Generation of augmented CT projection data from a subject of examination

Description

[0001] The invention relates to a method for generating augmented CT projection data of an examination object. Furthermore, the invention relates to a method for generating CT image data. The invention also relates to a projection data augmentation device. Finally, the invention relates to a computed tomography system.

[0002] In computed tomography, three-dimensional images of an object are obtained by rotating an X-ray source around the object and acquiring numerous projections at different angular positions of the source using circularly arranged detector arrays. These projections are then arranged in a sinogram. The properties of the Radon transform and the Fourier slice theorem allow the spatial distribution of the X-ray attenuation values ​​to be determined from these projections. This is usually achieved using filtered backprojection, and in recent years increasingly with iterative methods.

[0003] A general disadvantage of CT imaging is the high radiation exposure resulting from the numerous scans. One way to reduce radiation exposure is to decrease the number of acquired projections (views). However, too few individual scans or projections lead to streaking artifacts in the filtered back projection. This is FIG 1 The image is shown on the left. The sinogram (top) and the corresponding image reconstruction (bottom) with 360 views or projections are shown. A sinogram recorded with a lower angular resolution (second from the left) leads to strong streaking artifacts in the image reconstruction.

[0004] Improved reconstruction can be achieved by interpolating missing lines of a sinogram. Various interpolation methods, some more effective than others, are compared in Kim et al., 2018, "Image enhancement for computed tomography using directional interpolation for sparsely-sampled sinogram", Optik 166(2018) 227-235.

[0005] Recently, reconstructions based on neural networks have been proposed. These novel reconstruction methods interpolate additional projections using encoder-decoder architectures, particularly U-Net architectures. Such methods are described in Lee et al., "Deep-Neural-Network-Based Sinogram Synthesis for Sparse-View CT Image Reconstruction", IEEE Transactions on Radiation and Plasma Medical Sciences, Vol. 4, No. 2, March 2019, and in Chao et al., "Sparse-View Cone Beam CT Reconstruction using dual CNNs in Projection Domain and Image Domain", Neurocomputing 493 (2022) 536–547.

[0006] Als Stand der Technik sind ferner Adishesha Amogh Subbakrishna et al.: "Sinogram Domain Angular Upsampling of Sparse-View Micro-CT with Dense Residual Hierarchical Transformer and Noise-Aware Loss", bioRxiv, 12. Mai 2023, und Christiansen C, Zeng GL: "Sinogram Interpolation Inspired by Single-Image Super Resolution", Journal of biotechnology and its applications 2023, Bd. 2, Nr. 1, 15. Juni 2023, zu nennen.

[0007] The task is therefore to develop a method in connection with CT imaging that achieves improved image quality in CT imaging with a constant number of projections or with a constant resolution of the X-ray detector used to record the projections, or with constant image quality a reduction in the number of projections and thus a reduction in the radiation exposure of the imaged object compared to conventional methods, or with constant image quality a reduction in the data stream or data transfer from the X-ray detector used to record the projections.

[0008] This task is accomplished by a method for generating augmented CT projection data of an object under investigation according to claim 1. 1, a method for generating CT image data according to claim 1 8,a projection data augmentation device according to claim 9 and by a computed tomography system according to claim 10.

[0009] In the inventive method for generating augmented CT projection data from a test object, undersampled CT projection data are received from the test object (also referred to as step i)). Undersampling is understood to mean a measurement with reduced measurement information. Such measurement information includes, in particular, a reduced number of projections and / or a reduced resolution with which an X-ray detector acquires CT projection data. "Reduced" here is understood to mean relative to a predetermined reference level. Such a reference level includes, in particular, the maximum technically possible level of detail in the measurement information.This maximum size is preferably determined by the maximum number of projections per 360° rotation of the X-ray source and the X-ray detector opposite the X-ray source in a CT system, and / or by the number of detector pixels of the X-ray detector in the line and channel directions. Primarily, reducing the number of projections initially reduces the resolution of the CT image data acquired based on the projection data. The transition between reduced resolution and the appearance of artifacts is then gradual. If the number of acquired projections per 360° falls below a certain point, line artifacts appear.

[0010] The CT projection data comprises projection measurement data acquired from a test object through a measurement process. The CT projection data can be acquired directly from the test object using a computed tomography system and subsequently augmented using the method according to the invention. Alternatively, the CT projection data can be obtained from a database and / or via a data network.

[0011] Furthermore, synthetic CT projection data are generated by applying an AI-based model based on a transformer architecture (AI being an acronym for "Artificial Intelligence" and encompassing methods that employ algorithms capable of achieving results similar to those of human intelligence; this includes concepts such as machine learning, artificial neural networks, and deep learning) to the undersampled CT projection data (also referred to as step ii)). The use of a transformer architecture for generating synthetic CT projection data has the advantage over conventional models, which only utilize immediately adjacent projections or projection datasets, that it also considers CT projection datasets that are not directly adjacent.Preferably, CT projection datasets acquired rotated by 180° are also taken into account when generating synthetic CT projection data or CT projection datasets. In general terms, features from the entire time series of a CT projection dataset can be used to determine the synthetic CT projection data. This advantageously broadens the information base for determining the synthetic projection data.

[0012] Details of the transformer architecture are described in Vaswani A. et al., "Attention Is All You Need", 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA, which is incorporated into the present patent application by reference.

[0013] In this process, additional CT projection datasets lying between the adjacent CT projection datasets of the undersampled CT projection data are obtained by interpolation between the adjacent CT projection datasets (which is also part of step ii)). CT projection datasets comprise CT projection data that each correspond exactly to a single projection or were acquired with a measurement at a single angle of the X-ray detector.

[0014] Finally, augmented CT projection data are obtained by nested combination of the undersampled CT projection data and the synthetic CT projection data (also referred to as step iii)). Nested combination is understood to mean an alternating arrangement of the individual projection datasets of the synthetic CT projection data and the undersampled projection data. In a particularly preferred embodiment, individual "lines" of the sinograms of the synthetic CT projection data and the undersampled projection data, or the CT projection data associated with these lines, are to alternate during their nested combination.

[0015] As already indicated above, a lower detector resolution can also be compensated for, either additionally or alternatively, by inserting additional projection data sets (or subsets thereof) based on interpolation between projection data sets (which are also included under the term CT projection data sets used above) that are arranged adjacent to each other in the scan direction or line direction (also called z-direction). It should be noted here that an X-ray detector is two-dimensional and has both adjacent detector rows in the scan direction and channel rows perpendicular to them in the channel direction. A projection data set (or subset thereof) preferably comprises CT projection data that is assigned to such a channel row or detector row.

[0016] An increase in the resolution of CT projection data can now be achieved not only by inserting an entire CT projection dataset, which is assigned to a new angle, into the CT projection data, but the resolution within the CT projection datasets can also be increased by creating and inserting projection data between the projection data of neighboring detector rows and / or channel rows, figuratively speaking, through interpolation.

[0017] Therefore, a lower detector resolution can be achieved by a corresponding augmentation of projection data (sub)sets, which are assigned to detector rows or to adjacent channel rows running perpendicular to the detector rows and in the direction of the fan.

[0018] Advantageously, while maintaining image quality, the number of projections acquired from a single object can be reduced, thereby proportionally reducing the patient's radiation exposure. Alternatively, image quality can be improved while maintaining the same number of projections and thus the same radiation exposure. A further advantage of the method according to the invention is that the multiplication or augmentation takes place in the projection data space as a preliminary step before the actual image reconstruction, meaning the image reconstruction step does not need to be modified. As already mentioned, the resolution within individual projections can also be increased using the method according to the invention.

[0019] In the inventive method for generating CT image data, augmented CT projection data are acquired from an object of investigation based on the inventive method for generating augmented projection data. Furthermore, CT image data are reconstructed and acquired based on the augmented CT projection data. The inventive method for generating CT projection data from an object of investigation shares the advantages of the inventive method for generating augmented projection data from an object of investigation.

[0020] The projection data augmentation device according to the invention comprises an input interface which is configured to receive undersampled CT projection data from an object under investigation, which includes neighboring CT projection data sets.

[0021] The projection data augmentation device according to the invention also includes an interpolation unit, which is configured to generate synthetic CT projection data by applying an AI-based model based on a transformer architecture to the undersampled CT projection data. In this process, additional CT projection data sets lying between the adjacent CT projection data sets of the undersampled CT projection data are obtained by interpolation between these adjacent CT projection data sets.

[0022] The projection data augmentation device according to the invention also includes an augmentation unit for nested combination of the undersampled CT projection data and the synthetic CT projection data, whereby augmented CT projection data are obtained. The projection data augmentation device according to the invention shares the advantages of the inventive method for generating augmented projection data of a test object.

[0023] This disclosure further relates to a method for training an AI-based model based on a transformer architecture to generate augmented CT projection data of a subject. Training data is generated which, as input data, consists of undersampled CT projection data of a subject and, as reference output data, consists of CT projection data with a predetermined resolution, preferably that of the augmented CT projection data to be generated. The predetermined resolution is higher than the resolution of the undersampled projection data and is equal to the resolution of the augmented CT projection data. As already explained, the resolution relates to both a resolution in the projection direction (different angles of projection) and in the detector direction (which typically comprises a two-dimensional line direction and a channel direction).

[0024] The undersampled projection data are preferably obtained by "cutting out" or extracting them from fully sampled CT projection data or reference result data, which then serve as the target dataset or reference result data. Alternatively, true training data pairs must be generated by acquiring data with a scan unit of a CT system with a correspondingly reduced number of projections or a correspondingly reduced resolution within the individual CT projections or CT projection datasets.

[0025] Subsequently, the AI-based model based on a transformer architecture is trained by applying the AI-based model based on a transformer architecture to the input data of the training data and by processing the resulting data together with the reference result data, for example by comparing the generated augmented CT projection data with the reference result data.

[0026] Finally, the trained AI-based model, based on a transformer architecture, is used to generate augmented CT projection data of a study object. CT projection data from any CT scan with high angular and / or high detector resolution can be used as training data. For the input data, a portion of the projection data is omitted, while the complete projection dataset is used for the reference output data. In addition to CT projection data obtained through measurements, simulated high-resolution CT projection data (high resolution can be achieved in the angular direction as well as in the detector direction, particularly in the line direction and channel direction), generated, for example, based on mathematical phantoms, can also be used as training data.This has the advantage that, generally, an unlimited amount of training data is available. New training data can also be generated during the training of an AI-based model. New training data can be generated after each training epoch to improve the quality of the training.

[0027] The computed tomography system according to the invention comprises a scan unit for acquiring CT projection data from an object under investigation and a control unit for controlling the scan unit and for generating image data based on the acquired CT projection data. The computed tomography system according to the invention also includes a projection data augmentation device, which is configured to generate augmented CT projection data based on the acquired CT projection data. The computed tomography system according to the invention shares the advantages of the inventive method for generating augmented CT projection data from an object under investigation.

[0028] The input data consists preferably of high-resolution CT projection data, artificially undersampled by extracting every second or nth row, or every second or nth projection dataset or projection (sub)dataset, and assigning it to the input data. The reference output data can then be either the complete dataset or the CT projection data not assigned to the input data, depending on whether the reference output data is to be compared with the synthetic CT projection data or the augmented CT projection data. When using a neural network, a "comparison" involves applying an error function to the actual output (i.e., the determined result data) and to the target output (i.e., the reference output data). Based on the error determined by the error function, the weights of the neural network are then backpropagated.The training process, also referred to as "learning," ultimately involves minimizing the error function by adjusting the weights of the neural network. Thus, an AI-based model designed for superresolution application can also be trained. For generating the training data, the originally high-resolution sinogram is preferably used as the reference projection dataset or reference output data, and the input dataset is preferably reduced in size along the detector spatial direction or detector direction. Undersampled projection data in the detector spatial direction (i.e., in the row direction and / or channel direction) can also be used as input data, with projection data with increased resolution in the detector direction being used as the reference output data. The concept of superresolution is described in Zhang et al., "Deep Learning for Single Image Super-Resolution: A Brief Review," arXiv:1808.03344v3 [cs.CV] 12 Jul 2019. The content of this document is hereby incorporated into the present patent application by reference thereto.

[0029] The computer program product according to the invention comprises program code sections with which all steps of the inventive method for generating augmented CT projection data of an examination object or of the inventive method for generating CT image data are executed when the program is executed in a control unit of a computed tomography system.

[0030] A largely software-based implementation has the advantage that existing computed tomography systems or their control units can be easily retrofitted by means of a software update to operate in the manner according to the invention.

[0031] A large part of the aforementioned components of the projection data augmentation device according to the invention can be implemented wholly or partially in the form of software modules in a processor of a corresponding computer system. z. B. from a control unit of a computed tomography system or a computer used to control such a system. A largely software-based implementation has the advantage that even previously used computer systems can be easily retrofitted by a software update to operate in the manner of the invention. In this respect, the problem is also solved by a corresponding computer program product with a computer program that can be directly loaded into a computer system, containing program sections to execute the steps of the inventive method for generating augmented CT projection data of an examination object or the steps of the inventive method for generating CT image data when the program is executed in the computer system. In addition to the computer program, such a computer program product may optionally include additional components such as z. B. documentation and / or additional components, including hardware components, such as z.B. Hardware keys (dongles, etc.) for using the software are included.

[0032] For transport to the computer system or control unit and / or for storage on or in the computer system or control unit, a computer-readable medium may be used. z. B. A memory stick, a hard drive, or other portable or permanently installed data storage medium serves as the storage medium for the program sections of the computer program that can be read and executed by a computing system. The computing system can z. B. This includes one or more cooperating microprocessors or the like.

[0033] The dependent claims and the subsequent description each contain particularly advantageous embodiments and further developments of the invention. In particular, the claims of one claim category may also be further developed analogously to the dependent claims of another claim category. Furthermore, within the scope of the invention, the various features of different embodiments and claims may also be combined to form new embodiments.

[0034] As briefly explained above, a CT projection dataset can generally comprise one of the following dataset types: Projection data comprising measurement data of a detector row of an X-ray detector of a CT system with which the projection data were acquired; projection data comprising measurement data of a detector channel row, wherein the detector channel row runs perpendicular to the detector rows of the X-ray detector; projection data comprising a line of a sinogram to which a projection angle is assigned.

[0035] Such a CT projection dataset can therefore correspond both to the two-dimensional resolution of the X-ray detector and to the number of lines in the sinogram of the CT projection data. All three directions (angular direction, line direction, channel direction) influence the spatial resolution of the CT image data reconstructed from the CT projection data. It should be mentioned here that the concept of the model based on a transformer architecture is applicable both for augmenting the CT projection data in the angular direction and for implementing superresolution, i.e., increasing the (preferably two-dimensional) resolution of individual projections.

[0036] According to the invention, a CT projection dataset comprises projection data assigned to a line in a sinogram to which a projection angle is assigned. Furthermore, in the step of nested combination of the undersampled CT projection data and the synthetic CT projection data, additional CT projection datasets of the synthetic CT projection data assigned to intermediate angles between two adjacent angles of the undersampled CT projection data are inserted between the two adjacent angles of the undersampled CT projection data, so that an augmentation in the angular direction takes place.In simpler terms, gaps in the undersampled projection data are filled in the angular direction to increase the resolution of the CT image data resulting from the CT projection data and to reduce artifacts in the CT image data to be reconstructed later based on the projection data, thus improving the image quality.

[0037] In one embodiment of the inventive method for generating augmented CT projection data from an object under investigation, steps ii) and iii) are d.h. The steps for AI-based generation of synthetic CT projection data and for the nested combination of the undersampled CT projection data and the synthetic CT projection data are repeated multiple times, wherein in step ii) the most recently acquired augmented CT projection data are used for interpolation instead of the undersampled CT projection data, and in step iii) the most recently acquired augmented CT projection data are also nested with the synthetic CT projection data instead of the undersampled CT projection data. In this sequential execution of the method according to the invention, the resolution is doubled with each loop iteration. Advantageously, the described beneficial effect of the method according to the invention for generating augmented projection data of a test object is particularly strongly enhanced.

[0038] In a particularly preferred embodiment of the inventive method for generating augmented projection data of a test object, steps ii) and iii) are repeated twice. In this embodiment, the resolution M in the detector direction is increased fourfold.

[0039] In an alternative embodiment of the method according to the invention, step ii) is performed in parallel based on the undersampled CT projection data, and in step iii) the synthetic CT projection data acquired in parallel are nested and combined with the undersampled CT projection data. In this embodiment, two or more different synthetic CT projection datasets are acquired for each adjacent pair of angles based on the undersampled projection data. Together with one of the undersampled CT projection datasets of adjacent angles, the synthetic CT projection datasets thus acquired are nested and combined to form an augmented CT projection dataset, so that augmented CT projection data are generated whose resolution depends on the number of synthetic CT projection datasets acquired in parallel.

[0040] In the alternative variant described above, step ii) is preferably performed twice in parallel. That is, two synthetic CT projection datasets are generated, each assigned a different projection angle.

[0041] Alternatively, step ii) can also be performed more than twice in parallel. This advantageously broadens the data basis for later reconstruction, thus further improving the image quality of the subsequently reconstructed image data.

[0042] Particularly preferred is the combination of angular augmentation with detector-direction augmentation, where the detector-direction augmentation is achieved by applying an AI-based model using a super-resolution approach. Advantageously, the resolution of the image data in the detector direction can be further improved. This approach is especially effective when the data transmission of the projection data from the X-ray detector to the control unit is severely limited, because the reduced data rate can then be compensated for by subsequent augmentation of the projection data on the part of the stationary control unit.

[0043] Preferably, an AI-based model based on a superresolution approach is applied to the undersampled CT projection data before step ii), thereby obtaining higher-resolution undersampled CT projection data. In steps ii) and iii), the higher-resolution undersampled CT projection data are then used instead of the lower-resolution data.

[0044] Alternatively, undersampling is first performed in the detector direction, and steps ii) and iii) are then carried out based on the CT projection data undersampled both angularly (also referred to as the projection direction) and in the detector direction. Thus, unlike the variant described above, the CT projection data are first augmented in the angular direction. Subsequently, the undersampling in the detector direction is compensated by applying the AI-based model using a superresolution approach. It should be mentioned here that the superresolution concept can also be implemented, and preferably is implemented, using a model based on a transformer network.

[0045] The described combination of angular augmentation of CT projection data based on a transformer approach and superresolution augmentation in the detector space direction can be useful in systems with a reduced number of detector elements or in scan modes where subsampling of the measured data already takes place at the detector level and only a reduced number of data samples per projection are transmitted.

[0046] One reason for such undersampling can be, for example, a bandwidth limitation of the data transmission link between the X-ray detector and the control unit of a CT system. Especially with novel photon-counting X-ray detectors, the required bandwidth between the X-ray detector and the image reconstruction computer (which is part of the control unit) is extremely high. Particularly when using scan modes with the highest spatial resolution (smallest detector pixels), it is currently not possible, due to the bandwidth limitation of the data transmission link, to transmit all measured data to the image reconstruction computer in full resolution.

[0047] To circumvent this limitation, only the measurement or projection data from one of the two energy thresholds of the X-ray detector are currently transmitted at full resolution. The second threshold (or potentially further spectral thresholds, depending on the technical specifications of the X-ray detector) is undersampled and then transmitted. This means that not all spectral information is available at the image reconstruction computer at full resolution, and spectral images can only be partially reconstructed at a lower resolution. This presents another application variant of the described use of a transformer-based and AI-based model, or a model based on the principle of superresolution and AI-based. In addition to the undersampled data of threshold 2, the high-resolution data of threshold 1 can also serve as input data for the respective model.The task of the transformer-based model and / or the superresolution-based model is then to generate the full-resolution data for threshold 2 from the two input datasets. As an alternative to the previously used subsampling of the detector images, subsampling in the projection direction or angular direction could also be considered to reduce the data rate. The transformer-based models described above and the models based on the principle of superresolution can also be used in this case to calculate the missing projections and thereby increase the resolution achievable in the reconstructed images or reduce artifacts.

[0048] The projection direction refers to the angular direction or the direction of gantry rotation. Typically, a fixed number of projections are acquired per 360° during the scan, which then need to be transferred from the X-ray detector to the reconstruction computer. Reducing the number of projections to be transferred per 360° correspondingly reduces the required bandwidth. CT scanners typically use a two-dimensional X-ray detector with a specific number of rows in the z-direction (or row direction) and a specific number of channels in the orthogonal direction. This latter direction is often also referred to as the fan direction.

[0049] The invention is explained in more detail below with reference to the accompanying figures and exemplary embodiments. The figures show: FIG 1 a schematic representation of sinograms with different sampling frequencies and associated reconstructed image data, FIG 2 a schematic representation of the basic structure of a transformer architecture, FIG 3 a diagram illustrating a method for generating augmented CT projection data of an object under investigation according to a first embodiment of the invention, FIG 4 a diagram illustrating a method for generating augmented CT projection data of an object under investigation with sequential application of a transformer network according to a second embodiment of the invention, FIG 5 a diagram illustrating a method for generating augmented CT projection data from an object under investigation with parallel application of a transformer network according to a third embodiment of the invention, FIG 6 a diagram illustrating a method for generating augmented CT projection data of a subject using the sequential application of a transformer network and a superresolution network (which may also be based on a transformer architecture) ba illustrated) FIG 7 a flowchart illustrating a method for generating CT image data according to an embodiment of the invention, FIG 8 a block diagram illustrating a projection data augmentation device according to an embodiment of the invention, FIG 9 a schematic representation of a computed tomography system according to an embodiment of the invention.

[0050] In FIG 1 Figure 10 shows a schematic representation of four sinograms with different sampling frequencies and associated reconstructed image data.

[0051] The first original sinogram OS, shown on the left in Figure 10, has 360 projections P. The reconstructed image data BD shown below it are intended to serve as a reference image for the other image representations.

[0052] A second, heavily undersampled sinogram SS is in the second position from the left in FIG 1 The image is displayed and has 90 projections P. So-called streaking artifacts, resulting from the pronounced undersampling, can be seen in the associated reconstructed image data arranged below.

[0053] A third interpolated sinogram IS1 with 180 projections P, which is based on the 90 projections P of the second, heavily undersampled sinogram SS and was obtained by interpolating the 90 projections P, is shown third from the left in FIG 1 The streaking artifacts have disappeared in the associated image data BD, which are shown below the sinogram IS1, compared to the image shown second from the left.

[0054] A fourth interpolated sinogram IS2 with 360 projections P, which is also based on the 90 projections P of the second strongly undersampled sinogram SS and was obtained by a twice-executed sequential interpolation of the 90 projections P of the second strongly undersampled sinogram SS, is shown on the far right in FIG 1 shown. Here too, the streaking artifacts have disappeared in the associated image data BD, which are shown under the sinogram IS2.

[0055] In FIG 2 Figure 20 illustrates a schematic representation of the basic structure of a Transformer Architecture 20. The Transformer Architecture 20 includes a so-called encoder EC, which is located in FIG 2 shown on the left side, and a DC decoder, which is in FIG 2 The encoder EC is shown on the right. It comprises K layers, each containing two sublayers. The first sublayer includes a multi-head self-attention mechanism (MHA), a residual connection, and a layer normalization (RCN). An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is calculated as a weighted sum of the values, with the weight assigned to each value determined by a compatibility function between the query and its associated key.

[0056] A "query" is a representation of the current input vector on which the model is focusing and is used to calculate attention scores. The "key" corresponds to all input vectors in the input sequence that the current input vector could pay attention to, and these are used to calculate compatibility scores with the query. The "value" represents the actual content of the input vectors in the input sequence, which are weighted according to the attention scores to produce the final output. Essentially, the model uses queries and keys to determine which values ​​(input vectors) should be paid attention to in a given context.

[0057] Multi-head attention allows the model to access information from different representational subspaces at different positions simultaneously. With a single attentional head, this is prevented by averaging.

[0058] In mathematical terms, this results in: MultiHead Q , K , V = Concat head 1 , . . . , head h W O , where head i = Attention (QW i Q< , KW i K< , VW i V< ), where the projections parameter matrices W i Q< ∈ R d< model ×d< k W i K< ∈ R d< model ×d< k , W i V< ∈ R d< model ×d< v and WO< ∈ R d< v × d< model with the dimensions d model and dk, and MultiHead represents a "multi-head" and "Attention" represents an attention function.

[0059] Due to the reduced dimensions of each head, the overall computational effort is reduced, similar to a single-head attention with full dimensionality.

[0060] The second sublayer comprises a so-called feed-forward algorithm FFN and a residual connection and layer normalization RCN.

[0061] In addition to the aforementioned sublayers found in the EC encoder, the DC decoder (on the right in the picture) also features a sublayer with a masked multi-head self-attention mechanism MMHA and a residual compound and layer normalization RCN.

[0062] A so-called position encoding PE is used at the input of both the encoder EC and the decoder DC.

[0063] Since the transformer model contains no repetition and no convolution, it can utilize positional coding. Some information about the relative or absolute position within the sequence must be inserted into the sequence order.

[0064] For this purpose, "position codes" PE are added to the input data on the bottom sides of the encoder and decoder stacks. The position codes have the same dimension d model as the input data ED, so the two can be summed. There are many possibilities for position codes, for example, "learned" and "fixed".

[0065] Input data (ED) is entered into the encoder (EC), and output data (AD) is entered into the decoder (DC). Output data (AD) is also output from the decoder (DC).

[0066] Further details on the Transformer architecture are described in Vaswani A. et al., "Attention Is All You Need", 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.

[0067] In FIG 3 Figure 30 illustrates a method for generating augmented CT projection data of an object under investigation according to a first embodiment of the invention. (Left in) FIG 3 A sinogram of undersampled projection data PD-U is shown. The number M, A natural number indicates the number of individual detector pixels used in the detector direction for generating a projection. The number 0.5N indicates the number of projections per 360° angle, where N and 0.5N are also natural numbers. Thus, in the projection direction or angular direction, the following occurs in the FIG 3 The example shown is a 50 percent undersampling. FIG 3 Furthermore, step TR symbolizes a step in which synthetic CT projection data S-PD is determined by applying an AI-based model based on a transformer architecture to the undersampled projection data PD-U. Additional CT projection datasets are determined between angularly successive or adjacent CT projection datasets of the undersampled projection data PD-U by interpolation between the successive or adjacent CT projection datasets of the undersampled projection data PD-U. Here, a "projection dataset" is understood to be a projection line that is located in FIG 3 extends in a vertical direction. In the FIG 3 In the illustrated embodiment, the undersampled projection data PD-U and the synthetic CT projection data S-PD each comprise 0.5 N lines and thus also 0.5 N projection data sets.

[0068] Furthermore, in FIG 3 On the right side, a step of a nested combination VK of the undersampled CT projection data PD-U with the synthetic CT projection data S-PD is shown, whereby augmented CT projection data PD-A are generated with twice the resolution N of the undersampled projection data PD-U in the projection direction.

[0069] In FIG 4 Figure 40 illustrates a method for generating augmented CT projection data of an object under investigation using a sequential application of a transformer network according to a second embodiment of the invention. In the FIG 4 In the second embodiment shown, the in FIG 3 The previously illustrated procedures are executed sequentially twice in succession, with the second augmentation generating synthetic augmented CT projection data S-PD-A with twice the resolution N in the projection direction and angular direction compared to the undersampled CT projection data PD-U. Subsequently, based on the augmented CT projection data PD-A and the synthetic augmented CT projection data S-PD-A, a nested combination VK generates further augmented projection data PD-A, but now with double the resolution again, i.e., a resolution of 2N, in the angular direction.

[0070] In FIG 5 Figure 50 illustrates a method for generating augmented CT projection data of a test object using a transformer network in parallel, according to a third embodiment of the invention. In this third embodiment, instead of sequential double augmentation, a parallel double augmentation is performed. The step of AI-based generation of synthetic CT projection data S-PD is thus performed in parallel, or rather, applied in parallel to one and the same undersampled CT projection data PD-U. The undersampled CT projection data PD-U are interpolated differently to obtain two different sets of synthetic CT projection data S-PD. For example, while the first synthetic CT projection data S-PD (upper branch in Figure 50) are generated using the first set of synthetic CT projection data S-PD, the second set is generated using the second set of synthetic CT projection data S-PD. FIG 5 ) each corresponds to a first line or a first angle between two adjacent angles of the undersampled projection data PD-U, the second synthetic CT projection data S-PD (lower branch in FIG 5 ) a second angle, offset from the first angle, between the two adjacent angles of the undersampled projection data PD-U. In this way, augmented CT projection data PD-A can be generated by nested combination of the undersampled CT projection data PD-U and the first and second synthetic CT projection data S-PD, with a resolution three times that of the undersampled CT projection data PD-U. The resolution in the projection direction (angular direction) thus corresponds to 1.5 N, where N represents the maximum resolution achievable by the scanning unit without interpolation, or the maximum number of projections per 360° rotation of the X-ray detector and the X-ray source.

[0071] In FIG 6 Figure 60 illustrates a method for generating augmented CT projection data of a subject using the sequential application of a transformer network and a superresolution network. In simpler terms, the process described in FIG 6 In the illustrated embodiment, the resolution is increased not only in the projection direction but also in the detector direction, where the X-ray detector is depicted as one-dimensional for simplicity. In reality, the sensor area of ​​an X-ray detector is two-dimensional. The value M is then divided into a value M1, representing the number of pixels in the row direction, and a value M2, representing the number of pixels in the channel direction, where the channel direction runs perpendicular to the row direction.

[0072] The in FIG 6 The illustrated variant can be useful if the amount of data transferred for the projection data is limited. For example, as shown in FIG 6 As shown, only every second pixel of the M pixels is captured in the detector direction and its signal data is transmitted. This results in a kind of undersampling in both the detector direction and the projection direction, which is shown in FIG 6 This is illustrated by the values ​​"0.5 M" and "0.5 N" in the vertical direction (detector direction) and horizontal direction (projection direction) of the sinogram of the undersampled CT projection data PD-U on the left of the image. First, the already mentioned FIG 3 The illustrated method involves increasing the resolution of the undersampled projection data PD-U in the projection direction, so that the in FIG 6 The sinogram shown in the center comprises N lines or N angles in the projection direction. Subsequently, an AI-based model is applied, which is based on a network trained for so-called super-resolution SR. In this way, the resolution in the detector direction is also increased from 0.5 M to the value M, generating ultra-high-resolution CT projection data PD-UH.

[0073] In FIG 7 A flowchart is shown which illustrates a method for generating CT image data according to an embodiment of the invention.

[0074] In steps 7.I to 7.III, this is discussed in connection with FIG 3 bis FIG 6 Detailed procedures for generating augmented projection data were applied.

[0075] In detail, step 7.I involves controlling a scan unit of a CT system and recording undersampled CT projection data PD-U of an examination object O located in the scan unit.

[0076] In step 7.II, an AI-based model AI-M based on a transformer architecture is applied to the undersampled CT projection data PD-U, whereby additional CT projection data sets lying between adjacent CT projection data sets PD-S of the undersampled CT projection data PD-U are obtained as synthetic CT projection data S-PD by interpolation between the adjacent CT projection data sets PD-S.

[0077] In step 7.III, augmented CT projection data PD-A are generated by nested combination of the undersampled CT projection data PD-U obtained in step 7.I and the synthetic CT projection data S-PD generated in step 7.II.

[0078] Finally, in step 7.IV, high-resolution CT image data BD are reconstructed based on the augmented CT projection data PD-A.

[0079] In FIG 8 A block diagram is shown illustrating a projection data augmentation device 80 according to an embodiment of the invention. The projection data augmentation device 80 has an input interface 81 which is configured to receive undersampled CT projection data PD-U from a test object O.

[0080] Part of the projection data augmentation device 80 according to the invention is also an interpolation unit 82. The interpolation unit 82 is configured to generate synthetic CT projection data S-PD by applying an AI-based model AI-M based on a transformer architecture to the undersampled CT projection data PD-U. Additional CT projection data sets lying between adjacent CT projection data sets PD-S of the undersampled CT projection data PD-U are thereby obtained by interpolation between the adjacent CT projection data sets PD-S.

[0081] The projection data augmentation device 80 also includes an augmentation unit 83 for nested combination of the undersampled CT projection data PD-U and the synthetic CT projection data S-PD, yielding augmented CT projection data PD-A.

[0082] In FIG 9Figure 90 shows a schematic representation of a computed tomography system according to an embodiment of the invention.

[0083] The computed tomography system 90 comprises a control unit 91 and a scan unit 92 controlled by the control unit 91. The control unit 91 includes a control unit 93 for generating control data SD and a control interface 94 for transmitting control signals STS, which are generated based on the control data SD, to the scan unit 92. The control unit 91 also includes an input interface 95, which is configured to receive CT projection data PD from the scan unit 92. The CT projection data PD is transmitted to a projection data augmentation unit 80, which is configured to generate augmented CT projection data PD-A with increased resolution based on the CT projection data PD.

[0084] The augmented projection data PD-A are transmitted to a reconstruction unit 96, which is also part of the control unit 91 and is set up to reconstruct image data from an investigation object O located in the scan unit 92.

[0085] It should be noted that the features of all embodiments or further developments disclosed in figures can be used in any combination.

[0086] Finally, it should be noted once again that the detailed methods and setups described above are exemplary embodiments and that the basic principle can be varied in many ways by those skilled in the art without departing from the scope of the invention, insofar as it is defined by the claims. For the sake of completeness, it should also be noted that the use of the indefinite articles "a" or "an" does not preclude the possibility that the features in question may be present multiple times. Likewise, the term "unit" does not preclude the possibility that it consists of several components, which may also be spatially distributed.

Claims

1. Method for generating augmented CT projection data (PD-A) from an examination object (O), having the steps: i) receiving undersampled CT projection data (PD-U), which comprises adjacent CT projection data sets (PD-S), from the examination object (O), wherein a CT projection data set (PD-S) comprises projection data, which is assigned to a line in a sinogram assigned to a projection angle, ii) AI-based generation of synthetic CT projection data (S-PD) by way of application of an AI-based model (KI-M) based on a transformer architecture to the undersampled CT projection data (PD-U), wherein additional CT projection data sets located between the adjacent CT projection data sets (PD-S) of the undersampled CT projection data (PD-U) are obtained by way of interpolation between the adjacent CT projection data sets (PD-S), iii) interleaved combination of the undersampled CT projection data (PD-U) and the synthetic CT projection data (S-PD), wherein augmented CT projection data (PD-A) is obtained, wherein additional CT projection data sets of the synthetic CT projection data (S-PD) assigned to intermediate angles between two adjacent angles of the undersampled CT projection data (PD-U) are integrated between CT projection data sets (PD-S) of the undersampled CT projection data (PD-U) assigned to the two adjacent angles, enabling an augmentation in the angular direction to take place.

2. Method according to claim 1, wherein steps ii) and iii) are repeated multiple times, wherein, during the repetition, in step ii), the most recently obtained augmented CT projection data (PD-A) is used for the interpolation instead of the undersampled CT projection data (PD-U) and in step iii), instead of the undersampled CT projection data (PD-U), the most recently obtained augmented CT projection data (PD-A) is likewise combined in an interleaved manner with the synthetic CT projection data (S-PD).

3. Method according to claim 2, wherein steps ii) and iii) are repeated twice.

4. Method according to one of the preceding claims, wherein step ii) is performed in parallel on the basis of the undersampled CT projection data (PD-U) and in step iii) the synthetic CT projection data (S-PD) obtained in parallel is combined in an interleaved manner with the undersampled CT projection data (PD-U).

5. Method according to claim 4, wherein step ii) is performed twice in parallel.

6. Method according to claim 4, wherein step ii) is performed more than twice in parallel.

7. Method according to one of the preceding claims, wherein an augmentation in the angular direction is combined with an augmentation in the detector direction, wherein the augmentation in the detector direction takes place using an AI-based model (KI-M) based on a super resolution approach.

8. Method for generating CT image data (BD), having the steps: - obtaining augmented CT projection data (PD-A) based on the method according to one of claims 1 to 7, - reconstructing CT image data (BD) based on the augmented CT projection data (PD-A).

9. Projection data augmentation facility (80), having: - an input interface (81) for receiving undersampled CT projection data (PD-U) which comprises adjacent CT projection data sets (PD-S), from an examination object (O), wherein a CT projection data set (PD-S) comprises projection data, which is assigned to a line in a sinogram assigned to a projection angle, - an interpolation unit (82) for AI-based generation of synthetic CT projection data (S-PD) by way of application of an AI-based model (KI-M) based on a transformer architecture to the undersampled CT projection data (PD-U), wherein additional CT projection data sets located between the adjacent CT projection data sets (PD-S) of the undersampled CT projection data (PD-U) are obtained by way of interpolation between the adjacent CT projection data sets (PD-S), - an augmentation unit (83) for interleaved combination of the undersampled CT projection data (PD-U) and the synthetic CT projection data (S-PD), wherein augmented CT projection data (PD-A) is obtained, in that additional CT projection data sets of the synthetic CT projection data (S-PD) assigned to intermediate angles between two adjacent angles of the undersampled CT projection data (PD-U) are integrated between CT projection data sets (PD-S) of the undersampled CT projection data (PD-U) assigned to the two adjacent angles, enabling an augmentation in the angular direction to take place.

10. Computed tomography system (90), having: - a scanning unit (92) for acquiring CT projection data (PD) from an examination object (O), - a control facility (91) for controlling the scanning unit (92) and for generating image data (BD) based on the CT projection data (PD), - a projection data augmentation facility (80) according to claim 9.

11. Computer program product with a computer program, which can be loaded directly into a memory of a control facility (91) of a computed tomography system (90), with program code sections for carrying out all steps of a method according to one of claims 1 to 8 when the program is executed in the control facility (91).

12. Computer-readable medium, on which program sections which can be executed by a computer unit are stored for carrying out the steps of a method according to one of claims 1 to 8, when the program sections are executed by the computer unit.