Method for determining subtraction images, image recording device, program and data carrier

Abstract

The invention relates to a method for determining a subtraction image (13) from a feature image (2) of an imaging region of an examination object and a plurality of mask images (3), in which a combination of the mask images (3) weighted by mask weights (10) is subtracted from the feature image (2) in order to determine the subtraction image (13), the mask weight (10) is determined in an optimization process for optimizing an image quality function of the test image, the test image is calculated by subtracting from the feature image (2, 2 ') a combination of the mask images (3, 3') determined by the test group of mask weights, characterized in that the form of convex quadratic programming of the dimension of the image quality function corresponding to the number of mask images (3, 3 ') is minimized.

Description

Technical Field

[0001] The invention relates to a computer-implemented method for determining a subtraction image from a feature image of an imaged region of an object to be examined and a plurality of mask images, wherein a combination of mask images weighted by mask weights is subtracted from the feature image to determine the subtraction image, wherein the mask weights are determined in an optimization process for optimizing an image quality function for a test image, the test image being calculated by subtracting the combination of mask images determined by a test set of mask weights from the feature image. The invention also relates to an image recording device, a computer program and an electronically readable data carrier. Background Art

[0002] In medical imaging, there are numerous fields of application in which, in order to improve the display of certain features, a subtracted image is obtained by subtracting a mask image, which contains, for example, the anatomical background, from a feature image showing the image to be displayed. A major known field of application is digital subtraction angiography, in which an imaging region of an object to be examined, in particular a patient, is recorded without the presence of a contrast agent in order to obtain a mask image, while a feature image (also called a fill image) shows the imaging region with contrast agent. The contrast agent is introduced into the patient's vascular system so that the subtracted image shows the patient's blood vessels.

[0003] In this imaging method, the movement of the object to be examined, especially the patient, may occur between the acquisition of the mask image and the acquisition of the characteristic image. Such movement includes, for example, macroscopic, conscious patient movements and periodic physiological activities, especially those caused by heartbeat and / or breathing. As a result, the image content between the mask image and the characteristic image may shift, which may result in errors in the subtraction image. Another problem is the difference in noise between the mask image and the characteristic image. This may lead to noise amplification effects, etc.

[0004] In order to counteract these problems, it has been proposed to record a series of mask images instead of just one. To achieve noise reduction, it can be provided that a plurality of mask images are averaged, but this does not work if the object under examination is moving. It is also provided that an optimized, weighted combination of mask images is determined and subtracted from the feature image, in particular in the case of motion in the mask images, the combination being formed by a sum of mask weights.

[0005] Document DE 10 2021 208 272 A1 discloses a method for generating a subtraction image of digital subtraction angiography. A mask image is obtained here, which shows the natural state of the object when the contrast agent has not yet diffused in the object, especially in the blood vessels. The mask image contains a contour that specifically reflects the structure of the object, but the contour is not meaningful or has little significance for the examination. After the contrast agent is administered to the object under examination, a view (characteristic image) of the object under examination is obtained, which reflects the effect of the diffused contrast agent. It is provided here that a first full image is formed by multiple mask images, in which the multiple mask images are multiplied by independent mask weights and then added, wherein the independent mask weights are automatically determined for each of the multiple mask images by an optimization method. The optimization method uses a suitable quality metric, that is, an image quality function. The quality metric can be energy, a total variable or a softplus function.

[0006] In particular, these methods may include using "interior point" algorithms to solve optimization problems (e.g., see Richard A. Waltz et al., "An interior algorithm for nonlinear optimization that combines line search and trust region steps", Mathematical Programming, 107.3 (2006): 391-408). In general, such algorithms are not real-time, as is desired in fluorescence microscopy applications. In addition, the available algorithms also involve complex implementations, which require huge R&D costs for the corresponding products. Summary of the invention

[0007] Therefore, the technical problem to be solved by the present invention is to provide a robust and efficient solution, in particular a solution applicable in real time, for determining optimized mask weights.

[0008] This object is achieved according to the invention by a computer-implemented method, an image recording device, a computer program and an electronically readable data carrier according to the independent claims. Advantageous embodiments are given by the dependent claims.

[0009] According to the invention, in a computer-implemented method of the type mentioned at the outset, it is provided that the image quality function is minimized in the form of a convex quadratic program of a dimension corresponding to the number of mask images.

[0010] In this case, the examination object can be, in particular, a human or animal patient or an examination model. A plurality of mask images and a mask image should be recorded using a medical image recording device, in particular an X-ray machine and / or a magnetic resonance device. In this case, the mask image does not show features that are shown in the subtraction image, such as contrast agents. The mask image can be recorded using the image recording device before or after the feature image, in particular when no features are present in the imaging region of the examination object. The feature image, the mask image and the subtraction image can be two-dimensional or three-dimensional. The method should be implemented using a control device that controls the image recording device, which can also control the image recording of the mask image and the feature image. In this case, the recording process for the feature image and the mask image does not necessarily constitute a necessary part of the method; at least it can be provided that the mask image and the feature image are provided in the control device (otherwise in the computing device that implements the method) via at least one interface.

[0011] In a particularly advantageous embodiment of the invention, it can be provided that the method is carried out within the scope of digital subtraction angiography. It can then be provided that the characteristic image is a fill image of a digital subtraction angiography with contrast agent. The mask image can be recorded when no contrast agent is present in the imaging region, in particular before the administration of contrast agent. The characteristic image is recorded after the administration of contrast agent, wherein the administration of contrast agent does not form part of the method according to the invention.

[0012] In this case, the mask image is a filling image after the contrast agent has been previously administered. In other words, the filling image shows the contrast agent in the imaging area, which is to be extracted as a feature by subtraction. In this case, the contrast agent is injected into the patient, preferably into the vascular system, so that the blood vessels filled with the contrast agent can be particularly clearly identified in the subtraction image. In particular, a method for determining a subtraction image of digital subtraction angiography from a feature image of an imaging area of ​​an examination object and a plurality of mask images is provided, wherein the feature image is constructed as a filling image, wherein a combination of mask images weighted by mask weights is subtracted from the feature image to determine the subtraction image, wherein the mask weights are determined in an optimization process for optimizing an image quality function of a test image, wherein the test image is calculated by subtracting a combination of mask images determined by a test set of mask weights from the feature image, wherein the image quality function is minimized in the form of a convex quadratic program of a dimension corresponding to the number of mask images.

[0013] In general, an optimization process is also used within the scope of the invention to determine the mask weights in order to determine the sum image, i.e. the combination of mask images. The optimization process is ultimately used to find the best balance between noise reduction and motion artifact reduction by optimizing an image quality indicator (described by an image quality function). For a test set of mask weights, a test image is calculated by subtracting the sum image using the test set from the feature image, and the image quality function is calculated for each of the test images. The result is minimized or maximized depending on the image quality indicator. In this way, temporal interpolation between mask images is also implicitly achieved within the scope of reducing motion artifacts.

[0014] However, in contrast to the prior art, it is currently provided that the optimization problem is transformed into a quadratic program, which allows the amount of calculation to be reduced by a factor of 100 to 1000 by reducing the dimensionality to the number of mask images. At the same time, a robust determination of the optimized mask weights is also achieved. In other words, a more efficient way to solve the problem of mask image weights is provided. The optimization problem is transformed into a low-dimensional program, which can be solved with less calculation. There are a variety of freely available solutions for the program, which significantly reduces the workload for the specific implementation. In particular, the establishment of a general algorithm that is efficient in terms of computing time for solving nonlinear constraint problems is avoided. Experiments have shown that the method according to the invention allows the determination of mask weights in a robust manner and method, with an extremely low real-time compatible computing method, for example on a laptop as a computing device.

[0015] In a particularly preferred embodiment of the invention, a weighted L2 norm is used as the image quality function. The L2 norm of the individual test images is minimized in the optimization process. The quadratic program is formulated by multiplying the subtraction terms of the test images and separating the quadratic and linear terms in the vector of mask weights.

[0016] In the following, let P be the number of pixels in the mask image M, feature image F, and subtraction image S, and let N be the number of mask images. In addition, let s, f, and m k are respectively the vectorized subtraction image, feature image and mask image (where P is a real value, i.e., image value). The mask weight of the i-th mask image is set to α i , where each mask weight lies in the interval from 0 to 1 and the sum of the mask weights is set to 1. Furthermore, let W be the weight matrix, which is the identity matrix in the case of the unweighted L2 norm and a diagonal matrix of dimension PxP in the case of the weighted L2 norm, whose diagonal contains the corresponding standard weights w i (For each pixel i). Then the optimization problem can be described as

[0017]

[0018] Furthermore, let M be a PxN matrix containing the vectorized mask image m i As a column, M = (m1...m N ), and α is the mask weight α with M entries i The vector is as follows:

[0019]

[0020] And among them

[0021]

[0022] The last two planning

[0023] α=argminα T Qα+c T α st∑ i α i =1 0≤α i ≤1 (5)

[0024] The quadratic program then only includes matrices and numbers of dimension N, i.e. the number of mask images. No original image of dimension P needs to be evaluated during the optimization, where N is much smaller than P, i.e. N<<P. Only when Q and c are calculated is the full image processed. The contrast Q is defined as (semi-) positive definite, making the optimization problem convex.

[0025] For this type of convex quadratic optimization problem, there are some special algorithms that are faster than general constrained nonlinear optimization algorithms. In a suitable design of the present invention, the Goldfarb-Idnani method can be used for optimization. In particular, the "Golden Idnani active set dual method" can be used, for example, this method is available in a toolbox called QuadProg. D. Goldfarb and A. Idnani's article "A numerically stable dual method for solving strictly convex quadratic programs", Mathematical Programming 27.1 (1983): 1-33 gives a basic description of the Golden Idnani method. In addition, for semi-positive definite problems, please refer to NL Boland's article: "A dual-active-set algorithm for positive semi-definite quadratic programming", Mathematical Programming 78 (1996): 1-27.

[0026] However, compared to robust image quality functions such as total variation, the quadratic norm is more susceptible to outliers. For example, pixels that have no attenuation during X-ray imaging or pixels where metal objects are located can have a large impact on the optimization. In particular, if the movement pattern of these image points is inconsistent with the general movement pattern (for example, in skull images, the movement of metal dental implants is independent of the skull), it will affect the robustness of the optimized mask weight determination.

[0027] Therefore, it is particularly preferred within the scope of the present invention that the L2 norm is weighted for each pixel by means of a standard weight, wherein the standard weight is determined by an image analysis process of the feature image and the mask image and / or based on additional information. According to the above formula, this means that the matrix W does not correspond to the identity matrix, but W=diag(w) with the P standard weights w in the vector w i Therefore, it is also provided that a suitable standard weight w is determined and used to reduce the influence of outlier pixels and / or focus on certain pixels or regions. Here, with regard to outliers, the standard weight is preferably automatically determined by image analysis during the image analysis process, so that the robustness can be significantly improved.

[0028] To this end, it can be specified that during the image analysis

[0029] - determining for each pixel the maximum deviation of the image value of the feature image from the image value of the mask image, and

[0030] - Assigning a smaller annotation weight to at least a portion of the image points with the largest deviation.

[0031] Since the convex combination of the image is optimized, the maximum impact of each subtracted pixel on the quadratic error term can be calculated. Specifically, for the maximum error of each pixel p, for example, each pixel in the two-dimensional image is equivalent to the image value f(p) of the feature image and the image value m of the mask image. i (p) maximum deviation, where

[0032] e max (p) = max((f(p) - m1(p)) 2 ,…,(f(p)-m N (p)) 2 ) (6)

[0033] Since we want to reduce the impact of outliers, it is recommended to reduce the weights for larger maximum deviations.

[0034] Specifically, it can be provided that a standard weight of zero is configured for a pixel whose maximum deviation is higher than a threshold, and / or that the standard weight is determined according to a monotonically decreasing function within a function range extending at least around the threshold. It is relatively conceivable here that the threshold is fixed, but a preferred embodiment of the present invention provides that the threshold is determined based on a histogram analysis of the maximum deviation, in particular the value of the maximum deviation that exceeds a predetermined proportion. Therefore, a threshold can be derived by analyzing the histogram to identify outlier image values ​​of certain pixels. For example, the threshold can correspond to a maximum deviation value, and 90% to 99% (for example 95%) of the maximum deviation values ​​are lower than this value. The standard weight can be configured as follows:

[0035]

[0036] Therein, t denotes the threshold value. On the other hand, it can be provided that a continuous monotonically decreasing function (weighting function) determines the standard weight around the threshold value.

[0037] However, an advantageous embodiment of the invention can also provide that the standard weights are determined during the image analysis at least partially using an image analysis function that has been trained in particular. This type of image analysis function can be used to detect and / or segment metal and / or non-attenuation areas. Metal areas and / or object-free areas can then be suppressed, for example, by selecting a low standard weight or a standard weight of zero, thereby achieving an optimization.

[0038] In particular, trained image analysis functions are used here, which have been trained by machine learning. In particular, trained image analysis functions have been proposed in the prior art for detecting and / or segmenting metals in image data, which can also be used well in the present invention.

[0039] Generally speaking, trained functions map the cognitive functions that humans associate with other human brains. Through training based on training data (machine learning), trained functions have the ability to adapt to new environments and can discover and infer patterns.

[0040] Generally speaking, the parameters of the training function can be adjusted by training. In particular, supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning and / or active learning can be used. In addition, representation learning (also called "feature learning") can also be used. In particular, the parameters of the training function can be repeatedly adjusted through multiple training steps.

[0041] For example, the training function may include a neural network, a support vector machine (SVM), a decision tree and / or a Bayesian network, and / or the training function may be based on mean clustering, Q-learning, a genetic algorithm and / or an association rule. In particular, the neural network may be a deep neural network, a convolutional neural network (CNN) or a deep CNN. In addition, the neural network may also be an adversarial network, a deep adversarial network and / or a generative adversarial network (GAN).

[0042] In terms of additional information, it can be provided that this additional information describes the regions of interest of the imaging area that should be weighted more highly. In this way, the optimization can be focused on the regions of real interest, wherein less interesting parts of the imaging area can then be weighted less. The additional information can be obtained from user input, for example, the user can mark the regions of interest. However, automatic derivation is also conceivable, for example based on existing information such as the selection and position of the imaging area and / or the imaging target. Trained derivation functions, i.e. artificial intelligence, can also be used for automatic derivation.

[0043] In a generally particularly preferred improvement, it can be provided that the feature image and the mask image are preprocessed by an edge-sharpening image processing operation for the optimization process. As an alternative or supplement to the use of weighting, the feature image and the mask image can also be preprocessed for the optimization process. Here, the edge should be strengthened so as to emphasize the edge during the optimization process and thereby reduce edge artifacts, especially when the object under examination moves. This is particularly advantageous when shooting a patient's head image, because in this way, the optimization process can better focus on the bone edge. Here, an edge-sharpening filter known in principle can be used, but it is particularly preferred that a gradient image is determined in an image processing operation for each of the feature image and a plurality of mask images, the edge-sharpened feature image is determined by a weighted linear combination of the original feature image and the assigned gradient image, and the edge-sharpened individual mask images are determined by a weighted linear combination of the original individual mask images and the assigned gradient image. That is, whether for the feature image or for the mask image, the corresponding image is mixed with the assigned gradient image, which achieves better highlighting of the edge. Due to the improved consideration of motion, the quality of the subtracted image is improved by taking edges into account.

[0044] An advantageous embodiment of the invention is based on the recognition that the weighting carried out using the above-described method can also lead to an increase in the noise level in individual cases without significantly reducing motion artifacts. This is explained in more detail below.

[0045] when represents the noise variance in the unsubtracted individual views, then the noise variance in the averaged mask is equivalent to

[0046]

[0047] And the noise variance in the optimized weights is equivalent to

[0048]

[0049] The subtraction image is obtained

[0050]

[0051] The noise growth in the silhouette image is correspondingly

[0052]

[0053] For example, in the case of 8 mask images (M=8), when each mask is given a weight of 1, the noise in the unfavorable case is increased by

[0054]

[0055] According to an advantageous embodiment of the invention, in order to avoid this problem, it is provided that the image quality function includes a regularization of the weights, the purpose of which is to eliminate an uneven distribution of the weights.

[0056] According to a further advantageous embodiment of the invention, the regularization is performed as a function of the Euclidean norm of the weights.

[0057] Since the noise increase is directly related to Proportional, Can be used as a regularization term to improve the known noise. For this purpose, reference can be made to the Tikhonov regularization known in linear regression, which is known from the publication of Hansen PC "The L-curve and its use in the numerical treatment of inverse problems". Department of Mathematical Modelling, Technical University of Denmark, 1999:

[0058]

[0059] For optimization, appropriate parameters must be chosen Since the energy in the projected image varies with the dose, and therefore with the range of Q values, the Euclidean norm of Q is chosen to correspond to the range, or for complex or real matrices, the Frobenius norm of Q is chosen:

[0060]

[0061] Stable parameters for large datasets It can be determined by L-curve analysis known from the publication of Hansen PC cited above. In one example case, L-curve analysis averaged over 14,000 subtracted images from 1,000 DSA data sets showed that the parameter λ=10 -4 In this example, using λ=10 -4 In the regularization of , the initial level of noise is approximately reached in the optimization. According to

[0062] (92 2 / 78 2 -1)*100%=39%

[0063] Using this regularization reduces the noise increase to 2.5%, which is consistent with the estimation based on Equation 11, which gives a reduction from 36% to 3%.

[0064] In principle, the method of the invention can be implemented by a common computing device when the feature image and the mask image are provided. However, it is particularly suitable for such a computing device to be particularly suitable as a control device for an image recording device, and the feature image and the mask image are recorded by the image recording device.

[0065] In addition to the method, the present invention also relates to an image recording device having a control device, the image recording device comprising:

[0066] - a photographing unit for controlling the photographing of a plurality of mask images and feature images of an imaging region of the inspection object,

[0067] - a determination unit for determining a silhouette image by subtracting a combination of the mask images weighted by the mask weights from the feature images, and

[0068] - an optimization unit for determining mask weights in an optimization process for optimizing an image quality function of a test image, the test image being calculated by subtracting a combination of the mask image determined by a test set of mask weights from the feature image,

[0069] It is characterized in that the optimization unit is designed to optimize the image quality function in the form of a convex quadratic programming of a dimension corresponding to the number of mask images.

[0070] The control device preferably has a processor and a memory and is designed to carry out the method according to the invention. The embodiments of the method according to the invention can be transferred to the image capture device according to the invention in an analogous manner, with which the above-mentioned advantages can also be achieved. The functional units of the control device can be implemented by hardware and / or software. Other functional units for implementing other steps can also be implemented, for example a standard weight determination unit for determining standard weights and / or a preprocessing unit for edge sharpening image processing.

[0071] The image recording device can preferably be an X-ray machine, such as a computed tomography scanner or a C-arm X-ray machine. However, it is also conceivable to equip other image recording devices, such as a magnetic resonance device, accordingly.

[0072] The computer program according to the invention can be directly loaded into a memory of a computing device, in particular a control device of an image recording device, and has program means which, when the computer program is executed on the computing device, causes the computing device to carry out the method according to the invention. The computer program can be stored on an electronically readable data carrier according to the invention, which thus includes control information stored thereon, which includes at least one computer program according to the invention and is designed to enable the computing device to carry out the method according to the invention when the data carrier is used in a computing device, in particular a control device of an image recording device. The data carrier is preferably an untrained data carrier. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Further advantages and details of the invention are apparent from the exemplary embodiments described below and with reference to the accompanying drawings. In the drawings:

[0074] Figure 1 A flow chart showing an embodiment of the method according to the present invention,

[0075] Figure 2 A schematic diagram is shown for explaining the process measures of the method according to the invention.

[0076] Figure 3 A schematic diagram of the principle of the image capture device according to the present invention is shown, and

[0077] Figure 4 A functional block diagram of a control device for an image recording device is shown. DETAILED DESCRIPTION

[0078] Figure 1A flow chart showing an embodiment of the method according to the invention is shown, wherein also the image recording step on standby is shown. Here, an embodiment of digital subtraction angiography using an X-ray machine is described, in particular for real-time applications, such as fluoroscopy and / or other applications, in which the flow of contrast agent in the imaging region of the examination object (here a patient) can be observed at least substantially in real time. In particular, the inflow and outflow of contrast agent can also be observed.

[0079] In step S1, a plurality of mask images of an imaging region of the examination object are recorded without contrast medium using an image recording device, here an X-ray machine. For example, a series of mask images can be recorded in at least one breathing cycle and / or at least one cardiac cycle of the patient, depending on the field of application and the activities that may occur.

[0080] In step S2 , a characteristic image, here a filling image, of the imaging region is recorded by means of an image recording device, specifically at a time when the contrast agent is in the imaging region, in particular in a blood vessel in the imaging region.

[0081] It should be noted that steps S1 and S2 can also be performed in reverse order, which means that, for example, the mask image can be recorded after the contrast agent has disappeared from the imaging region. It is possible to record the mask image both before and after the characteristic image.

[0082] In optional step S3, for the subsequent optimization process, preprocessing of the feature image and multiple mask images is implemented. Here, the mask image and the feature image undergo an edge-sharpening image processing operation in order to determine an edge-sharpened mask image and an edge-sharpened feature image. It should be noted here that the edge-sharpened image is only determined for the optimization process, but is not used for the actual determination of the subtraction image to be performed later, and the determination of the subtraction image is based on the non-edge-sharpened image. Edge sharpening is used to increase the attention to edges in the subsequent optimization process to avoid motion artifacts. Specifically, in this embodiment, a gradient image is determined for each image, and a corresponding edge-sharpened image is determined by a weighted linear combination of the original image and the respectively assigned gradient images. The use of step S3 is particularly advantageous for head shots and the bone edges that exist thereby.

[0083] In step S4, standard weights are determined for the weighted L2 norm used as an image quality function by image analysis. To this end, the maximum deviation is first determined according to general formula (6). Then a histogram analysis of the maximum deviation is performed to determine a threshold value. The threshold value can be determined, for example, as 5% of the maximum deviations are above the threshold value and 95% of the maximum deviations are below the threshold value. As an alternative, the threshold value can also be set fixedly. Here, the standard weight of pixels whose maximum deviations are above the threshold value is set to 0, and the standard weight of all other pixels is set to 1 (general formula (7)). The standard weight can also be defined using a continuous weighting function near the threshold value.

[0084] In other embodiments, the standard weights may also be determined at least partially based on the results of a trained image analysis function. For example, the trained image analysis function may segment metallic objects and / or pores, and the image points of the segmented area may be less weighted or their standard weights may even be set to zero. In addition, additional information, such as additional information describing a region of interest (ROI), may be taken into account in order to assign higher standard weights to pixels within the region of interest, thereby focusing the optimization process on these pixels.

[0085] In step S5, an optimization process is performed to determine the optimized mask weights, from which the appropriate sum image of the mask image to be subtracted from the feature image can be determined. In this case, the optimization is performed based on quadratic programming according to general formula (5). For this purpose, Q and c (general formula (4)) are calculated in advance, wherein the standard weights obtained in step S4 are used and then the actual optimization is performed according to the Goldfarb-Idnani method. For the test set of mask weights, a test image is determined and evaluated by the weighted L2 norm as an image processing indicator.

[0086] The resulting subtraction image is then determined in step S6 by subtracting the sum image from the (original, non-edge-sharpened) feature image. The sum image is formed as the sum of the (original, non-edge-sharpened) mask images determined by the optimized mask weights of step S5. The result is a high-quality, rapidly and robustly determined subtraction image that shows the feature, here the contrast agent.

[0087] Figure 2The processing sequence within the scope of the method shown is schematically illustrated. In steps S1 and S2, an original set of images 1 is obtained, including a feature image 2 and a plurality of mask images 3. According to the arrow 4 symbolized in step S3, an image set 5 is then determined, including an edge-sharpened feature image 2' and a plurality of edge-sharpened mask images 3'. On the one hand, it forms the basis for determining standard weights 6 (cf. the symbolized arrow 7 of step S4). Using the image set 5 and the standard weights 6 (cf. the symbolized arrows 8 and 9 of step 5), an optimized set of mask weights 10 is then determined in an optimization process, which minimizes the L2 norm weighted with the standard weights 6. Finally, the mask weights 10 and the image set 1 are used in a subtraction step S6 symbolized by arrows 11 and 12 in order to determine a subtracted image 13.

[0088] Figure 3 The image recording device 14 according to the invention is schematically shown, which is designed here as an X-ray machine with a C-arm 15, on which an X-ray emitter 16 and an X-ray detector 17 are arranged opposite each other. The C-arm is movable in order to adjust different projection geometries. The operation of the image recording device 14 is controlled by a control device 18, which is also designed to carry out the method according to the invention.

[0089] The functional structure of the control device 18 is Figure 4 As is known, the control device 18 has, in addition to the memory 19, essentially an image recording unit 20 which is designed to control the image recording operation of the image recording device 14, ie, in particular, to carry out steps S1 and S2.

[0090] The preprocessing unit 21 is designed to determine the image group 5 according to step S3. In the standard weight determination unit 22, the standard weight 6 can be determined according to step S4. The optimization unit 23 is designed to carry out the optimization process according to step S5. Finally, a determination unit 24 is also provided for determining the subtraction image 13 according to step S6.

[0091] Although the details of the present invention have been illustrated and described in detail through the preferred embodiments, the present invention is not limited to the disclosed examples, and those skilled in the art may derive other different variations therefrom without departing from the scope of protection of the present invention.

Claims

1. A computer-implemented method for determining a subtracted image (13) from a feature image (2) of an imaging region of an object under examination and a plurality of mask images (3), wherein: A subtraction image (13) is determined by subtracting a combination of mask images (3) weighted by mask weights (10) from a feature image (2), wherein the mask weights (10) are determined in an optimization process for optimizing an image quality function for a test image, the test image being calculated by subtracting a combination of mask images (3, 3') determined by a test set of mask weights from the feature images (2, 2'), characterized in that the image quality function is minimized in the form of a convex quadratic program with a dimension corresponding to the number of mask images (3, 3').

2. The method according to claim 1, characterized in that The L2 norm is used as the image quality function.

3. The method according to claim 2, characterized in that A quadratic program is formulated by multiplying the subtraction terms of the test image and separating the quadratic and linear terms in the vector of mask weights.

4. The method according to claim 2 or 3, characterized in that: The L2 norm is weighted for each pixel using a standard weight (6), wherein the standard weight (6) is determined by an image analysis process of the feature image (2, 2') and the mask image (3, 3') and / or based on additional information.

5. The method according to claim 4, characterized in that During image analysis - determining for each pixel the maximum deviation of the image value of the characteristic image (2, 2') from the image value of the mask image (3, 3'), and - Assigning a smaller labeling weight to at least a portion of the image points with the largest deviation (6).

6. The method according to claim 5, characterized in that Pixels whose maximum deviation is above a threshold value are assigned a standard weight (6) of zero and / or the standard weight (6) is determined in a function range extending at least around the threshold value according to a monotonically decreasing function.

7. The method according to claim 5 or 6, characterized in that: The standard weights (6) are determined during the image analysis process at least in part using a trained image analysis function.

8. The method according to any one of claims 4 to 7, characterized in that The additional information describes regions of interest of the imaging area which should be weighted higher.

9. The method according to any one of the preceding claims, characterized in that The feature images (2, 2') and the mask images (3, 3') are pre-processed by an image processing operation of edge sharpening for the optimization process.

10. The method according to claim 9, characterized in that A gradient image is determined in an image processing operation for a feature image (2) and for each of a plurality of mask images (3), and the edge-sharpened individual images (2', 3') are determined by a weighted linear combination of the original individual images (2, 3) and the associated gradient image.

11. The method according to any one of the preceding claims, characterized in that The optimization process was performed with the help of the Golden Idnani method.

12. The method according to any one of the preceding claims, characterized in that The characteristic images (2, 2') are filled images of digital subtraction angiography with contrast medium.

13. The method according to any one of the preceding claims, characterized in that The image quality function includes a regularization of the weights, which aims to eliminate uneven distribution of weights.

14. The method according to claim 13, characterized in that The regularization is performed based on the Euclidean norm of the weights.

15. An image capturing device (14) having a control device (18), the image capturing device comprising: - a photographing unit (20) for controlling the photographing of a plurality of mask images (3) and feature images (2) of an imaging region of an inspection object, - a determination unit (24) for determining a silhouette image (13) by subtracting a combination of the mask image (3) weighted by the mask weight (10) from the feature image (2), and - an optimization unit (23) for determining the mask weights (10) in an optimization process for optimizing the image quality function of a test image, the test image being calculated by subtracting a combination of the mask images (3, 3') determined by the test set of mask weights from the feature images (2, 2'), It is characterized in that the optimization unit (23) is designed to optimize the image quality function in the form of a convex quadratic programming of a dimension corresponding to the number of mask images (3, 3'). 16 . A computer program which, when executed in a computing unit, causes the computing unit to carry out the steps of the method according to claim 1 .

17. An electronically readable data carrier having stored thereon a computer program according to claim 16.

Images