COMPUTER-IMPLEMENTED METHODS AND SYSTEMS FOR PROVISION OF A CORRECTION ALGORITHM FOR AN X-RAY IMAGE AND FOR CORRECTION OF AN X-RAY IMAGE, X-RAY FACILITY, COMPUTER PROGRAM, AND ELECTRONICALLY READABLE DATA MEDIUM

Abstract

A computer-implemented method for provision of a correction algorithm for an x-ray image that was recorded with an x-ray source emitting an x-ray radiation field, a filter facility spatially modulating an x-ray radiation dose, and an x-ray detector is provided. The correction algorithm includes a trained first processing function that, from first input data that includes at least one first physical parameter describing the x-ray radiation field and/or the measurement and at least one second physical parameter describing the spatial modulation of the filter facility, determines first output data. The first output data includes a mask for brightness compensation with regard to the spatial modulation of the filter facility in the x-ray image. The method includes providing first training data, providing an autoencoder for masks, and training of the autoencoder using the first training data. The method also includes determining an assignment rule, and providing the trained first processing function.

Claims

1. A computer-implemented method for provision of a correction algorithm for an x-ray image that has been recorded with an x-ray source emitting an x-ray radiation field, a filter facility spatially modulating an x-ray radiation dose, and an x-ray detector, wherein the correction algorithm comprises a trained first processing function that, from first input data that includes at least one first physical parameter describing the x-ray radiation field, the measurement and at least one second physical parameter describing the spatial modulation of the filter facility, or a combination thereof, determines first output data that includes a mask for brightness compensation with regard to the spatial modulation of the filter facility in the x-ray image, the method comprising: providing first training data, the first training data including first training datasets each with a mask, wherein each of the first training datasets is assigned the at least one first physical parameter and the at least one second physical parameter of the first input data assigned for determining the mask; providing an autoencoder for masks, wherein the autoencoder includes an encoder configured to determine a latent space representation of the mask and a decoder configured to determine a comparison mask from the latent space representation; training the autoencoder using the first training data; determining an assignment rule between the at least one first physical parameter and the at least one second physical parameter of the first input data, which are assigned to the first training datasets, and the latent space representations of the masks of the respective first training dataset; providing the trained first processing function as a combination of the assignment rule and the trained decoder.

2. The method of claim 1, wherein: a larger number of latent space parameters of the latent space representation than physical parameters of the first input data are used; 3 to 30 physical parameters of the first input data, 3 to 30 latent space parameters of the latent space representation, or a combination thereof is used; at least one functional relationship is determined at least partly by fitting, interpolation, extrapolation, or any combination thereof; or any combination thereof.

3. The method of claim 2, wherein the at least one functional relationship includes an assignment rule exclusively including functional relationships.

4. The method of claim 1, wherein: the at least one first physical parameter includes: a tube voltage of the x-ray source; a tube current of the x-ray source; a pre-filter parameter, an aperture parameter, or the pre-filter parameter and the aperture parameter; a distance of the x-ray source to the x-ray detector; a distance of the x-ray source to the filter facility; a distance of the filter facility from the x-ray detector; a pulse length of an x-ray pulse creating the x-ray radiation field; a number of x-ray pulses since a beginning of a recording of a series of x-ray images; at least one focal parameter describing a geometry of a focus point; a zoom of the x-ray detector; an orientation of the x-ray detector; a frame rate of the x-ray detector; or a combination thereof, the at least one second physical parameter includes: a material of the filter facility; at least one filter thickness parameter; at least one change over time of time parameters describing the spatial modulation; or any combination thereof; or a combination thereof.

5. The method of claim 1, wherein the first training datasets are determined from x-ray images recorded with and without the filter facility.

6. The method of claim 1, wherein the correction algorithm further includes a second processing function downstream of the first trained processing function for refining the mask determined by the first processing function, wherein the second processing function includes a generator network configured to use, as second input data, an x-ray image, for which the refined mask is to be determined, and a mask to be refined determined using the second input data using the trained first processing function for the at least one first physical parameter and the at least one second physical parameter of the first input data of the x-ray image, wherein the method further comprises: training the second processing function, the training of the second processing function comprising: providing second training data comprising x-ray images of an object recorded with and without the filter facility with assigned physical parameters of the first input data; discriminating, by a discriminator network, between x-ray images recorded without the filter facility and corrected x-ray images obtained as second output data using the refined mask, for completion of a generative adversarial network, training the generator network and the discriminator network, the training of the generator network and the discriminator network comprising using an output of the discriminator network, by comparison of the corrected x-ray image of a second training dataset with the x-ray image recorded without the filter facility of the second training dataset, for fitting the generator network and the discriminator network; and providing the correction algorithm, the correction algorithm comprising a combination of the trained first processing function and the trained second processing function.

7. The method of claim 6, wherein the generator network, at least during the training of the generator network, receives, as further second input data, at least one boundary condition restricting the second output data.

8. The method of claim 7, wherein the at least one boundary condition includes: a boundary condition restricting a deviation of the refined mask from the mask to be refined, a space of the possible mask to be refined, or a combination thereof; a boundary condition providing a smoothness of the refined mask; a boundary condition restricting a type of arithmetical operations for obtaining the refined mask from the mask to be refined; or any combination thereof.

9. The method of claim 6, wherein the x-ray images are logarithmically transformed before being used as the second input data, the second training data, or the second input data and the second training data.

10. A computer-implemented method for correction of an x-ray image that was recorded with an x-ray source emitting an x-ray radiation field, a filter facility spatially modulating the x-ray radiation dose, and an x-ray detector, using a correction algorithm, wherein the correction algorithm comprises a trained first processing function that, from first input data that includes at least one first physical parameter describing the x-ray radiation field, the measurement and at least one second physical parameter describing the spatial modulation of the filter facility, or a combination thereof, determines first output data that includes a mask for brightness compensation with regard to the spatial modulation of the filter facility in the x-ray image, providing the correction algorithm comprising providing first training data, the first training data including first training datasets each with a mask, wherein each of the first training datasets is assigned the at least one first physical parameter and the at least one second physical parameter of the first input data assigned for determining the mask, providing the correction algorithm further comprising providing an autoencoder for masks, wherein the autoencoder includes an encoder configured to determine a latent space representation of the mask and a decoder configured to determine a comparison mask from the latent space representation, providing the correction algorithm further comprising training the autoencoder using the first training data, determining an assignment rule between the at least one first physical parameter and the at least one second physical parameter of the first input data, which are assigned to the first training datasets, and the latent space representations of the masks of the respective first training dataset, and providing the trained first processing function as a combination of the assignment rule and the trained decoder, the computer-implemented method comprising: determining a mask using the correction algorithm from at least the first input data assigned to the x-ray image to be corrected; and correcting the x-ray image to be corrected using the mask.

11. The computer-implemented method of claim 10, further comprising applying a denoising method to the corrected x-ray image.

12. A system for provision of a correction algorithm for an x-ray image that was recorded with an x-ray source emitting an x-ray radiation field, a filter facility spatially modulating the x-ray radiation dose, and an x-ray detector, wherein the correction algorithm comprises a trained first processing function that, from first input data that includes at least one first physical parameter describing the x-ray radiation field, the measurement and at least one second physical parameter describing spatial modulation of the filter facility, or a combination thereof, determines first output data that includes a mask for brightness compensation with regard to the spatial modulation of the filter facility in the x-ray image, the system comprising: a first training interface configured to provide first training data, the first training data comprising first training datasets each with a mask, wherein the at least one first physical parameter and the at least one second physical parameter of the first input data assigned for determining the mask are assigned to each of the first training datasets; a first training unit configured to train an autoencoder for masks, wherein the autoencoder has an encoder for determining a latent space representation of the mask and a decoder for determining a comparison mask from the latent space representation, using the first training data; a rule determination unit configured to determine an assignment rule between the at least one first physical parameter and the at least one second physical parameter of the first input data, which are assigned to the first training datasets, and the latent space representations of the masks of the respective first training dataset; and a second training interface configured to provide the trained first processing function as a combination of the assignment rule and the trained decoder.

13. The system of claim 12, wherein the filter facility includes a region-of-interest (ROI) filter.

14. A system for correction of an x-ray image that was recorded with an x-ray source emitting an x-ray radiation field, a filter facility spatially modulating the x-ray radiation dose, and an x-ray detector, the system comprising: a first application interface configured to accept a correction algorithm provided from a system for provision of the correction algorithm for the x-ray image that was recorded, wherein the correction algorithm comprises a trained first processing function that, from first input data that includes at least one first physical parameter describing the x-ray radiation field, the measurement and at least one second physical parameter describing spatial modulation of the filter facility, or a combination thereof, determines first output data that includes a mask for brightness compensation with regard to the spatial modulation of the filter facility in the x-ray image, wherein the system for provision comprises a first training interface configured to provide first training data, the first training data comprising first training datasets each with a mask, wherein the at least one first physical parameter and the at least one second physical parameter of the first input data assigned for determining the mask are assigned to each of the first training datasets, wherein the system for provision further comprises a first training unit configured to train an autoencoder for masks, wherein the autoencoder has an encoder for determining a latent space representation of the mask and a decoder for determining a comparison mask from the latent space representation, using the first training data, wherein the system for provision further comprises a rule determination unit configured to determine an assignment rule between the at least one first physical parameter and the at least one second physical parameter of the first input data, which are assigned to the first training datasets, and the latent space representations of the masks of the respective first training dataset, and a second training interface configured to provide the trained first processing function as a combination of the assignment rule and the trained decoder; a second application interface configured to accept the x-ray image to be corrected along with assigned first and second physical parameters; a mask determination unit configured to determine a mask using the correction algorithm from at least first input data assigned to the x-ray image to be corrected; a correction unit configured to correct the x-ray image to be corrected using the mask; and a third application interface configured to provide the corrected x-ray image.

15. The system of claim 14, wherein the filter facility includes a region-of-interest (ROI) filter.

16. An x-ray facility comprising: an x-ray source; an x-ray detector; a filter facility configured to spatially modulate an x-ray radiation dose; and a control facility comprising a system for correction of an x-ray image that was recorded with the x-ray source emitting an x-ray radiation field, the filter facility spatially modulating the x-ray radiation dose, and the x-ray detector, the system comprising: a first application interface configured to accept a correction algorithm provided from a system for provision of the correction algorithm for the x-ray image that was recorded, wherein the correction algorithm comprises a trained first processing function that, from first input data that includes at least one first physical parameter describing the x-ray radiation field, the measurement and at least one second physical parameter describing spatial modulation of the filter facility, or a combination thereof, determines first output data that includes a mask for brightness compensation with regard to the spatial modulation of the filter facility in the x-ray image, wherein the system for provision comprises a first training interface configured to provide first training data, the first training data comprising first training datasets each with a mask, wherein the at least one first physical parameter and the at least one second physical parameter of the first input data assigned for determining the mask are assigned to each of the first training datasets, wherein the system for provision further comprises a first training unit configured to train an autoencoder for masks, wherein the autoencoder has an encoder for determining a latent space representation of the mask and a decoder for determining a comparison mask from the latent space representation, using the first training data, wherein the system for provision further comprises a rule determination unit configured to determine an assignment rule between the at least one first physical parameter and the at least one second physical parameter of the first input data, which are assigned to the first training datasets, and the latent space representations of the masks of the respective first training dataset, and a second training interface configured to provide the trained first processing function as a combination of the assignment rule and the trained decoder; a second application interface configured to accept the x-ray image to be corrected along with assigned first and second physical parameters; a mask determination unit configured to determine a mask using the correction algorithm from at least first input data assigned to the x-ray image to be corrected; a correction unit configured to correct the x-ray image to be corrected using the mask; and a third application interface configured to provide the corrected x-ray image.

17. In a non-transitory computer-readable storage medium that stores instructions executable by one or more processors for provision of a correction algorithm for an x-ray image that has been recorded with an x-ray source emitting an x-ray radiation field, a filter facility spatially modulating an x-ray radiation dose, and an x-ray detector, wherein the correction algorithm comprises a trained first processing function that, from first input data that includes at least one first physical parameter describing the x-ray radiation field, the measurement and at least one second physical parameter describing the spatial modulation of the filter facility, or a combination thereof, determines first output data that includes a mask for brightness compensation with regard to the spatial modulation of the filter facility in the x-ray image, the instructions comprising: providing first training data, the first training data including first training datasets each with a mask, wherein each of the first training datasets is assigned the at least one first physical parameter and the at least one second physical parameter of the first input data assigned for determining the mask; providing an autoencoder for masks, wherein the autoencoder includes an encoder configured to determine a latent space representation of the mask and a decoder configured to determine a comparison mask from the latent space representation; training the autoencoder using the first training data; determining an assignment rule between the at least one first physical parameter and the at least one second physical parameter of the first input data, which are assigned to the first training datasets, and the latent space representations of the masks of the respective first training dataset; providing the trained first processing function as a combination of the assignment rule and the trained decoder.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Further advantages and details of the present invention emerge from the exemplary embodiments described below and also with the aid of the drawings. In the drawings:

[0057] FIG. 1 shows an exemplary embodiment of a neural network;

[0058] FIG. 2 shows an exemplary embodiment of a convolutional neural network (CNN);

[0059] FIG. 3 shows a basic sketch of an embodiment of an x-ray facility;

[0060] FIG. 4 shows a recording geometry produced during the recording of a patient;

[0061] FIG. 5 shows a filter facility embodied as an ROI filter;

[0062] FIG. 6 shows a schematic diagram of an x-ray image to be corrected recorded using the filter facility in accordance with FIG. 5;

[0063] FIG. 7 shows a diagram for determining a mask for correction of an x-ray image to be corrected;

[0064] FIG. 8 shows a flow diagram of a first stage of one embodiment of a provision method;

[0065] FIG. 9 shows a flow diagram of a second stage of one embodiment of a provision method;

[0066] FIG. 10 shows a flow diagram of one embodiment of a correction method;

[0067] FIG. 11 shows the functional structure of a provision system in accordance with an embodiment; and

[0068] FIG. 12 shows the functional structure of a correction system in accordance with an embodiment.

DETAILED DESCRIPTION

[0069] FIG. 1 shows an exemplary embodiment of an artificial intelligence neural network 1. Other expressions for such a network are “artificial neural network,” “neural network,” “artificial neural net,” or “neural net”.

[0070] The artificial neural network 1 includes 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. In one embodiment, the first node 6 to 18 and the second node 6 to 18 may be identical. For example, in FIG. 1, 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 referred to as the ingoing edge for the second node 6 to 18 and as the outgoing edge for the first node 6 to 18.

[0071] In this exemplary embodiment, the nodes 6 to 18 of the artificial intelligence neural network 1 may be arranged in layers 2 to 5, where the layers 2 to 5 may have an intrinsic order that is introduced by the edges 19 to 21 between the nodes 6 to 18. For example, edges 19 to 21 may only be provided between neighboring layers of nodes 6 to 18. In the exemplary embodiment shown, there exists an input layer 110 that only has the nodes 6, 7, 8, without an ingoing edge in each case. The output layer 5 includes only the nodes 17, 18, without outgoing edges in each case, where further hidden layers 3 and 4 lie between the input layer 2 and the output layer 5. In the general case, any number of hidden layers 3, 4 may be chosen. The number of the nodes 6, 7, 8 of the input layer 2 may correspond to the number of input values in the neural network 1, and the number of nodes 17, 18 in the output layer 5 may correspond to the number of output values of the neural network 1.

[0072] For example, a number (e.g., a real number) may be assigned to the nodes 6 to 18 of the neural network 1. In this case, x.sup.(n).sub.i refers to the value of the ith node 6 to 18 of the nth layer 2 to 5. The values of the nodes 6, 7, 8 of the input layer 2 are equivalent to the input value of the neural network 1, while the values of the nodes 17, 18 or the output layer 5 are equivalent to the output values of the neural network 1. Further, edge 19, 20, 21 may be assigned a weight in the form of a real number. For example, the weight is a real number in the interval [−1, 1] or in the interval [0, 1,]. In this case, w.sup.(m,n).sub.i,j refers to the weight of the edge between the ith node 6 to 18 of the mth layer 2 to 5 and the jth node 6 to 18 of the nth layer 2 to 5. The abbreviation w.sub.i,j.sup.(n) is further defined for the weight w.sub.i,j.sup.(n,n+1).

[0073] In order to calculate output values of the neural network 1, the input values are propagated through the neural network 1. For example, the values of the nodes 6 to 18 of the (n+1)th layer 2 to 5 may be calculated based on the values of the nodes 6 to 18 of the nth layer 2 to 5 by

x.sub.j.sup.(n+1)=f(Σ.sub.ix.sub.i.sup.(n).Math.w.sub.i,j.sup.(n)).

[0074] In this equation, f is a transfer function that may also be referred to as an activation function. Known transfer functions are step functions, Sigmoid functions (e.g., the logistical function, the generalized logistical function, the tangens hyperbolicus, the arcustangens, the error function, the smooth step function) or rectifier functions. The transfer function is essentially used for standardization purposes.

[0075] For example, the values are propagated layer-by-layer through the neural network 1, where values of the input layer 2 are given by the input data of the neural network 1. Values of the first hidden layer 3 may be calculated based on the values of the input layer 2 of the neural network 1, values of the second hidden layer 4 may be calculated based on the values in the first hidden layer 3, etc.

[0076] In order to be able to define the values w.sub.i,j.sup.(n) for the edges 19 to 21, the neural network 1 is to be trained using training data. For example, training data includes training input data and training output data, which are referred to below as t.sub.i. For a training step, the neural network 1 is applied to the training input data in order to determine calculated output data. For example, the training output data and the calculated output data include a number of values, where the number is determined as the number of the nodes 17, 18 of the output layer 5.

[0077] For example, a comparison between the calculated output data and the training output data is used in order to recursively fit the weights within the neural network 1 (e.g., back propagation algorithm). For example, the weights may be changed in accordance with

w′.sub.i,j.sup.(n)=w.sub.i,j.sup.(n)−γ.Math.δ.sub.j.sup.(n).Math.x.sub.i.sup.(n)

where γ is a learning rate, and the numbers δ.sub.j.sup.(n) may be calculated recursively as

δ.sub.j.sup.(n)=(Σ.sub.kδ.sub.k.sup.(n+1.Math.w.sub.j,k.sup.(n+1)).Math.f′(Σ.sub.ix.sub.i.sup.(n).Math.w.sub.i,j.sup.(n))

based on δ.sub.j.sup.(n+1), when the (n+1)th layer is not the output layer 5, and

δ.sub.j.sup.(n)=(x.sub.k.sup.(n+1)−t.sub.j.sup.(n+1)).Math.f′(Σ.sub.ix.sub.i.sup.(n).Math.w.sub.i,j.sup.(n))

the (n+1)th layer is the output layer 5, where f′ is the first derivation of the activation function, and yj(n+1) is the comparison training value for the jth node 17, 18 of the output layer 5.

[0078] Also given below with respect to FIG. 2 is an example for a convolutional neural network (CNN). In this example, the expression “layer” in the figure is employed in a slightly different way than for classical neural networks. For a classical neural network, the expression “layer” only refers to the set of nodes that forms a layer, and thus, to a specific generation of nodes. For a convolutional neural network, the expression “layer” is often used as an object that actively changes data (e.g., as a set of nodes of the same generation and either the set of ingoing or outgoing edges).

[0079] FIG. 2 shows an exemplary embodiment of a convolutional neural network 22. In the exemplary embodiment shown, the convolutional neural network 22 includes an input layer 23, a convolutional layer 24, a pooling layer 25, a fully connected layer 26, and an output layer 27. In alternate embodiments, the convolutional neural network 22 may contain a number of convolutional layers 24, a number of pooling layers 25, and a number of fully connected layers 26, just like other types of layers. A given sequence of the layers may be selected, where usually fully connected layers 26 form the last layers before the output layer 27.

[0080] For example, within a convolutional neural network 22, the nodes 28 to 32 of one of the layers 23 to 27 may be arranged in a d-dimensional matrix or as a d-dimensional image. For example, in the two-dimensional case, the value of a node 28 to 32 may be referred to with the indices i, j in the nth layer 23 to 27 as x.sup.(n)[i,j]. The arrangement of the nodes 28 to 31 of a layer 23 to 27 does not have any effect as such on the calculations within the convolutional neural network 22 as such, since these effects are exclusively produced by the structure and the weights of the edges.

[0081] A convolutional layer 24 is, for example, characterized in that the structure and the weights of the ingoing edges form a convolution operation based on a specific number of kernels. For example, the structure and the weights of the ingoing edges may be selected so that the values x.sub.k.sup.(n) of the nodes 29 of the convolutional layer 24 are determined as a convolution x.sub.k.sup.(n)=K.sub.k*x.sup.(n−1) based on the values x.sup.(n−1) of the node 28 of the preceding layer 23, where the convolution * in the two-dimensional case may be defined as

x.sub.k.sup.(n)[i,j]=(K.sub.k*x.sup.(n−1))[i,j]=Σ.sub.i′Σ.sub.j′K.sub.k[i′,j′].Math.x.sup.(n−1)[i−i′, j−j′].

[0082] In this equation, the kth kernel K.sub.k is a d-dimensional matrix (e.g., a two-dimensional matrix) that may be small by comparison with the number of the nodes 28 to 32 (e.g., a 3×3 matrix or a 5×5 matrix). For example, this implies that weights of the ingoing edges are not independent, but are selected so that the weights create the convolution equation above. In the example for a kernel that forms a 3×3 matrix, there exist only nine independent weights (e.g., where each entry of the kernel matrix corresponds to an independent weight), regardless of the number of the nodes 28 to 32 in the corresponding layers 23 to 27. For example, for a convolutional layer 24, the number of the nodes 29 in the convolutional layer 24 is equivalent to the number of the nodes 28 in the preceding layer 23 multiplied by the number of the convolution kernels.

[0083] When the nodes 28 of the preceding layer 23 are arranged as a d-dimensional matrix, the use of the plurality of kernels may be understood as the insertion of a further dimension, which is also referred to as a depth dimension, so that the nodes 29 of the convolutional layer 24 are arranged as a (d+1)-dimensional matrix. When 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 convolution kernels may be understood as an expansion along the depth dimension, so that the nodes 29 of the convolutional layer 24 are equally arranged as a (d+1)-dimensional matrix. The size of the (d+1)-dimensional matrix in the depth dimension is greater by the factor formed by the number of the kernels than in the preceding layer 23.

[0084] The advantage of using convolution kernels 24 is that the spatially local correlation of the input data may be utilized by a local connection pattern between nodes of neighboring layers being created (e.g., in that each node only has connections to a small area of the node of the preceding layer).

[0085] In the exemplary embodiment shown, the input layer 23 includes thirty-six nodes 28 that are arranged as a two-dimensional 6×6 matrix. The convolutional layer 24 includes seventy-two nodes 29 that are arranged as two two-dimensional 6×6-matrices, where each of the two matrices is the result of a convolution of the values of the input layer 23 with a convolution kernel. In the same way, the nodes 29 of the convolutional layer 24 may be understood as being arranged as a three-dimensional 6×6×2 matrix, where the last-mentioned dimension is the depth dimension.

[0086] A pooling layer 25 is characterized in that the structure and the weights of the ingoing edges as well as the activation function of its nodes 30 define a pooling operation based on a non-linear pooling function f. For example, in the two-dimensional case, the values x.sup.(n) of the nodes 30 of the pooling layer 25 may be calculated, based on the values x.sup.(n+1) of the nodes 29 of the preceding layer 24, as

x.sup.(n)[i,j]=f(x.sup.(n−1)[id.sub.1, jd.sub.2], . . . , x.sup.(n−1)[id.sub.1+d.sub.1−1, jd.sub.2+d.sub.2−1]).

[0087] In other words, the number of nodes 29, 30 may be reduced by the use of a pooling layer 25, in that a number of d.sub.1×d.sub.2 of neighboring nodes 29 in the preceding layer 24 is replaced by a single node 30 that is calculated as a function of the values of the the number of neighboring nodes 29. For example, the pooling function f may be a maximum function, an averaging or the L2 norm. For example, for a pooling layer 25, the weights of the ingoing edges may be defined and not modified by training.

[0088] The advantage of using a pooling layer 25 is that the number of nodes 29, 30 and the number of parameters is reduced. This leads to a reduction in the amount of calculations necessary within the convolutional neural network 22 and thus to a control of the overfitting.

[0089] In the exemplary embodiment shown, the pooling layer 25 involves a max pooling layer, in which four neighboring nodes are replaced by just one single node, the value of which is formed by the maximum of the values of the four neighboring nodes. The max pooling is applied to each d-dimensional matrix of the preceding layer; in this exemplary embodiment, the max pooling is applied to each of the two two-dimensional matrices, so that the number of nodes is reduced from seventy-two to eighteen.

[0090] A fully connected layer 26 is characterized by a plurality (e.g., all) edges being present between the nodes 30 of the preceding layer 25 and the nodes 31 of the fully connected layer 26, where the weight of each of the edges may be fitted individually. In this exemplary embodiment, the nodes 30 of the preceding layer 25 and the fully connected layer 26 are both shown as two-dimensional matrices and also as non-contiguous nodes (shown as a row of nodes, where the number of the nodes has been reduced so that the nodes may be shown more easily). In this exemplary embodiment, the number of nodes 31 in the fully connected layer 26 is equal to the number of the nodes 30 in the preceding layer 25. In alternate forms of embodiment, the number of the nodes 30, 31 may be different.

[0091] Further, in this exemplary embodiment, the values of the nodes 32 of the output layer 27 are determined by the softmax function being applied to the values of the nodes 31 of the preceding layer 26. Through application of 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 a real number between 0 and 1. When the convolutional neural network 22 is used for classification of input data, the values of the output layer 27, for example, may be interpreted as the probability of the input data falling into one of the different classes.

[0092] A convolutional neural network 22 may likewise have a ReLU layer, where ReLU is an acronym for “rectified linear units”. For example, the number of the nodes and the structure of the nodes within a ReLU layer is equivalent to the number of the nodes and the structures of the nodes of the preceding layer. The value of each node in the ReLU layer may be calculated, for example, by application of a rectifier function to the value of the corresponding node of the preceding layer. Examples of rectifier functions are f(x)=max(0,x), the tangens hyperbolicus, or the Sigmoid function.

[0093] Convolutional neural networks 22 may be trained, for example, based on the back propagation algorithm. In order to avoid an overfitting, methods of regularization may be employed (e.g., dropout of individual nodes 28 to 32), stochastic pooling, use of artificial intelligence data, weight decomposition based on the L1 or the L2 standard, or maximum standard restrictions.

[0094] FIG. 3 shows a basic sketch of one embodiment of an x-ray facility 33. The x-ray facility 33 has a C-arm 34, on which an x-ray source 35 (e.g., an x-ray tube) and an x-ray detector 36 are arranged opposite one another. The x-ray facility 33 further includes a filter facility 37 (e.g., an ROI filter) that may be used, for example, for x-ray monitoring of medical interventions (e.g., minimally invasive interventions) for reducing the x-ray radiation dose for the patient and the medical personnel. The operation of the x-ray facility 33 is controlled by a control facility 38 that is embodied as a correction system 39 for correction of x-ray images for balancing brightness recorded with the filter facility 37, since the filter facility 37 is embodied for spatial modulation of the x-ray radiation dose. Optionally, the control facility 38 may also include a provision system 40. While the correction system 39 is embodied for carrying out a correction method of one or more of the present embodiments, the provision system 40 is embodied for carrying out a provision method of one of more of the present embodiments, which will be explained in greater detail below.

[0095] FIG. 4 shows in greater detail the recording geometry when the filter facility 37 is used. The x-ray source 35, starting from a focus point 41, creates an x-ray radiation field 42 in order to record a recording region of a patient 43 only shown schematically here. This may also result in effects lying outside the filter facility 37, which relate to the mapping of the filter facility 37 on the x-ray detector 36; in accordance with the present embodiments, these are likewise to be taken into account in a brightness compensation (e.g., the determination of a brightness correction mask). Thus, with an x-ray tube, for example, this may result in a “jittering” of the focus point 41, which, due to the filter facility 37 being arranged close to the focus point 41 (e.g., at a distance of 5 cm) may be mapped in a relevant way as a “wobbling” of specific regions. It may further also be relevant that the focus point 41 is mostly not completely punctiform, but has a specific geometry. Also, the patient 43, by beam hardening, scattered radiation, and the like, may have an influence on the mapping of the filter facility 37 at the x-ray detector 36, where the setting of the x-ray detector 36 itself is to be further mentioned. For determining a correct brightness correction mask to be used for brightness compensation, not only knowledge about the filter facility 37 and its state during imaging is relevant, but also about other aspects (e.g., system settings for dynamics of the situation and influences of the patient 43).

[0096] FIG. 5 shows an exemplary embodiment of the filter facility 37 as an ROI filter in more detail. The filter facility 37 consists of a filter material 44 (e.g., tungsten) that has a central passage opening 46 that defines an ROI 45, in which the standard dose of x-ray radiation is still present, while outside the ROI 45, a strong attenuation of the x-ray radiation through the filter material 44 (e.g., a thin tungsten layer) is present. The filter facility 37 may further have an actuator system 47, by which the size and/or position of the ROI 45 may be able to be changed (e.g., by exchanging various filter plates available, movement of the filter material 44, or the like). The actuator system 47 may be actuated, for example, as a function of user inputs, for tracking a medical instrument (e.g., a catheter), and/or as a function of eye tracking data.

[0097] FIG. 6 shows a schematic of an x-ray image 48 to be corrected that was recorded using the filter facility 37 of FIG. 5. The ROI 45 is mapped clearly visibly in the x-ray image 48, since there, the anatomy 49 of the patient 43 appears far brighter than in the outer regions surrounding the ROI 45, as is indicated by the cross hatching. As has already been explained, a smooth transition is mostly produced in this case, where, however, the geometrical form of the opening 46, due to dynamic effects, such as the jittering of the focus point 41, is not necessarily reproduced in the exact correct form.

[0098] FIG. 7 shows an option for how a mask may be determined, such as at least with regard to system settings and their effects (e.g., in a calibration measurement). In this case, two x-ray images 50, 51 without a patient 43 with specific, fixed system settings, described by first and second physical parameters, may be recorded. The x-ray image 50 is a flat-field x-ray image without the filter facility 37 as object, while the x-ray image 51 is a flat-field x-ray image for which the filter facility 37 has been used. If the x-ray image 51 is now subtracted from the x-ray image 50 (cf., act 52), a mask 53 is produced. In this case, the x-ray images 50, 51 are logarithmically transformed, as usual, before determination of the mask 53. The mask 53 may be employed for additive correction of an x-ray image 48 to be corrected, when exactly the same system settings are present, where, however, the described dynamic effects as well as the effects by the patient 43 may then not be taken into account. This, however, makes possible the procedure in accordance with the exemplary embodiments described here, with which extremely expensive calibration measurements for all conceivable system settings are avoided.

[0099] A provision method in accordance with the present embodiments is now described with reference to FIGS. 8 and 9, where FIG. 8 relates to the training of the first processing function, thus to the first stage, and FIG. 9 relates to the training of the second stage, thus to the second processing function. By combination of the trained first processing function and the trained second processing function, a highly accurate mask for correction of x-ray images 48, encompassing many effects for use in a correction algorithm of the present embodiments, may be provided.

[0100] The first processing algorithm relates to the characteristics of the measurement itself (e.g., the technical settings or system settings, such as creation of the x-ray radiation, geometry, detector operation, characteristics, or settings of the filter facility). In order to train the first processing algorithm, in acts Si and S2, using the method described for FIG. 7, first training data is recorded, where each training dataset includes a mask 53 and also the first and second physical parameters 54, 55 used in the recording of the mask 53, to be used as first input data.

[0101] The first and second physical parameters describe system settings of the x-ray facility 33 used. For example, first physical parameters are concerned with the x-ray radiation field, its creation, and the measurement, also including the recording geometry and the operation of the x-ray detector 36, and second physical parameters 55 are explicitly concerned with the characteristics of the filter facility 37. For example, the second physical parameters 55 may relate to the filter material, the filter material thickness, as well as the size and the location of the ROI 45, if necessary also temporarily by a time parameter. First physical parameters 54 may, for example, relate to settings of the x-ray source (e.g., tube current, tube voltage, and pulse length), focus point settings (e.g., focus point sizes and angles), detector settings (e.g., zoom, orientation, and frame rate), as well as the recording geometry (e.g., SID, SOD and OID, where the filter facility 37 counts as the object). In this case, the physical parameters 54, 55 to be used in the calibration are, where possible, chosen so that the physical parameters 54, 55 cover the setting space relevant for the actual recordings.

[0102] With the settings of the first and second physical parameters, in act S1, x-ray images 50 and 51 are recorded, as described for FIG. 7. In act S2, respective masks 53 are determined from the x-ray images 50 and 51, as described. With the assignment of the first and second physical parameters 54, 55 to these first masks, the first training data is determined, so that an autoencoder 56 for the masks 53, as is basically known, may be trained, which is indicated by the arrow 57. The autoencoder 56 includes, as is basically known, at least one CNN, where an encoder 58 and a decoder 59 are formed. A latent space representation 60 with just a few latent space parameters is determined by the encoder 58, which should be sufficient for the mask 53 to be reproduced as exactly as possible by the decoder 59. In other words, the decoder 59, during the training in accordance with arrow 57, delivers comparison masks 61 that may be compared with the respective input mask 53 of the respective training data, where the deviation is minimized.

[0103] A relationship exists between the latent space representation 60 and the first and second physical parameters. Sets of first and second physical parameters, between which there are only few or small changes, deliver similar latent space representations 60 in this case. If a physical parameter 54, 55 changes the recording, latent space parameters also having a relationship with this physical parameter, which indeed map relevant characteristics, change. If, for example, the distance changes between the x-ray source 35 and the x-ray detector 36 (SID), where the distance from the x-ray source 35 to the filter facility 37 (SOD) remains constant, however, the opening 46 is mapped larger, so that changes in latent space parameters relating to this characteristic occur. The fact that a relationship exists at least with a part of the first and second physical parameters 54, 55 is utilized in act S3 in order to determine an assignment rule 62 from latent space representations 60 (e.g., in specific terms, latent space parameters) to any given sets of first and second physical parameters 54, 55, which may contain functional relationships between the first and second physical parameters 54, 55 and the latent space parameters of the latent space representation 60. The functional relationships may be parameterized by fitting and/or interpolation and/or extrapolation. Accordingly, the assignment rule 62 for any given system settings (e.g., any given first and second physical parameters 54, 55) allows the associated latent space representation 60 to be determined, so that using the trained decoder 59, a rough estimation of a mask used in training of the second processing function accordance with FIG. 9 may be determined. Accordingly, for the first stage, as output of the trained first processing function 63 (cf., FIG. 9), a combination of the assignment rule 62 and the decoder 59 may be provided. This provides that the trained first processing function uses as input data the first and second physical parameters 54, 55 in order to output as first output data a mask 64 still to be refined by the second stage.

[0104] In accordance with FIG. 9, the second processing function is trained and provided. In order to now first determine the second training data, in act S4 taking place over a longer period of time, pairs of images of x-ray images 65 of different patients without filter facility 37 and x-ray images of the same patient in each case using the same physical parameters 54, 55 with filter facility 13 are recorded and in act S5 logarithmically transformed, as is basically known. This provides that the second training data includes a pair of images consisting of one x-ray image 65 that was recorded without filter facility 13 and an x-ray image 66 that was recorded with filter facility 37, as well as the first and second physical parameters 54, 55 used for the recording, associated first input data (shown by way of summary in FIG. 9 as first input data with the reference character 67). The x-ray images 65, 66 in exemplary embodiments do not necessarily have to show the same object and/or the same point in time, but certainly may.

[0105] Due to the trained first processing function 63, it is possible, for each pair of images 65, 66, using this mask 64 to be refined, to determine a generator network 69 that forms the second processing function 70 as first output data. These, together with boundary conditions 68 still to be discussed as well as the respective x-ray images 66 to be corrected, recorded with the filter facility 37, form second input data for the second processing function 70, which as second output data 69, should deliver a refined mask 71. In order to train the second processing function 70, a generative adversarial network (GAN) is created, in that a discriminator network 72 that serves to discriminate between x-ray images 74 corrected by the refined mask 71 present by addition 73 and true x-ray images 65 is inserted. As output, the discriminator network 72 delivers an adversarial loss 75 or classification loss, which, in simple terms, expresses how unrealistic the corrected x-ray image 74 still is by comparison with the realistic x-ray image 65. This adversarial loss is to be minimized and is thus used during the training process 76 for fitting the generator network 69 as well as the discriminator network 72, as is basically known.

[0106] The result of the training process 76 is then the trained second processing function 70, which delivers refined masks 71.

[0107] The boundary conditions serve various purposes and may, for example, make sure that refined masks 71 do not deviate too much from the specification or the space of possible masks 64, that the smoothness of the mask 71 is provided and that only a specific set of arithmetical operations may be carried out in order to get from the input mask 64 to the refined mask 71. In this way, the generator network 69 (e.g., based on physical observations, such as an analysis of the space of masks 53 obtained in the acts S1 and S2) is prevented from creating unrealistic outputs, which may then irritate medical personnel if applied (e.g., by causing artifacts to arise).

[0108] The addition of the trained second processing function 70 for correction allows aspects, such as, where necessary, variable geometry and position of the ROI 45 due to the jittering of the focus point 41 at low pulse lengths, movements of the ROI 45 (e.g., due to eye tracking), effects caused by patients recorded, such as beam hardening and the like, to be taken into consideration. At the same time, the mask 64 may be adapted with respect to the general brightness stabilization.

[0109] FIG. 10 shows a flow diagram of an exemplary embodiment of a correction method. According to the method, in act S6, using first and second physical parameters 54, 55 (e.g., first input data 67), an x-ray image 48 to be corrected is recorded and logarithmically transformed in act S7. The trained first processing function 63 is then used in order to determine the mask 64 to be refined, in that first, in act S8, the assignment rule 62 is used in order to determine the latent space representation 60; after this, in act S9, the trained decoder 59 is used in order to determine the mask 64 to be refined from the latent space representation 60 determined. The x-ray image 48 to be corrected, the mask 64 to be refined, and also the boundary condition 68 form the second input data for the generator network 69 provided as second trained processing function, which thus delivers the refined mask 71 as output data. In act S10, the refined mask 71, in order to achieve the brightness compensation, is corrected to the x-ray image 48 to be corrected, so that a corrected x-ray image 77 arises.

[0110] A denoising may then be applied to this corrected x-ray image 77 in order to further improve the image quality and in this way possibly to allow a further reduction of the x-ray dose.

[0111] FIG. 11 explains the functional layout of one embodiment of a provision system 40 that may also be referred to as a training system. The provision system 50 has a first training interface 78, via which the first training data (e.g., masks 53 with assigned physical parameters 54, 55) are accepted. In a first training unit 79, the autoencoder 56 is trained as described in accordance with FIG. 8 (cf., arrow 57). After this, in a rule determination unit 80, the assignment rule in accordance with act S3 may be determined.

[0112] A second training interface 81 in the present case involves an internal interface, via which the trained first processing function 63 may be provided as a combination of the assignment rule 62 and the trained decoder 59.

[0113] The provision system 40 now also has a third training interface 82, via which the second training data for the second stage is accepted (e.g., the x-ray images 65 and 66 as well as the first input data 67 (physical parameters 54, 55) in accordance with FIG. 9). A second training unit 83 trains the generator network 69 as the second processing function 70 using this second training data, as has been described in detail with regard to FIG. 9, in the sense of an adversarial training. The first processing function 62 has then been trained by representation learning. The correction algorithm, including the trained first processing function 62 and the trained second processing function 70 is then provided via a fourth training interface 84.

[0114] The correction system 39 in accordance with FIG. 12 includes a first connection interface 85, via which the correction algorithm may be accepted from the fourth training interface 84 of the provision system 40. X-ray images to be corrected 48 may be accepted via a second application interface 86, to which the corresponding system settings (e.g., the first and second physical parameters 54, 55) are then assigned accordingly. This makes it possible, in a mask determination unit 87, to determine the refined correction mask 71, in that the trained first and second processing functions 63, 70 are applied accordingly. In a correction unit 88, the mask 71 may then be used in order to correct the x-ray image 48 to be corrected. In the correction unit 88, a denoising of the corrected x-ray image 77 may also be undertaken in addition, before the image is provided to the third application interface 89.

[0115] The correction algorithm also contains the actual correction act (e.g., the addition of mask 71 and x-ray image 48 to be corrected).

[0116] Although the invention has been illustrated and described in greater detail by the exemplary embodiments, the invention is not restricted by the disclosed examples, and other variations may be derived herefrom by the person skilled in the art without departing from the scope of protection of the invention.

[0117] The elements and features recited in the appended claims may be combined in different ways to produce new claims that likewise fall within the scope of the present invention. Thus, whereas the dependent claims appended below depend from only a single independent or dependent claim, it is to be understood that these dependent claims may, alternatively, be made to depend in the alternative from any preceding or following claim, whether independent or dependent. Such new combinations are to be understood as forming a part of the present specification.

[0118] While the present invention has been described above by reference to various embodiments, it should be understood that many changes and modifications can be made to the described embodiments. It is therefore intended that the foregoing description be regarded as illustrative rather than limiting, and that it be understood that all equivalents and/or combinations of embodiments are intended to be included in this description.