either the set of incoming or outgoing edges.

[0108] Fig.5 shows an embodiment of a convolutional neural network 22. In the illustrated embodiment, the convolutional neural network 22 comprises an input layer 23, a convolutional layer 24, a pooling layer 25, a fully connected layer 26, and an output layer 27. In alternative embodiments, the convolutional neural network 22 may contain multiple convolutional layers 24, multiple pooling layers 25, and multiple fully connected layers 26, as well as other types of layers. The order of the layers can be chosen arbitrarily, with fully connected layers 26 typically forming the last layers before the output layer 27.

[0109] In particular, within a convolutional neural network 22, the nodes 28 to 32 of one of the layers 23 to 27 can be understood as being arranged in a d-dimensional matrix or as a d-dimensional image. In particular, in the two-dimensional case, the value of a node 28 to 32 with the indices i, j in the n-th layer 23 to 27 can be represented as x (n) [i,j]. It should be noted that the arrangement of nodes 28 to 31 of a layer 23 to 27 has no effect on the calculations within the convolutional neural network 22 as such, since these effects are determined exclusively by the structure and weights of the edges.

[0110] A convolutional layer 24 is particularly characterized in that the structure and weights of the incoming edges form a convolution operation based on a specific number of kernels. In particular, the structure and weights of the incoming edges can be chosen such that the values xk(n) the node 29 of the convolutional layer 24 as a convolution xk(n)=Kk*x(n−1) based on the values ​​x (n-1) the node 28 of the previous layer 23 can be determined, where the convolution * in the two-dimensional case can be defined as xk(n)[i,j]=(Kk*x(n−1))[i,j]=∑i'∑j'Kk[i',j']⋅x(n−1)[i−i', j−j'].

[0111] In it, the k-th kernel K ka d-dimensional matrix, in this embodiment a two-dimensional matrix, which is usually small compared to the number of nodes 28 to 32, for example a 3x3 matrix or a 5x5 matrix. In particular, this implies that the weights of the incoming edges are not independent, but are chosen to generate the above convolution equation. In the example for a kernel forming a 3x3 matrix, only nine independent weights exist (where each entry of the kernel matrix corresponds to an independent weight), regardless of the number of nodes 28 to 32 in the corresponding layer 23 to 27. In particular, for a convolutional layer 24, the number of nodes 29 in the convolutional layer 24 is equivalent to the number of nodes 28 in the preceding layer 23 multiplied by the number of convolution kernels.

[0112] If the nodes 28 of the preceding layer 23 are arranged as a d-dimensional matrix, the use of the plurality of kernels can be understood as the addition of a further dimension, also referred to as the depth dimension, so that the nodes 29 of the convolutional layer 24 are arranged as a (d+1)-dimensional matrix. If the nodes 28 of the preceding layer 23 are already arranged as a (d+1)-dimensional matrix with a depth dimension, the use of a plurality of convolutional kernels can be understood as an expansion along the depth dimension, so that the nodes 29 of the convolutional layer 24 are likewise arranged as a (d+1)-dimensional matrix, wherein the size of the (d+1)-dimensional matrix in the depth dimension is larger than in the preceding layer 23 by the factor formed by the number of kernels.

[0113] The advantage of using convolutional layers 24 is that the spatially local correlation of the input data can be exploited by creating a local connection pattern between nodes of adjacent layers, in particular by each node having connections only to a small range of the nodes of the previous layer.

[0114] In the illustrated embodiment, the input layer 23 comprises thirty-six nodes 28 arranged as a two-dimensional 6x6 matrix. The convolutional layer 24 comprises seventy-two nodes 29 arranged as two two-dimensional 6x6 matrices, each of which is the result of convolving the values ​​of the input layer 23 with a convolution kernel. Similarly, the nodes 29 of the convolutional layer 24 can be understood as being arranged in a three-dimensional 6x6x2 matrix, with the latter dimension being the depth dimension.

[0115] A pooling layer 25 is characterized in that the structure and weights of the incoming edges as well as the activation function of its nodes 30 define a pooling operation based on a nonlinear pooling function f. For example, in the two-dimensional case, the values ​​x (n) the node 30 of the pooling layer 25 based on the values ​​x (n+1) the node 29 of the previous layer 24 as x(n)[i,j]=f(x(n−1)[id1, jd2], …, x(n−1)[id1+d1−1, jd2+d2−1]) In other words, by using a pooling layer 25, the number of nodes 29, 30 can be reduced by replacing a number of d1 × d2 neighboring nodes 29 in the preceding layer 24 with a single node 30, which is calculated as a function of the values ​​of said number of neighboring nodes 29. In particular, the pooling function f can be a maximum function, an averaging function, or the L2 norm. In particular, for a pooling layer 25, the weights of the incoming edges can be fixed and not modified by training.

[0116] The advantage of using a pooling layer 25 is that the number of nodes 29, 30 and the number of parameters are reduced. This leads to a reduction in the amount of computation required within the convolutional neural network 22 and thus to a control of overfitting.

[0117] In the illustrated embodiment, pooling layer 25 is a max-pooling layer in which four neighboring nodes are replaced with a single node whose value is the maximum of the values ​​of the four neighboring nodes. Max-pooling is applied to each d-dimensional matrix of the previous layer; in this embodiment, max-pooling is applied to each of the two two-dimensional matrices, reducing the number of nodes from seventy-two to eighteen.

[0118] A fully connected layer 26 is characterized in that a majority, in particular all, of the edges are present between the nodes 30 of the previous layer 25 and the nodes 31 of the fully connected layer 26, wherein the weight of each of the edges can be individually adjusted. In this embodiment, the nodes 30 of the previous layer 25 and the fully connected layer 26 are shown both as two-dimensional matrices and as non-connected nodes (shown as a row of nodes, with the number of nodes reduced for clarity). In this embodiment, the number of nodes 31 in the fully connected layer 26 is equal to the number of nodes 30 in the previous layer 25. In alternative embodiments, the number of nodes 30, 31 may be different.

[0119] Furthermore, in this embodiment, the values ​​of the nodes 32 of the output layer 27 are determined by applying the softmax function to the values ​​of the nodes 31 of the preceding layer 26. By applying the softmax function, the sum of the values ​​of all nodes 32 of the output layer 27 is one, and all values ​​of all nodes 32 of the output layer are real numbers between 0 and 1. When the convolutional neural network 22 is used to classify input data, the values ​​of the output layer 27, in particular, can be interpreted as the probability that the input data falls into one of the different classes.

[0120] A convolutional neural network 22 can also have a ReLU layer, where ReLU is an acronym for "rectified linear units." In particular, the number of nodes and the structure of the nodes within a ReLU layer are equivalent to the number of nodes and the structure of the nodes in the previous layer. The value of each node in the ReLU layer can be calculated, in particular, by applying a rectifier function to the value of the corresponding node in the previous layer. Examples of rectifier functions are f(x)=max(0,x), the hyperbolic tangent, or the sigmoid function.

[0121] Convolutional neural networks 22 can be trained, in particular, based on the backpropagation algorithm. To avoid overfitting, regularization methods can be used, such as dropout of individual nodes 28 to 32, stochastic pooling, use of artificial data, weight decay based on the L1 or L2 norm, or maximum norm constraints.

[0122] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.

Claims

[1] Computer-implemented method for processing three-dimensional medical input image data (33), comprising the steps of: - Obtaining the input image data (33) describing the image values ​​(34) for a plurality of input image voxels (35), - determining a respective vessel-likeness measure (36) for at least parts of the input image voxels (35), wherein the vessel-likeness measure (36) describes a determined degree of similarity of a structure imaged in the respective input image voxels (35) to an image of a vessel, and - applying a filter algorithm (38) to the input image data (33) to provide filtered image data (39), wherein the parameter value (41) of at least one filter parameter (42) of the filter algorithm (38) is predetermined as a function of at least one of the vascularity measures (36), wherein the filter algorithm (38) is an anisotropic filter algorithm (38), the filter parameter (42) of which is an anisotropy parameter (46) or the filter parameters (42) of which comprise at least one anisotropy parameter (46), wherein the respective anisotropy parameter controls a degree and / or a direction of the anisotropy (42) of the filter algorithm (38), so that for at least some of the possible parameter values ​​(41) of the anisotropy parameter (46), the filter algorithm (38) has a different filter behavior for two spatial directions, wherein each filtered voxel (44) is assigned one of the Input image voxel (35) is permanently assigned,wherein the input image voxel (35) selected for determining the anisotropy parameter (46) of the respective filtered voxel (44) is selected from a subset (45) of the input image voxels (35) selected on the basis of the vascularity measures (36), depending on the distance (67) of the respective input image voxel (35) of the subset (45) to the input image voxel (35) assigned to the respective filtered voxel (44). [2] Computer-implemented method according to claim 1, characterized by in that the filter algorithm (38) determines the image value (43) of a respective filtered voxel (44) of the filtered image data (39) in each case as a function of the image values ​​(34) of a plurality of the input image voxels (35), wherein to determine the image values ​​(43) of at least two of the filtered voxels (44) different parameter values ​​(41) for the filter parameter (42) or at least one of the filter parameters (42) are used. [3] Computer-implemented method according to claim 2, characterized by that a subset (45) of the input image voxels (35) which are each part of a vessel-like structure according to their vascularity measure (36) is selected, wherein the parameter value (41) of the filter parameter (42) or at least one of the filter parameters (42) for the respective filtered voxel (44) is selected depending on whether an input image voxel (42) assigned to the respective filtered voxel (44) is an element of the subset (45) and / or how far the assigned input image voxel (35) is from the nearest input image voxel (35) which is an element of the subset (45). [4] Computer-implemented method according to one of the preceding claims, characterized byin that, in order to calculate the respective vascularity measure (36), a value matrix (47) assigned to the respective input image voxel (35) is determined as an intermediate step, wherein the respective vascularity measure (36) is determined from eigenvalues ​​(48) of the respective value matrix (47). [5] Computer-implemented method according to one of the preceding claims, characterized by that the value matrix (47) is determined by convolving the Hesse matrix of a Gaussian kernel or the Hesse matrix of another filter kernel with the input image data (33). [6] Computer-implemented method according to one of claims 4 or 5, characterized byin that to determine the image values ​​(43) of all filtered voxels (44) of the filtered image data (39) or of a subgroup (49) of the filtered voxels (44), a parameter value (41) of the anisotropy parameter (46) or at least one of the anisotropy parameters (46) is used, which is determined in each case as a function of at least one eigenvector (50) of that value matrix (47) which is assigned to a respective selected one of the input image voxels (35). [7] Computer-implemented method according to one of the preceding claims, characterized bythat the subgroup (49) of the filtered voxels (44) for which the parameter value (41) of the anisotropy parameter (46) is determined as a function of the respective eigenvector (50) comprises exclusively filtered voxels (44) whose respective associated input image voxel (35) is either itself an element of the selected subset (45) of the input image voxels (35) or has a distance (67) to at least one input image voxel (35) of this subset (45) which is smaller than or corresponds to a predetermined distance limit value (68). [8] Computer-implemented method according to one of the preceding claims, characterized by that the three-dimensional medical input image data (33) are based on a magnetic resonance angiography without a contrast agent administration and / or with a field strength of the main magnetic field of less than 1.5 T or less than 1 T. [9] Computer-implemented method according to one of the preceding claims, characterized bythat the image values ​​(43) of the filtered voxels (44) of the filtered image data (39) are scaled in order to provide scaled image data (37), wherein a respective scaling factor (40) used to scale the respective filtered voxel (44) is predetermined as a function of at least one of the vascularity measures (36), or that a filter algorithm (38) is applied to the scaled image data (37) in order to provide filtered image data (39), wherein the parameter value (41) of at least one filter parameter (42) of the filter algorithm (38) is predetermined as a function of at least one of the vascularity measures (36). [10] Computer-implemented method according to one of the preceding claims, characterized bythat a determination algorithm (51) trained by machine learning is used as the filter algorithm (38), wherein both the image values ​​(34) of the input image voxels (35) and the filter parameters (42) are used as input variables (52) of the trained determination algorithm (51). [11] Computer-implemented method according to claim 10, characterized by that the vascularity measures (48) of the input image voxels (35) and / or a three-dimensional vascularity map determined from these are used as filter parameters (42) or as part of the filter parameters (42) for the trained determination algorithm (51). [12] Processing device, characterized by that it is designed to carry out the computer-implemented method according to one of the preceding claims. [13] Computer program with instructions which are designed to carry out the computer-implemented method according to one of claims 1 to 11 when executed on a data processing device (60).

Images