Method and apparatus for reconstructing a CT image

Abstract

The invention relates to a method for reconstructing a CT image, comprising the following steps: providing CT recording data (RD); reconstructing overlapping sub-images (S1, S2, S3); determining a displacement vector (V) for the registration of overlapping regions of the sub-images (S1, S2, S3); interpolating a displacement vector field (VF) for each sub-image (S1, S2, S3) from the displacement vector sets (V) of the two flank regions; generating an output image data set (AD) on the basis of the CT recording data (RD) and the displacement vector field (VF); outputting the output image data set (AD). The invention also relates to a corresponding device, control apparatus and CT system.

Description

Technical Field

[0001] This invention relates to a method and apparatus for reconstructing CT images, and particularly to a method for correcting CT stacking artifacts. Background Technology

[0002] In CT imaging (computed tomography) of the chest region, motion such as heartbeat or lung changes caused by breathing can lead to inconsistencies in the recorded data. In CT images reconstructed from inconsistent recorded data, these inconsistencies manifest as artifacts in the form of "smeared" structures or discontinuities (such as abrupt changes in transitions).

[0003] These artifacts can be reduced, for example, by reducing the time required to cover the appropriate rotational angle interval in the recording through high rotational speeds or dual-source techniques. If the typical length scale that changes within this time interval is small compared to the voxel distance (approximately 1 mm) achieved later, the artifacts are also smaller but still present.

[0004] To reduce these artifacts, images are typically captured under limited motion conditions when the motion pattern is known. To achieve this, for example in cardiac recordings using CT, in addition to temporal resolution, controlling recording and reconstruction, particularly by analyzing the ECG signals acquired during recording, plays a crucial role. Therefore, data acquired, especially during periods when the heart is in a limited resting position, is typically recorded or incorporated into the reconstruction.

[0005] There are particular challenges when imaging large areas. The recording area of ​​a CT detector does not necessarily completely cover the desired recording area along the CT axis (e.g., the recording area of ​​the heart). Complete coverage can be achieved, for example, through continuous table feeds during recording (spiral CT) or through table feeds between consecutive images of a stationary patient (sequential CT). As a result, not all locations entering the image along the CT axis contain data from the same shooting time point. Especially when imaging the heart, it is possible that for two different locations, only data from different cardiac cycles exist. Conversely, for the same location, data from multiple cardiac cycles may exist. It is possible that recorded data from similar areas from multiple cardiac cycles may still have spatial differences, for example, due to superimposed respiratory movements or other differences between cardiac cycles that may not necessarily be present and identifiable in the ECG. Thus, despite high temporal resolution, inconsistent data may be generated, which may manifest, for example, as superimposed structures from multiple cardiac cycles visible in the reconstructed CT image. However, discontinuities between slices may also occur, for example, because parts of the patient are so strongly displaced that they cannot be detected by the detector area in any cardiac cycle.

[0006] In certain SOMATOM CT scanners, it is possible to perform so-called "true stacking" reconstruction, where even if a slice has recording data from multiple cardiac cycles, each CT slice is reconstructed only from a single, defined cardiac cycle. This generates a coherent stack of slices that can be precisely assigned to a specific cardiac cycle (i.e., a point in time). Multiple such stacks can then be generated sequentially. As mentioned above, within each stack, it can be assumed that motion-induced artifacts only occur when the time interval occupying that stack is too long.

[0007] If inconsistencies exist between stacks (e.g., due to respiratory motion), this can be manifested by discontinuities at the boundary between two stacks. These discontinuities can be characterized, for example, by structures or double structures along the CT axis.

[0008] An improved method for addressing discontinuities at stack boundaries is known from Lebdev et al.'s "Stack Transition Artifact Removal (STAR) in Cardiac CT" (Medical Physics 46(11), 4777-4791; November 2019). This method is based on determining a displacement vector field through image registration, thereby reducing discontinuities at stack boundaries by deforming the image structure corresponding to the thus obtained vector field. Furthermore, it utilizes the fact that stacking typically has a higher coverage than the final coverage shown along the CT axis. This results in overlapping regions of stacked image data, where image registration can then be performed. Summary of the Invention

[0009] The purpose of this invention is to provide an alternative and more comfortable method and apparatus for reconstructing CT images, thereby avoiding the aforementioned disadvantages.

[0010] This objective is achieved by the method, apparatus, control device, and computed tomography system according to the invention.

[0011] This invention is used for reconstructing CT images, particularly for CT stacking artifact correction methods. If reconstruction is performed via a CT system or diagnostic unit, this invention also controls the associated CT system or diagnostic unit. Specifically, this invention modifies image data to enable the creation of uninterrupted combined images from multiple individual stacks.

[0012] The method according to the present invention includes the following steps:

[0013] – Provides CT recording data, which includes multiple overlapping sub-image volumes.

[0014] – Reconstruct a working image dataset from CT recordings, wherein the working image dataset comprises multiple sub-images, each sub-image having an overlapping region with at least one other sub-image.

[0015] – Calculate the displacement vectors used to register overlapping regions of sub-images to each other, wherein a set of displacement vectors is assigned to each sub-image of the working image dataset for every two opposing side regions, wherein a predetermined set of displacement vectors is assigned to a side region if one side region is not registered with another side region.

[0016] – From the set of displacement vectors in the two lateral regions, interpolate the displacement vector field for each sub-image, where the two sides of the displacement vector field correspond to the obtained sets of displacement vectors for that sub-image, and perform interpolation of the displacement vectors between the two sides based on a predetermined transformation function from one set of displacement vectors to another.

[0017] – Generate an output image dataset based on CT recording data and displacement vector fields.

[0018] - Output this image dataset.

[0019] The CT recordings provided include unreconstructed data. This can be, for example, raw or preprocessed data. It can be provided by recording data using a CT scanner or, for example, by accessing recorded data on a storage device using a PACS (Picture Archiving and Communication System).

[0020] CT recordings comprise or consist of multiple records of overlapping sub-image volumes that collectively present a single subject. This is particularly relevant to records of cardiac regions, such as ECG-controlled CT recordings of the lungs or heart. Here, each record can be specifically assigned a separate motion state of its corresponding sub-volume, particularly a defined cardiac cycle.

[0021] The structure of CT recording data can be visualized in such a way that it comprises “raw sub-images” that combine to produce the entire subject. However, since CT data must be complexly reconstructed to produce images, these raw sub-images can be thought of as datasets of raw data or preprocessed raw data that will be used to generate sub-images through reconstruction.

[0022] A particular advantage of this invention lies in reconstructing images with motion artifacts. Here, the imaging should be performed in a "motion-triggered" manner, for example, by taking sub-volume images at predetermined time points, such as during a specific state of the cardiac cycle or during respiration. In particular, the motion should be known in advance, for example, because it is periodic or its course is well-known. However, in principle, the invention can also be used for other CT recording data, for example, to compensate for motion artifacts caused by interference from the CT scanner. CT recording data specifically contains image data that can be assigned to a unique, fixed cardiac cycle of the patient. Here, the recording of the patient's heart is performed over multiple cycles, typically 2 to 4 cycles (in the triggered sequence) or 5 to 8 cycles (in the case of a spiral). In both techniques, the recording ensures overlap of the sub-image data. This overlap is typically about 10% of the sub-volume length (up to 20% in a spiral), thus making it easy to find overlapping areas.

[0023] Sub-images (also known as "stacks") are now reconstructed from the CT recording data. These sub-images represent the volume of the captured sub-images (part of the overall subject) and together form the working image dataset. These sub-images (or the working image dataset) are not yet intended for diagnostic purposes but only for implementing the method. These sub-images do not necessarily have optimal resolution and can have a lower resolution than the output image dataset created later. The working image dataset preferably has properties required by the registration method that are unaffected by the CT user.

[0024] In particular, the fixed image resolution can be mentioned, for example, through the fixed predetermined selection of the reconstructed convolutional kernel and layer thickness, as well as the interlayer spacing and xy grid spacing that adapt to the voxels within the layer. There is also typically a minimum size requirement for image detail areas so that a sufficiently large image region can be provided for registration, even if the user selects very small image details. Standardization of the production of evaluable records is advantageous if the user cannot change (all) these characteristics to avoid accidentally altering critical diagnostic prerequisites.

[0025] Each sub-image in the working image dataset must overlap with at least one other sub-image. Here, a single sub-image overlapping only one unique sub-image should be located at an edge, while a sub-image not located at an edge (such as in the middle) should overlap with at least two sub-images. It is generally advantageous that all boundaries of the sub-images representing the desired portion of the subject constitute the overlapping region between at least two sub-images. An “overlapping region” or “overlap” refers to the partial volume of the recorded area (of the subject) in each relevant sub-image shown therein. The overlapping region should preferably be at least 1%, and particularly at least 5%, of the sub-image. The overlapping region is typically located at the edges of the sub-images.

[0026] When these sub-images are available, they are combined using this method. A first difficulty arises here: the sub-images often do not perfectly match in their overlapping areas because the subject changes between the captures of the sub-images, such as changes in the heart due to a heartbeat. To overcome these difficulties, a displacement vector can be calculated for the overlapping areas used to register the sub-images. In a rough approximation, the displacement vector can be considered to indicate which image points (or preferably which image values) must be shifted to where so that the sub-images match in their overlapping areas.

[0027] In this process, a set of displacement vectors is assigned to each sub-image of the working image dataset for every two opposing lateral regions. However, it is crucial to distinguish which sub-images are located at edges and which are not. If the lateral regions are overlapping, the displacement vectors can be simply calculated. If the lateral regions are boundaries (non-overlapping regions), a predetermined set of displacement vectors, such as the zero vector, is assigned to them.

[0028] So far, the sub-images are now fitted together, but they may be distorted in overlapping areas due to displacement vectors, causing artifacts to simply extend further from the image edges to the center of the sub-images. This problem is solved by implementing the following steps to adjust the entire sub-image.

[0029] To achieve "harmony" in the sub-image, the displacement vector field is interpolated using displacement vectors. The term "interpolation" refers to creating a continuous transformation from displacement vectors in one side region to those in another. Here, the two sides of the displacement vector field (in this case, the locations of the side regions of the sub-image) correspond to the various sets of displacement vectors obtained for that sub-image. Between these sets, the displacement vectors are interpolated based on a transformation function from one set of displacement vectors to another, thus creating a transformation from one side to the other. For example, all three spatial coordinates of the displacement vector between the sides can be calculated from the coordinates of two displacement vectors at the side using a weighted function, where the weighting takes into account the distances of the calculated displacement vector to two other displacement vectors at the side.

[0030] To ensure a harmonious image impression, it is preferable that adjacent vectors (not only in overlapping regions but also in other locations within sub-images) have a predetermined similarity, i.e., there is no over-adjustment, but rather a "smoothness" of the displacement vector field.

[0031] Therefore, by using a displacement vector field, sub-images can be corrected to fit adjacent sub-images, and an output image dataset can be generated from CT recording data. The output image dataset can now have the resolution required for diagnosis. Reconstruction can be performed by either distorting the sub-images with the help of the displacement vector field and then combining them, or by starting a new reconstruction of the CT recording data using the information from the displacement vector field.

[0032] The output image dataset is then output to a storage device (e.g., PACS) and / or a display unit, where it can be viewed by the diagnostician.

[0033] The device according to the invention comprises the following components:

[0034] – Data interface, designed to receive CT recording data, which includes records of multiple overlapping sub-image volumes.

[0035] – A working reconstruction unit, designed to reconstruct a working image dataset from CT recording data, wherein the working image dataset comprises multiple sub-images, each sub-image having an overlapping region with at least one other sub-image.

[0036] – A displacement unit, designed to obtain displacement vectors for aligning overlapping regions of sub-images to each other, wherein a set of displacement vectors is assigned to each sub-image of the working image dataset for every two opposing side regions, wherein a predetermined set of displacement vectors is assigned to a side region in the case where one side region is not aligned with another side region.

[0037] – A vector field unit is designed to interpolate the displacement vector field for each sub-image from the set of displacement vectors in the two lateral regions, wherein the two sides of the displacement vector field correspond to the obtained sets of displacement vectors for the sub-image, and the displacement vectors between the two sides are interpolated based on a predetermined transformation function from one set of displacement vectors to another.

[0038] – The second reconstruction unit is designed to generate an output image dataset based on CT recording data and displacement vector fields.

[0039] – Data interface, which is used to output the output image dataset.

[0040] The control device for controlling a computed tomography system according to the present invention is designed to perform the method according to the present invention and / or include the means according to the present invention.

[0041] The computed tomography system according to the present invention includes a control device according to the present invention.

[0042] Most of the aforementioned components of the device can be implemented, wholly or partially, as software modules within the processor of a corresponding computing system, such as the control equipment of a computed tomography system. This primarily software-based implementation offers the advantage that even previously used computing systems can be easily retrofitted via software updates to operate according to the invention. In this regard, this objective is also achieved by a corresponding computer program product having a computer program that can be directly loaded into the computing system, comprising program segments that, when run on the computing system, perform the steps of the method according to the invention (at least computer-executable steps). In addition to the computer program, such a computer program product may, if necessary, include additional components and / or additional parts such as documentation, or hardware components such as hardware keys (dongles, etc.) for using the software.

[0043] For transmission to or storage within a computing system or control device, a computer-readable medium, such as a memory stick, hard disk, or other removable or fixed-mount data carrier, may be used, on which program segments of a computer program are stored that can be read and executed by the computing system. For this purpose, the computing system may, for example, have one or more cooperating microprocessors.

[0044] Furthermore, other particularly advantageous designs and modifications of the invention are described below, in which claims of one claim class may be modified in a manner similar to claims and descriptions of another claim class, and in particular, individual features of different embodiments or variations may be combined to form new embodiments or variations.

[0045] According to a preferred method, a two-dimensional boundary layer in the overlapping region is determined to obtain a displacement vector for registration of the overlapping region between the first and second sub-images. The boundary layer is preferably planar, and particularly preferably arranged perpendicular to the CT axis, especially at its center within the boundary of the overlapping region.

[0046] Here and below, the CT axis is considered to be the axis along which the stage is fed, or the axis along which the patient moves relative to the CT scanner. Therefore, when multiple sub-images to be combined are captured, the CT axis is the axis that extends from top to bottom through the sub-images, that is, substantially orthogonally through the overlapping area.

[0047] Preferably, the boundary layer is first determined by assuming a set of displacement vectors with predetermined values ​​(e.g., zero), and the position of the boundary layer along the CT axis is determined such that the squared difference of the image values ​​(HU values ​​of image points in each image) is as small as possible from the outset. Advantageously, a displacement vector of zero length is started, and the difference between two images from each stack is examined, which should in principle be "identical" and congruent. If this is not the case (e.g., due to actual offset between the two images), the two images are moved until the difference is zero or at least minimized. Here, the offset vector is generated in particular by gradient descent, with minimizing the cost function ("difference between the two images") as the objective.

[0048] It should be noted that this method can provide particularly good results when the boundary layer location is chosen appropriately. Thereafter, it is preferred to optimize the individual displacement vectors, particularly by minimizing the residual squared difference. This should be done, in particular, by considering the regularization term in the cost function.

[0049] According to a preferred method, a regular 2D grid of sampling points is defined in the overlapping region of the sub-images, and a displacement vector is assigned to each sampling point, which points from the sampling point to a position in the sub-image, and the image value should be assigned to the origin position of the displacement vector.

[0050] For example, a plane (boundary layer) that should coincide in the two sub-images is defined, for instance, precisely at the center of the overlapping region of sub-images S1 and S2. Then, a displacement vector V with an origin R (on the boundary layer) specifies that the value of voxel R+V should be assigned to the value R in the boundary layer. Thus, in this example, the location pointed to by the displacement vector serves as the source of the comparison input to the boundary layer, thereby helping to minimize the aforementioned differences.

[0051] Even though mathematically a vector is related to displacement from one point to another, in the field of registration, it is common practice to move the vector away from the relevant point r to represent the desired deformation. Thus, the transformed image region I' (after deformation of image region I) is determined by the vector v, where I'(r) = I(r+v). Generally, there is a deformation model r' = D(r) that provides a corresponding point for each point in space, such that the deformed image I'(r) = I(D(r)). Therefore, in the specific case considered here, D(r) = r+v(r) is the displacement of vector v from point r (e.g., the sampling point) towards point r+v. Here, the resulting image impression indicates that the image point has moved from r+v towards r. The background is that in practice, the point r for which the deformed image values ​​are typically specified is usually chosen. Here, it is preferably located on a regular grid. In other methods (starting from the originally regularly arranged grid points and "moving" them), it may result in an irregularly arranged grid of image points after deformation. This yields a significant advantage: no interpolation is required, but it necessitates converting the deformed mesh back to a regular mesh, or performing interpolation within the irregular mesh. This type of interpolation is far more complex than directly interpolating within a still regularly arranged mesh and then directly obtaining a regularly arranged mesh.

[0052] The side regions of the sub-image are preferably defined by sampling points, and a set of displacement vectors are assigned to the side regions, including displacement vectors corresponding to these sampling points.

[0053] The sampling points are preferably arranged at equal intervals and / or form the aforementioned boundary layer. For each of the two sub-images defining the boundary layer, each sampling point preferably has a vector indicating its position in the respective stack, wherein the displacement vector is obtained by means of an image registration method. A shifted image point is obtained from each displacement vector, wherein the displacement vector can start from the center point or a corner of the image point. Each displacement vector should point exactly to an image point in the stack, and each shifted image point should lie on the corresponding mapping of the same image point in the other sub-images.

[0054] According to a preferred method, a displacement vector is selected in the overlapping region between the first sub-image and the second sub-image, such that sampling points of the first sub-image are assigned a first displacement vector in the overlapping region, and corresponding sampling points of the second sub-image are assigned a second displacement vector. Here, the term "corresponding" means that the sampling points respectively display the same image area of ​​the subject. Here, the second displacement vector preferably represents the inverse vector of the first displacement vector. Here, the inverse vector is a vector with equal component values ​​but opposite signs.

[0055] Preferably, a predetermined metric of the similarity between the image regions of the two sub-images is used to determine the two displacement vectors, specifically based on the squared difference of the image values. However, as an alternative to the squared difference, other metrics can be used to assess similarity, such as local cross-correlation (LCC) or interaction information (MI).

[0056] Particularly preferably, based on the current image values ​​and image value gradients of the two sub-images and a predetermined step size at each shifted sampling point, a minimization vector is progressively determined using an iterative method and added to the displacement vector, wherein the minimization vector is based on the aforementioned metric. Thus, in each step n, there exists an initial displacement vector v. n And after this step, the next displacement vector v is obtained. n+1 For v n+1 =v n +u n , where u n This is the minimized vector determined in step n.

[0057] In practice, it is noteworthy that very strong low-frequency fluctuations in image values ​​(e.g., those with units of HU) can be identified proportionally below predetermined boundary values ​​(which are quite low in practice). These are considered undesirable artifacts. For example, these artifacts can be seen in areas of air outside the patient and in areas of the lungs with high ventilation. These low-frequency fluctuations cause gradients in the image, which strongly affect the determination of the aforementioned minimization vector. Here, it is preferable not to ignore (or process) these image regions, but simply to assume “saturated image values” instead of actual image values. Here, saturation in the sense of a step function can be very difficult to achieve (i.e., all image values ​​below the boundary value are simply set to the boundary value, and all image values ​​above the boundary value remain unchanged) or a function that makes the transition to the boundary value continuously differentiable can be used, such as using an error function, a sigmoid function, or a trigonometric function. The aim is to set the differences in fluctuations of image values ​​below the threshold and their gradients as close to zero as possible, so that they no longer affect the determination of the minimization vector.

[0058] Therefore, according to a preferred method, the displacement vector of an image region in a sub-image whose image value is below a predetermined boundary value is set to a predetermined value (e.g., adjusted to the boundary value as described above or adapted according to a function). The expression "its image value" refers to the value of the image region (e.g., a voxel) shifted by the displacement vector. This has the advantage of requiring less computation time even for irrelevant regions that have a minor impact on the result.

[0059] The boundary value is preferably located above the air CT value, particularly above -1000 HU and preferably below -800 HU. This method is particularly advantageous in determining the aforementioned minimized vector. It has the particular benefit that low-frequency image value fluctuations no longer cause severe distortion.

[0060] According to a preferred method, when interpolating the displacement vector of the displacement vector field of image regions (e.g., voxels) between side regions,

[0061] – Determine the distance from the relevant image region to the lateral region.

[0062] – Determine the ratio of distances to lateral regions, particularly the weighted ratio, where closer lateral regions preferably receive higher weights.

[0063] – The displacement vector is calculated by adding the first displacement vector in one side region of the subimage and the second displacement vector in another side region, and the ratio thereto.

[0064] Preferably, the displacement vector found in the overlapping region (e.g., at the boundary layer) continues to extend towards the center of the associated sub-image (i.e., towards the other side region) with a predetermined attenuation from the outside. During this attenuation process, the length of the displacement vector decreases to a minimum (e.g., zero). This ensures a continuous transition between the overlapping region (e.g., the boundary layer) and the remaining sub-images. The length of this attenuation transition can be determined or predetermined by the user based on experience. This length is preferably between 5 mm and 10 mm.

[0065] According to a preferred method, the displacement vectors in the displacement vector field between the overlapping regions and / or the lateral regions are smoothed. Preferably, in this smoothing process, smoothing convolution or low-pass filtering, particularly convolution with a Gaussian window, is performed in each of the aforementioned iterative steps used to determine the displacement vectors in the overlapping regions (particularly the boundary layer). Alternatively or additionally, it is preferred that convolution with a Gaussian window be performed in the overlapping regions, particularly in their boundary layers, before interpolating the displacement vectors of the two lateral regions of the displacement vector field.

[0066] There are different types of displacement vector interpolation, each with different frequency characteristics. Preferably, displacement vector interpolation is performed at an arbitrary point in the space between two boundary layers. The two boundary layers are first smoothed by convolution (optionally), the smoothing width being, for example, proportional to the distance between the boundary layers and the point to be interpolated. The position of this point is then first vertically projected onto both boundary layers. Bilinear interpolation is then performed at the projected location, resulting in two image values. These two image values ​​are then weighted and summed under a boundary condition with a total weight of 1. Since the weights are linearly related to the distance between the boundary layers, an equivalent value of linear interpolation between the boundary layers is obtained. Overall, this can be considered as trilinear interpolation. In particular, if the sampling points on the (2D) boundary layers are located at the same xy position, the process can also be considered as trilinear interpolation in a 3D mesh, where the mesh in the z-direction comprises only two mesh points respectively.

[0067] According to a preferred method, within the scope of generating the output image dataset, sub-images are first reconstructed, then the image regions of the sub-images are moved according to their assigned displacement vector fields, and finally the sub-images are combined.

[0068] According to an alternative preferred method, particularly considering the displacement vector field in the reconstruction process within the scope of back projection, weighting is performed on the corresponding contributions of relevant sub-images when the region of the output image dataset can be formed by information from more than one sub-image.

[0069] Preferably, the output image dataset has a higher image resolution than the working image dataset. This allows the method to be executed faster.

[0070] According to a preferred method, the output image dataset and / or sub-images are reconstructed by weighting different time fractions of the CT recording data. In this case, weight values ​​are preferably determined for multiple image points (e.g., voxels) of the output image dataset or sub-images, with each displacement vector assigned exactly one weight. For this purpose, a 3D image point grid is particularly preferred, with each grid point or group of grid points assigned a weight value. In this way, portions of the CT recording data that are more accurate than others can be considered in the reconstruction with higher weights.

[0071] A preferred method can be applied when, in the overlapping region of a first sub-image and a second sub-image, there are image points in the first sub-image that cannot be reconstructed or can only be reconstructed with limitations due to missing data. If corresponding data exists in the second sub-image, these missing data are preferably supplemented based on the data and displacement vector field of the second sub-image to reconstruct the pixels of the first sub-image. Alternatively, relevant image points from the second sub-image (rather than unfavorable data from the first sub-image) are preferably used in the reconstruction of the output image dataset.

[0072] In 3D geometries with fan-shaped detector geometry, the data present in the ECG-triggered recordings for reconstruction is typically less than that required for 360° coverage. Therefore, image voxels see a complete set of 180° readings within the parallel geometry, resulting in a complete image contribution. These voxels do not fill a simple cuboid geometry but only within the inner core, and then only partially within the complex geometry. Here, the principle of reconstructing voxels from information in other sub-images or the output image dataset applies.

[0073] A preferred specific method for practical application includes the following steps:

[0074] 1. ECG-controlled CT recording of the heart (as spiral CT or sequential CT recording).

[0075] 2. Reconstruct the registration image dataset from the CT recording data, from which the coherent sub-image volumes of all stacks contained in the recording (sub-images are referred to as such below, as they are called in practice) can be extracted, wherein a single defined cardiac cycle can be assigned to each stack, wherein the registration image dataset (multiple stacks or working image datasets) can have characteristics that are not affected by the CT user, especially required by the registration method.

[0076] 3. Determine the overlapping regions of the continuously stacked structures, and determine the (two-dimensional) boundary layer in each overlapping region.

[0077] 4. Determine the displacement vector starting from the sampling points arranged at equal intervals within the boundary layer, where for each of the two stacks defining the boundary layer, there is a vector for each sampling point, which shows the position in each stack, wherein the displacement vector is obtained by means of an image registration method.

[0078] 5. Generate displacement vectors and corresponding cardiac cycles for any image point in the space above and between boundary layers (i.e., the displacement vector field), where all image points in the same layer along the CT axis always originate from the same cardiac cycle.

[0079] 6. A final image dataset is generated by considering the center of image points shifted according to the displacement vector, wherein the final image dataset has specific attributes determined by the CT user.

[0080] Regarding step 2:

[0081] In the product implementation, two CT image volumes are reconstructed using the same image point grid. Thus, each layer of this grid along the CT axis is reconstructed twice. If a layer contains recording data from two cardiac cycles, the recording data from the first cardiac cycle is input into the layer of the first volume, and the recording data from the second cardiac cycle is input into the layer of the second volume. Therefore, the image data from the two cycles of that layer exist independently. Furthermore, for subsequent stack extraction, information about which cycle each layer originates from is retained. The stack (or a portion thereof) can then be recovered from the two image volumes through the corresponding reconstruction of the layers from the two image volumes.

[0082] The following implementation method is also preferred, wherein the complete image volume is reconstructed specifically for each cardiac cycle. The stacked image data is then directly separated into individual image volumes through reconstruction.

[0083] In the reconstruction of the registered image dataset, reconstruction is performed in a manner that makes the image data particularly suitable for the registration purpose. For this, a predetermined image resolution and appropriate image detail are particularly relevant. The predetermined image resolution is achieved by specifying a pre-defined reconstruction kernel within the layers to achieve a favorable image point spacing. The layer thickness and inter-layer spacing are also predetermined. Here, distance selection is performed by considering the sampling theorem to obtain aliasing-free sampling of the image signal. The image detail is selected as a "safe range" around the desired image detail for the CT user. Thus, inaccuracies arising in image registration at edge regions due to subsequent convolution operations at image boundaries are minimized by keeping the image detail away from the user's image detail.

[0084] Regarding step 3:

[0085] The overlapping region can be determined from the known boundaries of the stack. In the product implementation, a boundary layer perpendicular to the CT axis is selected at the center position within the boundary of the overlapping region.

[0086] Regarding step 4:

[0087] For each boundary layer, image registration is performed between the two stacks of the boundary layer.

[0088] In the boundary layer, a regular 2D grid of sampling points is assumed. A displacement vector is assigned to each sampling point, pointing from its center to its location in the first stack. The location in the second stack is correspondingly determined by a vector with equal component values ​​but opposite signs compared to the displacement vector to the first stack. The goal of registration is to match the image values ​​at the two locations. Here, the square of the difference is considered as a metric for matching. Another boundary condition is that adjacent vectors in the layer should be similar, thus avoiding over-adjustment to ensure a "smooth" displacement vector field.

[0089] Here, in the iterative method, preferably, a minimization vector is determined at each step based on the current image value and image value gradient at the shifted sampling point, along with a predetermined step size, and added to the displacement vector that minimizes the square of the difference. Furthermore, to ensure field smoothness, a smooth convolution or low-pass filtering (e.g., convolution with a Gaussian window) is performed on the vector field at each iteration step. In the method used here, the step size and the width of the Gaussian window are adjusted (reduced) at each iteration step. This aims to allow larger differences in image values ​​at the beginning to be compensated for by larger displacements in larger regions. As the number of steps increases, finer details can be adjusted.

[0090] Another adjustment is to limit the generated displacement vector. Here, an appropriate mapping function with "saturation properties" is used to ensure that the displacement vector can only exceed the stacking boundary within a limited range.

[0091] If the displacement vector still exceeds the stack boundary, it is preferable to continue assuming that the image value at the stack boundary is perpendicular to the boundary.

[0092] Regarding step 5:

[0093] The displacement vectors are known at the regularly arranged sampling points on the boundary surfaces. If a displacement vector is needed at an image point different from these sampling points, it must be generated from the known displacement vectors. First, it must be determined which boundary layers the desired point lies between. It is possible that it is not between boundary layers, but rather "before" or "after" the "first" or "last" boundary layer. Then, it is assumed that there is a boundary layer at the beginning or end of the final image volume, where all displacement vectors are equal to zero. From this location, a clear assignment is also obtained regarding which stack the image point belongs to. It is also known which vectors must be considered on the boundary layers, because at this point there are two distinct vectors for each image point, one pointing to the "before" stack and the other to the "after" stack.

[0094] Then, the distance from the image points to the boundary surface is determined. The weighting ratio used to sum the values ​​of the two boundary layers is then derived from the ratio of these distances. Here, it applies that closer layers receive higher weights. This is essentially equivalent to linear interpolation between the two boundary layers, at least under the further condition that the sum of the weights is 1 and the weights are linearly related to the distance. Of course, other forms of weighting can generally be used, and the sum does not necessarily have to be 1.

[0095] Within the boundary layer, interpolation (bilinear) is also performed at the locations of image points projected onto the boundary layer. This yields a trilinear interpolation of the displacement vector between the two boundary layers.

[0096] However, interpolation itself can result in an "unnatural" image appearance. Strong distortion at the boundary layer propagates very uniformly along the CT axis. A more natural image appearance can be achieved through additional smoothing. This is done by performing boundary layer convolution with a Gaussian window before interpolation. The spatial width of the window increases (e.g., linearly) with the distance from the image point to the boundary layer. This significantly "relaxes" the displacement vector field as the distance to the boundary layer increases.

[0097] Regarding step 6:

[0098] As a specific implementation method, two variant schemes are preferred here.

[0099] In the image-based method, two image volumes are generated based on the reconstruction in step 2). The reconstruction is performed according to the CT user's presets. The image points in these two volumes exist within a regular sampling grid according to these presets. To obtain the final image dataset, the center of each image point is shifted by a known heart rate cycle based on the displacement vector obtained in step 5), and then interpolation is performed between the two image volumes in the image data at the position of the shifted center, using the heart rate cycle assigned to that image point. Noise-preserving interpolation is preferably used in the interpolation to avoid locally altering the image impression (sharpness and noise characteristics).

[0100] In the reconstruction-based method, the displacement vector from step 5 is used directly so that the displacement center of the image points is already considered in the reconstruction of the CT image data (e.g., through filtered backprojection). Furthermore, information about the selected cardiac cycle of the image points can be directly considered through appropriate weighting of the recording region. However, the final image points are then displayed on a conventional grid. In this variant, the final image volume with characteristics preset by the CT user is directly generated from the reconstruction. This variant requires a more complex reconstruction implementation because the displaced image points must be considered. However, when implementing this method, the implementation time is nearly halved compared to the image-based method because only half the number of image points need to be generated.

[0101] Typically, preferred practices can be described as follows:

[0102] a. Create CT recordings such that sub-image volumes (stacks) can be reconstructed from the CT recording data, where each stack is characterized by different weights of the time fractions of the CT recording data.

[0103] b. Estimate spatial allocation, which allows the positions of image elements in one stack to be assigned to those image elements in another stack (if they exist).

[0104] c. Generate a set of weights and displacement vectors for each image point in the final 3D image point mesh, wherein the shift center of the image point is obtained from each displacement vector, wherein each displacement vector points precisely to an image point in a stack, and wherein each vector is assigned exactly one weight, and wherein each shift center points to a mapping of the same element in a different stack.

[0105] d. Reconstruct the final image volume, where an image value and shift center are determined for each final image point, and these image values ​​are summed in a weighted manner to obtain the final image value of the image point, which is displayed in a regular grid of the final image volume.

[0106] The steps of the more specific method described above can be mapped to the steps of the more general method as follows: 1. and 2. correspond to a., 3. and 4. correspond to b., 5. correspond to c., and 6. correspond to d.

[0107] More specific methods are envisioned for applications such as ECG-controlled recording of cardiac images. More general methods can also be used in other applications, such as respiratory-controlled CT recordings. Here, for example, a respiratory signal is present, which allows the recorded data to be "divided" into stacks. However, ECG signals may be missing, thus discontinuities may also occur, particularly in the cardiac region. Furthermore, the spatial arrangement of thoracic organs cannot be reliably reproduced at all times using respiratory signals, resulting in completely different spatial arrangements even with identical respiratory signals, and consequently, discontinuities may arise.

[0108] A particular extension is that data integrity can be determined in addition to the actual image values. Thus, it's possible that in the edge regions of stacked image volumes, there exist image points that can still be approximately reconstructed for correct reconstruction, but the range of CT rotation angles available through weighting is limited, resulting in incompleteness. It is feasible to supplement these image points with the portion of the rotation angles available in the weighting window from adjacent layers. The supplemented data integrity then allows for the identification of these regions and their avoidance when necessary. However, supplementing these image points in this way is advantageous due to edge effects.

[0109] Regarding step b:

[0110] Generally, various registration methods are applicable here. Image registration is preferably performed throughout the overlapping region or even beyond. Therefore, registration or spatial assignment is achieved not only for the boundary layers but also beyond. This can be used, for example, to consider not only one stack of recorded data in the image points, but at least two. This improves the dose efficiency of CT recordings. It is also preferred to use a deep learning-based registration method trained to establish the spatial assignment.

[0111] Furthermore, it is preferable that when determining the boundary layer, all displacement vectors are initially zero, but the position of the boundary layer along the CT axis is determined such that the squared difference is as small as possible from the outset. Therefore, the smallest possible displacement vector can then be expected, which will produce a more natural overall image impression.

[0112] Even though the symmetry of the displacement vectors in the boundary layer is preferred, i.e., the displacements toward the second stack are exactly opposite in magnitude to the displacements toward the first stack, it is advantageous for the image impression to determine the two vectors independently or at least with weaker coupling. However, this increases the challenge regarding stability due to the increased number of free parameters.

[0113] If data integrity is available, it can be considered, for example, as a boundary condition in the registration method.

[0114] Regarding step c:

[0115] In summary, it is possible that the recorded data from multiple stacks can influence the final image points. To do this, for example, the central position of the image must be found by determining the centroid, based on the spatial allocation of common image elements in the stacks. Then, starting from the centroid, the displacement vector of the image element in all stacks where the image element can be found can be determined. These centroids must then be transferred to a regular grid so that a vector set of the image element can be generated at each grid point.

[0116] More specifically, it is preferable to register multiple layers (i.e., 3D meshes instead of 2D meshes) in the overlapping region. This provides a denser set of displacement vectors in the overlapping region, from which the displacement vector field can then be interpolated.

[0117] Regarding step d:

[0118] Reconstruction can be performed again by interpolating at the shift center point location in image space or by considering the shift center point in CT reconstruction (e.g., through backprojection). Preferably, more than one stack contributes to the final image value. However, for each contributing stack, a similar approach can be used, and the final image value is created as a weighted sum of the amounts of a single stack. For example, it is conceivable that, with very low pitch in helical recording, even more than two cardiac cycles could be covered at one location along the CT axis.

[0119] All the features of the practical method described herein can also be applied to general methods.

[0120] For the method according to the invention, an AI-based approach (AI: "Artificial Intelligence") is preferably used. Artificial Intelligence is based on the principles of machine learning and typically utilizes a learning algorithm that has been trained accordingly. Machine-based learning is generally referred to as machine learning, which also includes the principles of deep learning. For example, a deep convolutional neural network (DCNN) is trained to recover a high-dose condition from a low-dose CT image (thus having low noise). Here, the high-dose condition is known during training. However, the problem here is the error estimation in subsequent applications to unknown data, because it is impossible to be perfect no matter how it is trained, and it is generally difficult to prove generalization.

[0121] The components of this invention are preferably used as a "cloud service." This cloud service is used for data processing, particularly with the aid of artificial intelligence, but can also be a service based on traditional algorithms or a service where human evaluation is performed in the background. Typically, a cloud service (hereinafter also simply "the cloud") is an IT infrastructure in which storage space or computing power and / or application software are provided via a network, for example. Here, communication between the user and the cloud occurs through a data interface and / or data transmission protocol. In the present context, it is particularly preferred that the cloud service provides not only computing performance but also application software.

[0122] Within a preferred approach, data is provided to cloud services via a network. Cloud services include computing systems, such as computer clusters, which typically do not include the user's local computer. The cloud may be provided, particularly by healthcare institutions that also provide medical technology systems. For example, image acquisition data is sent to a (remote) computer system (cloud) via a RIS (Radiology Information System) or PACS. The cloud's computing system, network, and medical technology system preferably constitute a network in a data technology sense. Here, the method can be implemented via instructions within the network. The data computed in the cloud ("result data") is then sent to the user's local computer via the network. Attached Figure Description

[0123] The invention will now be explained in more detail with reference to the accompanying drawings and embodiments. Here, in the different drawings, the same parts have the same reference numerals. The drawings are generally not drawn to scale. Wherein:

[0124] Figure 1 A rough schematic diagram of a computed tomography system with a control device according to an embodiment is shown, the control device having means for performing a method according to the invention.

[0125] Figure 2 The recording of ECG control for the sub-image is shown.

[0126] Figure 3 A schematic diagram of the method according to the present invention is shown.

[0127] Figure 4 Three sub-images with boundary layers and displacement vectors are shown.

[0128] Figure 5 The interpolation of the displacement vector field is shown.

[0129] Figure 6 A comparison of an output image dataset with an image dataset based on existing techniques is shown.

[0130] Figure 7 A comparison of an output image dataset with an image dataset based on existing technologies is shown.

[0131] In the following description, it is assumed that the imaging system is a computed tomography (CT) system. However, in principle, this method can also be used with other imaging systems. Detailed Implementation

[0132] Figure 1 A computed tomography (CT) system 1 with control equipment 11 for performing the method according to the invention is shown schematically. The CT system 1 typically includes a scanner 2 with a gantry, X-ray sources 3 rotating within the gantry and irradiating a patient, who is pushed into the measurement chamber of the gantry by means of a bed 5, so that the rays are incident on detectors 4 respectively opposite to the X-ray sources 3. It should be clearly noted that the embodiment according to the figures is merely one example of CT, and the invention can also be used in any CT configuration, such as those with a ring-shaped fixed X-ray detector and / or multiple X-ray sources.

[0133] Similarly, only components essential to the explanation of the invention are shown in control device 11. Essentially, this CT system and associated control device are known to those skilled in the art and therefore do not require detailed explanation.

[0134] Here, the core component of the control device 11 is a processor, on which various components are implemented as software modules. In addition, the control device 11 also has a terminal interface 14, where a terminal 20 is connected, allowing the operator to operate the control device 11 and thus the computed tomography (CT) system 1. Another interface 15 is a network interface for connecting to the data bus 21, thereby establishing a connection with a RIS (Radio-Information System) or PACS (Picture Archiving and Communication System).

[0135] The scanner 2 can be controlled by the control device 11 via the control interface 13, for example, controlling the rotation speed of the gantry, the displacement of the patient bed 5, and the X-ray source 3 itself. Raw data RD is read from the detector 4 via the acquisition interface 12. Furthermore, the control device 11 has a memory unit 16 that stores various measurement protocols.

[0136] As a software component, a measurement control unit is implemented on the processor. This measurement control unit is controlled via control interface 13 based on one or more selected measurement protocols that have been modified by the operator through terminal 20 as necessary, in order to perform measurements and acquire data.

[0137] Another component on the processor is an image data reconstruction unit 18, which reconstructs the desired image data from the raw data RD received through the data acquisition interface 12. Here, the image data reconstruction unit 18 is designed as an apparatus 18 according to the invention and includes the following components. The functions of these components can be found in the description of the method (see...). Figure 3 ).

[0138] Data interface 6 is used to receive CT recording data RD, which includes records of multiple overlapping sub-image volumes. Such records are as follows: Figure 2 As shown.

[0139] Working reconstruction unit 7 is used to reconstruct working image dataset A from CT recording data RD, wherein working image dataset A includes multiple sub-images S1, S2, S3 (e.g. Figure 2 As shown), each sub-image S1, S2, S3 has an overlapping region U with at least one other sub-image S1, S2, S3.

[0140] Displacement element 8 is used to obtain the displacement vector V for mutual registration of overlapping regions of sub-images S1, S2, and S3 (see, for example, [link to relevant documentation]). Figure 4 ), wherein a set of displacement vectors V is assigned to each sub-image S1, S2 and S3 of the working image dataset A for every two opposing side regions, wherein a predetermined set of displacement vectors V is assigned to a side region if a side region is not registered with another side region.

[0141] Vector field unit 9 is used to interpolate the displacement vector field VF for each sub-image S1, S2, S3 from the displacement vector group V of the two side regions (see example). Figure 5 The displacement vector field VF corresponds to the obtained sets of displacement vectors V in sub-images S1, S2 and S3 on both sides, and the displacement vector V between the two sides is interpolated based on a predetermined transformation function from one set of displacement vectors V to another set of displacement vectors V.

[0142] The second reconstruction unit 10 is used to generate the output image dataset AD based on the CT recording data RD and the displacement vector field VF. Alternatively, theoretically, the working reconstruction unit 7 can also be used if the appropriate design is implemented.

[0143] Here, the data interface 6 used for receiving data can also be used to output the output image dataset AD. However, the terminal interface 14 can also be used.

[0144] Figure 2 The ECG-controlled recordings of sub-images S1, S2, and S3 are shown. The ECG heart curves are displayed, with rectangles and arrows indicating recording time points. During recording, raw data RD is recorded, which can be reconstructed into sub-images S1, S2, and S3 as shown in the figure.

[0145] Figure 3 A schematic diagram of a method for reconstructing CT images according to the present invention is shown.

[0146] In step I, CT recording data RD is provided, which includes records of multiple overlapping sub-image volumes, such as... Figure 2 As shown.

[0147] In step II, the working image dataset A is reconstructed based on the CT recording data RD, wherein the working image dataset A includes multiple sub-images S1, S2, S3, and each sub-image S1, S2, S3 has an overlapping region U with at least one other sub-image S1, S2, S3.

[0148] In step III, the displacement vector V is obtained for mutual registration of the overlapping regions of sub-images S1, S2, and S3 (see example). Figure 4 ), wherein a set of displacement vectors V is assigned to each sub-image S1, S2, S3 of the working image dataset A for every two opposing side regions, wherein a set of predetermined displacement vectors V is assigned to the side region if the side region is not registered with the other side region.

[0149] In step IV, the displacement vector field VF is interpolated from the displacement vector group V of the two side regions for each sub-image S1, S2, S3 (see also...). Figure 5 The displacement vector field VF corresponds to the two sides of the sub-images S1, S2, S3, and the displacement vector V between the two sides is interpolated based on a predetermined transformation function from one set of displacement vectors V to another set of displacement vectors V.

[0150] In step V, the output image dataset AD is generated based on the CT recording data RD and the displacement vector field VF.

[0151] In step VI, the output image dataset AD is output.

[0152] Figure 4Three sub-images S1, S2, and S3 are shown, each having a boundary layer G and a displacement vector V in the overlapping region U of each sub-image S1, S2, and S3. Here, the lateral regions are located at the upper and lower parts, respectively, and are not labeled because they are only part of the sub-images S1, S2, and S3. For example, the upper edge and the overlapping region U with the second sub-image, or the boundary layer G in that overlapping region U if necessary, can be considered as the lateral region of the upper sub-image S1. Generally, the term "lateral region" is only intended to combine cases where there are two overlapping regions U or one overlapping region and one edge region (non-overlapping).

[0153] Displacement vectors V originate in pairs from sampling points T in the planar boundary layer, and move away from the boundary layer G (opposite to each other). These displacement vectors point in pairs to the image regions in their corresponding sub-images S1, S2, and S3, with the upper displacement vectors pointing to the upper sub-images S1 and S2, and the lower displacement vectors pointing to the lower sub-images S2 and S3. The pointed-to image regions are those in sub-images S1, S2, and S3 whose image values ​​should be assigned to the corresponding scan points T of the foot points forming the vector pairs. Therefore, these related image regions should display the same area of ​​the recorded subject.

[0154] This presents a challenge: obtaining results given the conditions (opposite vector pairs, a planar boundary layer G in the middle of the overlapping region U). Thus, the only degree of freedom is the vector coordinates, which are constrained by conditions pointing to the same region as the subject. However, this can be achieved through appropriate minimization methods.

[0155] Figure 5 It shows according to Figure 4 The displacement vector sets V of the two lateral regions of the intermediate sub-image S2 are interpolated with the displacement vector field VF. A boundary layer G is visible above and below, from which the two sets of displacement vectors V originate. Between these layers, additional displacement vectors V are interpolated between the two lateral regions, thus representing a smooth transition from one set of displacement vectors V to the other.

[0156] Figure 6 and Figure 7 The images show a comparison between the output image dataset AD (right) after “real stacking” reconstruction and the image dataset according to existing techniques (left). In the left image, a horizontal “line” can be seen along the arrow, indicating a significant offset between the corresponding upper and lower sub-images. This offset is no longer visible in the right image.

[0157] Finally, it should be reiterated that the methods described in detail above and the computed tomography system 1 shown are merely embodiments, and those skilled in the art can modify them in various ways without departing from the scope of the invention. Furthermore, the use of the indefinite articles "a" or "an" does not preclude the possibility that the described features may be repeated. Similarly, the terms "unit" and "module" do not preclude the possibility that a related component may consist of multiple interacting sub-components, which may also be spatially distributed if necessary. The word "multiple" should be understood as "at least one".

Claims

1. Method for reconstructing a CT image, comprising the following steps: - providing CT recording data (RD) comprising recordings of a plurality of overlapping sub-image volumes; - reconstructing a working image data set (A) from the CT recording data (RD), wherein the working image data set (A) comprises a plurality of sub-images (S1, S2, S3), wherein each sub-image (S1, S2, S3) has an overlap region (U) with at least one further sub-image (S1, S2, S3); - determining displacement vectors (V) for registering the overlap regions of the sub- images (S1, S2, S3) with each other, wherein for each sub-image (S1, S2, S3) of the working image data set (A) of each two opposing side regions a set of displacement vectors (V) is assigned, wherein in case one side region is not registered with another side region, a set of predetermined displacement vectors (V) is assigned to the side region; - interpolating a displacement vector field (VF) for each sub-image (S1, S2, S3) from the sets of displacement vectors (V) of two of the side regions, wherein the two sides of the displacement vector field (VF) correspond to the determined sets of displacement vectors (V) of the sub-image (S1, S2, S3) and the displacement vectors (V) between the two sides are interpolated based on a predetermined conversion function from one set of displacement vectors (V) to another set of displacement vectors (V), and wherein before interpolating the displacement vectors (V) of the two side regions of the displacement vector field (VF) of each sub-image (S1, S2, S3), a convolution with a Gaussian window in the overlap region (U) is performed; - generating an output image data set (AD) based on the CT recording data (RD) and the displacement vector field (VF); - outputting the output image data set (AD).

2. Method according to claim 1, wherein a boundary layer (G) of two dimensions in the overlap region (U) is determined to determine displacement vectors (V) for registering the overlap region (U) between a first sub-image (S1, S2, S3) and a second sub-image (S1, S2, S3).

3. Method according to claim 1 or 2, wherein a regular 2D grid of sampling points (T) is defined in the overlap region (U) of a sub-image (S1, S2, S3) and for each sampling point (T) a displacement vector (V) is assigned, which points from the sampling point (T) to a position in the sub-image (S1, S2, S3).

4. Method according to claim 1 or 2, wherein in the overlap region (U) between a first sub-image (S1, S2, S3) and a second sub-image (S1, S2, S3) displacement vectors (V) are chosen such that in the overlap region (U) a sampling point (T) of the first sub-image (S1, S2, S3) is assigned a first displacement vector (V) and a corresponding sampling point (T) of the second sub-image (S1, S2, S3) is assigned a second displacement vector (V). ​ ​ ​ ​ ​ ​ 5. The method according to claim 1 or 2, wherein in forming the displacement vector field (VF) displacement vectors (V) of image regions in sub-images (SI, S2, S3) whose image values are below a predetermined boundary value are set to a predetermined value.

6. The method according to claim 1 or 2, wherein in interpolating displacement vectors (V) of the displacement vector field (VF) of image regions between the side regions, - the distance of the relevant image region to the side region is determined, - a ratio of the distance to the side region is determined, - the displacement vector (V) is calculated from the vector addition of a first displacement vector (V) in one side region of the sub-image (SI, S2, S3) and a second displacement vector (V) in the other side region and the ratio.

7. The method according to claim 1 or 2, wherein displacement vectors (V) in the displacement vector field (VF) in the overlap region and / or between the side regions are smoothed.

8. The method according to claim 1 or 2, - wherein within the scope of generating the output image data set (AD) sub- images (SI, S2, S3) are first reconstructed and then image regions of the sub- images (SI, S2, S3) are moved according to their assigned displacement vector field (VF) and then the sub-images (SI, S2, S3) are combined, or - wherein the displacement vector field (VF) is taken into account in the reconstruction process, wherein in the case of a region of the output image data set (AD) which can be formed from information of more than one sub-image (SI, S2, S3), the respective contribution of the relevant sub-images (SI, S2, S3) is weighted.

9. The method according to claim 1 or 2, wherein the output image data set (AD) and / or sub-images (SI, S2, S3) are reconstructed by means of different weighting of the temporal fractions of the CT recording data, wherein a plurality of displacement vectors (V) are assigned a weight.

10. The method according to claim 1 or 2, wherein in the overlap region of a first sub- image (SI, S2, S3) and a second sub-image (SI, S2, S3) in the case of an image point in the first sub-image (SI, S2, S3) which cannot be reconstructed or can only be reconstructed with limitations due to missing data, the missing data is supplemented on the basis of the data and displacement vector field of the second sub-image (SI, S2, S3) in order to reconstruct the image point or the relevant image point of the second sub-image (SI, S2, S3) is used to reconstruct the output image data set.

11. The method according to claim 2, wherein the boundary layer (G) is planar.

12. The method according to claim 2, wherein the boundary layer (G) is arranged perpendicular to the CT axis.

13. The method according to claim 2, wherein the boundary layer (G) is arranged perpendicular to the CT axis in a central position within the boundary of the overlap region (U).

14. The method according to claim 2, wherein a set of displacement vectors (V) is first assumed to have predetermined values to determine the boundary layer (G), and the position of the boundary layer (G) is determined along the CT axis such that the squared difference of image values is as small as possible from the start.

15. The method according to claim 2, wherein a regular 2D grid of sampling points (T) is defined in the overlap region (U) of sub-images (S1, S2, S3), and each sampling point (T) is assigned a displacement vector (V) pointing from this sampling point (T) to a position in the sub-image (S1, S2, S3), wherein the sampling points (T) are arranged equidistantly and / or form the boundary layer (G).

16. The method according to claim 3, wherein the side regions of the sub-images (S1, S2, S3) are defined by the sampling points (T), and a set of displacement vectors (V) is assigned to the side regions, which set of displacement vectors comprises the displacement vectors (V) assigned to the sampling points (T).

17. The method according to claim 4, wherein the second displacement vector (V) represents the inverse of the first displacement vector (V).

18. The method according to claim 4, wherein a predetermined measure for image region similarity of the two sub-images (S1, S2, S3) is used for determining two displacement vectors (V), wherein a minimization vector is determined step by step and added to the displacement vectors by means of an iterative method based on the current image values and image value gradients of the two sub-images (S1, S2, S3) and a predetermined step length at the respective shifted sampling points (T), wherein the minimization vector is based on the measure.

19. The method according to claim 18, wherein the measure is based on the squared difference of image values.

20. The method according to claim 5, wherein the boundary value is above the air CT value.

21. The method according to claim 5, wherein the boundary value is higher than -1000 HU and lower than -800 HU.

22. The method of claim 6, wherein said determining a ratio of distances to said side regions comprises: determining a weighted ratio of the distances to the side regions, wherein nearer side regions obtain higher weights.

23. The method according to claim 6, wherein a minimization vector is determined step by step and added to the displacement vectors based on the current image values and image value gradients of the relevant sub-image and a predetermined step length at the respective shifted sampling points, wherein the minimization vector is based on a predetermined measure.

24. The method according to claim 7, wherein a smoothing convolution or low-pass filtering is performed in the iteration step for deriving displacement vectors (V) in the overlap region (U).

25. The method according to claim 24, wherein the overlap region is a boundary layer (G).

26. The method according to claim 24, wherein the convolution is performed with a Gaussian window.

27. The method according to claim 8, wherein the output image data set (AD) has a higher image resolution than the working image data set (A).

28. The method of claim 9, wherein weight values are determined for a plurality of image points of the output image data set (AD) or the sub-images (SI, S2, S3).

29. An apparatus (18) for reconstructing a CT image, comprising: - a data interface (6) designed to receive CT recording data (RD) comprising recordings of a plurality of overlapping sub-image volumes; - a working reconstruction unit (7) designed to reconstruct a working image data set (A) from the CT recording data (RD), wherein the working image data set (A) comprises a plurality of sub-images (SI, S2, S3), wherein each sub-image (SI, S2, S3) has an overlap region (U) with at least one further sub-image (SI, S2, S3); - a displacement unit (8) designed to determine displacement vectors (V) for registering the overlap regions of the sub-images (SI, S2, S3) with each other, wherein for each sub-image (SI, S2, S3) of the working image data set (A) of each two opposing side regions a set of displacement vectors (V) is assigned, wherein in case a side region is not registered with another side region a set of predetermined displacement vectors (V) is assigned to the side region; - a vector field unit (9) designed to interpolate for each sub-image (SI, S2, S3) a displacement vector field (VF) from the sets of displacement vectors (V) of two of the side regions, wherein the two sides of the displacement vector field (VF) correspond to the determined sets of displacement vectors (V) of the sub-image (SI, S2, S3) and the displacement vectors (V) between the two sides are interpolated based on a predetermined conversion function from one set of displacement vectors (V) to another set of displacement vectors (V), and wherein before interpolating the displacement vectors (V) of the two side regions of the displacement vector field (VF) of each sub-image (SI, S2, S3) a convolution with a Gaussian window is performed in the overlap region (U); - a second reconstruction unit (10) designed to generate an output image data set (AD) based on the CT recording data (RD) and the displacement vector field (VF); - a data interface (6) for outputting the output image data set (AD).

30. A control device (11) for controlling a computed tomography system (1), comprising an apparatus (18) according to claim 29 and / or being designed to perform a method according to any one of claims 1 to 28.

31. A computed tomography system (1), comprising a control device (11) according to claim 30.

32. A computer program product comprising instructions which, when the program is run by a computer, cause the computer to carry out the steps of a method according to any one of claims 1 to 28.

33. A computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of a method according to any one of claims 1 to 28.

Images