Processing of projection data produced by a computed tomography scanner

Abstract

A mechanism for processing projection data generated by a computed tomography scanner, where the projection data is processed by a machine learning algorithm that is trained to perform upsampling or super-resolution techniques on the input data to generate higher resolution projection data.

Description

[Technical field]

[0001] The present disclosure relates to the field of medical imaging, and in particular to the field of processing projection data produced by a computed tomography scanner. [Background technology]

[0002] Computed tomography (CT) scanners are well-established medical imaging devices that use a detector device formed from an array of detectors or pixels to detect the interaction between X-ray radiation energy and irradiated material to generate medical imaging data. Summary of the Invention [Problem to be solved by the invention]

[0003] There is a continuing interest in increasing the resolution of images produced by CT scanners while maintaining image quality. Techniques for improving the resolution of CT scans include scanning the object with a small focal spot, small detector / pixel size, and sufficient sampling in both angular and radial directions.

[0004] In particular, radial sampling has a significant effect on the resolution of the images produced by a CT scanner. For a detector configuration with a detector / pixel size of W in the fan direction (projected to the isocenter), the resolution limit associated with averaging the intensity over the pixel size is f W = 1 / W, which is the spatial frequency at which the convolution of the signal over the pixel width has its first zero crossing. Since the detector / pixel spacing roughly matches the detector / pixel size (ignoring any spacers between pixels), the Nyquist frequency associated with radial sampling is therefore f S = 1 / 2W. In a typical CT system, f W is about 13 lp / cm, and f S is about 6.5 lp / cm (where lp refers to the line pair). Therefore, without further technical effort, fS Since efforts to reconstruct an image with frequency components beyond φ will result in aliasing artifacts, the spatial resolution is limited by the radial sampling.

[0005] One known approach to overcome this problem is to use dual focal spot (DFS) acquisition, where the focal spot of the x-ray tube is deflected along the source path. By appropriate selection of the distance, the radial sampling pattern is shifted by half the detector / pixel spacing for every other view. Parallel rebinning allows data from two focal spot positions to be interleaved (e.g., to form a single view), and the resulting Nyquist frequency of the radial sampling matches the resolution limit associated with the pixel size.

[0006] Another sampling condition that must be observed is that classical sampling theory requires that when an N×N image is to be reconstructed, Nπ / 2 angular samples are measured over 180 degrees, resulting in approximately 800 angular samples for the typically used 512×512 image size.

[0007] It is desirable to further improve image resolution, enabling the generation of high resolution images for various forms of CT imaging modalities.

[0008] EP 3 447 731 A1 discloses an approach for generating enhanced tomographic images of an object.

[0009] US Patent Application Publication No. 2013 / 051519 A1 discloses an approach for increasing the resolution of 2D projection images for cone-beam computed tomography data. [Means for solving the problem]

[0010] The invention is defined by the claims.

[0011] According to an embodiment of the present invention, a computer-implemented method for processing projection data generated by a CT scanner is provided.

[0012] The computer-implemented method includes acquiring projection data generated by a CT scanner, processing the projection data using a machine learning algorithm configured to perform a super-resolution imaging technique on the projection data to increase an apparent sampling of the projection data in at least one dimension, and outputting the processed projection data.

[0013] The machine learning algorithm is trained using a training dataset, the training dataset including an input training dataset formed from a plurality of input training data entries, each including low-resolution projection data of an imaged object, and an output training dataset formed from a plurality of output training data entries, each output training data entry corresponding to a respective input training data entry and including high-resolution projection data of the same imaged object of the respective input training data entry.

[0014] The high-resolution projection data of each output training data entry is generated by a training CT scanner that generates the high-resolution projection data, the training CT scanner using a dual focus acquisition technique to generate the intermediate data, the intermediate data including interleaved first and second sample sets, each sample set acquired using a different focus, and parallel binning is performed on the interleaved first and second sample sets to generate the high-resolution projection data.

[0015] The low-resolution projection data is generated by discarding the first or second sample set of the intermediate data, such that only the first or only the second sample set forms the low-resolution projection data.

[0016] The proposed approach thereby utilises machine learning algorithms to improve the resolution and / or apparent sampling of the projection data, and it is recognised herein that machine learning algorithms may be advantageously trained to increase the resolution of the projection data.

[0017] The use of machine learning algorithms can be used to complement and / or replace one or more resolution improvement techniques known in the art, such as DFS acquisition. This is particularly advantageous in scenarios where such previously known techniques are difficult to implement during imaging procedures. As an example, when kVp switching is performed, DFS techniques are technically difficult to implement.

[0018] In particular, machine learning algorithms are effectively trained to emulate dual focus acquisition techniques, and this means that the benefits of DFS techniques can be utilized even in situations where the true performance of such techniques is difficult, e.g., when implementing kVp switching.

[0019] This approach has advantages over other forms of generating training data sets because it does not introduce artificial noise to generate the low-resolution projection data (e.g., resulting from using blurring or downsampling techniques) and does not require data to be removed from the projection data acquired to clinically represent the scene. Furthermore, the low-resolution projection data continues to represent clinically useful and relevant medical information, improving the relevance and accuracy of the super-resolution performed by machine learning methods.

[0020] The machine learning algorithm may be configured to increase the apparent sampling of the projection data in a radial direction. In some examples, the machine learning algorithm is configured to increase the apparent sampling of the projection data in an angular direction. Of course, in still other examples, the machine learning algorithm increases the apparent sampling in both directions.

[0021] The projection data may be projection data generated by a CT scanner using a dual focus acquisition approach.

[0022] In at least one example, the machine learning algorithm is further trained using the modified training dataset, the modified training dataset including a modified input training dataset in which noise is added to the low-resolution projection data of each input training data entry of the training dataset, and an output training dataset of the training dataset, which introduces robustness against noise for the machine learning algorithm, such that the machine learning algorithm is trained to perform both the denoising and super-resolution procedures.

[0023] In some examples, the projection data includes a plurality of samples, each sample associated with a different combination of parameter r and angle Φ, where the parameter r represents a minimum distance between an isocenter of the computed tomography scanner and a ray used by the computed tomography scanner to generate the sample, and the angle Φ represents an angle between the ray used by the computed tomography scanner to generate the sample and a predetermined plane, and the machine learning algorithm is configured to generate a plurality of new samples for the projection data, each new sample generated from a different subset of the plurality of samples, where the subset includes only samples for which an absolute difference between a value of the parameter r of the sample and a value of the parameter r of the new sample is below a first predetermined threshold and / or includes only samples for which an absolute difference between a value of the angle Φ of the sample and a value of the angle Φ of the new sample is below a second predetermined threshold.

[0024] This approach effectively limits or reduces the size of the receptive field when processing the projection data. It has been recognized that the most valuable information for predicting new projection samples, especially in CT projection data, is found in nearby existing projection samples. Limiting the size of the receptive field increases the efficiency of processing the projection data using machine learning algorithms.

[0025] In some examples, the machine learning algorithm is further trained using a second training dataset comprising a second input training dataset formed from a plurality of second input training data entries, each comprising low-resolution image data of a scene, and a second output training dataset formed from a plurality of second output training data entries, corresponding to each second input training data entry and comprising high-resolution image data of the same scene of each second input training data entry.

[0026] This embodiment recognizes that machine learning methods trained to perform super-resolution imaging or up-sampling of image data (e.g., images) can be repurposed or designed to perform super-resolution on projection data. This approach can leverage existing databases of image data (high and low resolution) to provide a large databank of examples for accurate up-sampling.

[0027] The image data is preferably unprojected image data. This approach recognises that standard super-resolution machine learning methods can achieve improved resolution of projected data.

[0028] Preferably, the values ​​of the low-resolution image data and the high-resolution image data are scaled to correspond to the range of possible values ​​of the projection data produced by the CT scanner. This embodiment provides more relevant training data designed to improve the resolution of the projection data.

[0029] Also proposed is a computer-implemented method for generating a training data set for training a machine learning algorithm that performs a super-resolution imaging technique on projection data generated by a CT scanner.

[0030] The computer-implemented method includes generating, for the training dataset, an output training dataset formed from a plurality of output training data entries, each including high-resolution projection data of an imaged subject, where the training dataset receives intermediate data generated by a training CT scanner and generates the intermediate data, the intermediate data including interleaved first and second sample sets, each sample set acquired using a different focus, and performing parallel binning on the interleaved first and second sample sets to generate the high-resolution projection data.

[0031] The computer-implemented method also includes generating an input training data set formed from a plurality of input training data entries, each input training data entry corresponding to a respective output training data entry, each including low-resolution image projection data of the same image object in the respective output training data entry, and generating the input training data set by discarding the first sample set or the second sample set of the intermediate data.

[0032] In a preferred example, the step of generating the output training data set further comprises controlling the training computer tomography using a dual focus acquisition technique to generate the intermediate data.

[0033] Also proposed is a non-transitory computer program product comprising computer program code which, when executed on a computing device having a processing system, causes the processing system to perform all of the steps of any of the methods described herein.

[0034] A processing system for processing projection data generated by a CT scanner is also proposed, the processing system being configured to acquire projection data generated by the CT scanner, process the projection data using a machine learning algorithm configured to perform a super-resolution imaging technique on the projection data to increase an apparent sampling of the projection data in at least one dimension, and output the processed projection data.

[0035] Also proposed is an imaging system comprising a processing system and a CT scanner.

[0036] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

[0037] For a better understanding of the present invention and to show more clearly how it may be carried into effect, reference will now be made, by way of example only, to the accompanying drawings in which: [Brief description of the drawings]

[0038] [Figure 1] 1 shows an imaging system. [Diagram 2] 1 conceptually illustrates the radiation output from a radiation source in a CT scanner. [Diagram 3] 1 conceptually illustrates projection data generated by a CT scanner. [Figure 4] 1 conceptually illustrates projection data generated by a CT scanner using dual focus acquisition technique. [Diagram 5] The proposed approach is presented. [Figure 6] 1 shows an exemplary receptive field for the machine learning method used in the proposed approach. [Figure 7] The effectiveness of the proposed approach is shown. [Figure 8] Here's how. [Figure 9] 1 shows a processing system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0039] The present invention will now be described with reference to the drawings.

[0040] It should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the devices, systems, and methods, are for purposes of illustration only and are not intended to limit the scope of the invention. These and other features, aspects, and advantages of the devices, systems, and methods of the present invention will become better understood from the following description, appended claims, and accompanying drawings. It should be understood that the figures are merely schematic and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the figures to denote the same or similar parts.

[0041] The present disclosure provides mechanisms for processing projection data generated by a CT scanner, where the projection data is processed by machine learning algorithms trained to perform upsampling or super-resolution techniques on input data to generate higher resolution projection data.

[0042] The approach is based on the recognition that machine learning algorithms can be appropriately trained to upsample or improve the resolution of projection data. This reduces the need to modify imaging techniques or CT scanners to provide higher resolution projection data. The embodiments can be used, for example, to provide high quality projection data even in cases where other sampling / resolution improvement techniques cannot be implemented.

[0043] The embodiments may be used in any imaging system that has a CT scanner, such as those used in clinical or healthcare settings.

[0044] 1 illustrates an imaging system 100, in particular a computed tomography (CT) imaging system in which embodiments of the present invention may be used. The CT imaging system includes a CT scanner 101 and a processing interface 111 for processing and performing operations using data generated by the CT scanner.

[0045] The CT scanner 101 generally includes a stationary gantry 102 and a rotating gantry 104. The rotating gantry 104 is rotatably supported by the stationary gantry 102 and rotates about an examination region 106 about a vertical or z-axis, which may be defined as being perpendicular to the xz-axis and in a horizontal plane (e.g., perpendicular to gravity). The y-axis may also be defined as being perpendicular to the z-axis and x-axis, i.e., the vertical axis or an axis aligned with gravity.

[0046] A patient support 120, such as a couch, supports an object, such as a human patient, in the examination region 106. The support 120 is configured to move the object or subject for loading, scanning, and / or removing the object or subject.

[0047] A radiation source 108, such as an x-ray tube, is rotatably supported by the rotating gantry 104. The radiation source 108 rotates with the rotating gantry 104 and emits radiation that traverses an examination region 106.

[0048] A radiation sensitive detector array 110 subtends an angular arc across the examination region 106 opposite the radiation source 108. The detector array 110 includes one or more rows, e.g., two or more rows, of detectors extending along a z-axis direction to detect radiation traversing the examination region 106 and generate projection data indicative thereof. The projection data may be, for example, cone beam projection data. Each detector represents a pixel of the detector array.

[0049] As the rotating gantry rotates, the detector array captures a series of sample sets of the received radiation. Each sample set represents a portion of the projection data for a different view. More specifically, a single view represents a portion of the projection data obtained when the radiation source is at a single position. The portions of the projection data from the different views may be combined, e.g., concatenated together, to form the projection data provided by the detector array. In this manner, the projection data includes a plurality of samples, each sample captured by the detector at a different position relative to the examination region. The plurality of samples may be divided into a plurality of sample sets, each sample set including samples captured at a different projection view.

[0050] The angular increment of rotation of the rotating gantry between different projection views, and thus the capture of different sample sets by the detector array, is referred to as the angular sampling density.

[0051] The processing interface 111 processes the projection data generated by the CT scanner and may also facilitate user control over the operation of the CT scanner.

[0052] In particular, the processing interface serves as an operator console 112 and may comprise a general-purpose computing system or computer including input devices 114, such as a mouse, keyboard, and / or the like, and output devices 116, such as a display monitor, filmer, or the like. The console 112 allows an operator to control the operation of the system 100. Data generated by the processing interface may be displayed through at least one display monitor of the output devices 116.

[0053] The processing interface processes the projection data and utilizes a reconstructor 118 to reconstruct image data, for example, for display on an output device 116 or for storage in an external server or database. It should be appreciated that the reconstructor 118 may be implemented via a microprocessor executing computer readable instructions encoded or embedded on a computer readable storage medium, such as physical memory and other non-transitory media. Additionally or alternatively, the microprocessor may execute computer readable instructions carried by carrier waves, signals, and other transitory or non-transitory media.

[0054] The reconstructor 118 may use filtered backprojection (FBP) reconstruction, image domain and / or projection domain reduced noise reconstruction algorithms (eg, iterative reconstruction), and / or other algorithms to perform the reconstruction.

[0055] FIG. 2 shows radiation traversing an examination region for purposes of understanding the nomenclature used throughout this disclosure.

[0056] 2 shows multiple light rays emitted by radiation source 108. Each light ray represents radiation incident on a different detector 201 of detector array 110.

[0057] Each ray from the radiation source 108 to the detector 201, and the samples produced by the detector in response to that ray, is characterized by its values ​​of r and Φ.

[0058] The parameter r represents the distance of the ray to the isocenter 210 of the computed tomography system 106. The isocenter 210 represents the center of the region about which the rotating gantry of the computed tomography system rotates, i.e., the center of rotation. In particular, the parameter r represents the distance of an imaginary line connecting the isocenter 210 to the center, where the imaginary line is connected perpendicular to the center.

[0059] The angle Φ represents the angle of the ray from the x-axis. In particular, the angle Φ represents the angle between an imaginary line connecting the isocenter 210 to the center, which is connected to the x-axis and perpendicular to the center.

[0060] For convenience, the distance of the parameter r is often considered to be the negative for rays on the left half of the detector array, for example from a given viewpoint at the front of the CT scanner.

[0061] The radial sampling density is often simply called radial sampling and is defined as the difference in radial values ​​r associated with adjacent detectors in a detector array. Thus, if one of the centers of radiation passes through the isocenter 210, as shown in FIG. 2, the radial sampling density is the distance r of a second imaginary line connecting the isocenter and the closest center that does not pass through the isocenter. S or length, and a second imaginary line is perpendicular to said nearest center.

[0062] FIG. 3 conceptually illustrates samples contained in the projection data.

[0063] It will be appreciated that each sample taken by a detector in the detector array is associated with a different value of the parameter r and the angle Φ. Thus, as shown, the samples of projection data can be plotted in rΦ space, with each sample 301 being represented diagrammatically using a circle.

[0064] Samples captured simultaneously, i.e. for the same effective projection view, belong to the same sample set and lie on the same tilt line 302. When the projection view changes, i.e. after the rotating gantry rotates, different sample sets are captured. This means that the angular sampling density 303 can also be derived based on the difference in the angular value Φ between two adjacent samples having the same value for r, i.e. representing samples captured at successive time steps by the same detector of the detector array.

[0065] The present invention relates to an approach for improving the resolution of image data, i.e. images, produced by a computed tomography scanner. It will be readily apparent that the greater the resolution of the projection data, the greater the resolution of the image data reconstructed from the projection data.

[0066] In particular, the present invention can provide an approach for increasing the apparent sampling of the detector array, for example, by appending, inserting, or interpolating additional samples to the projection data.

[0067] Existing approaches to improving the resolution of image data have focused on improving the sampling density in generating the projection data.

[0068] Some approaches that may be adopted include decreasing the spacing between detectors in the detector array and / or decreasing the size of the detectors in the detector array to allow a larger number of detectors to fit in the same area. Increasing the number of detectors in a detector array of the same size has the natural consequence of increasing the sampling density by the detector array. However, smaller detector sizes are more expensive to manufacture and have higher handling costs, for example, because additional bandwidth between the gantry and the processing interface is required. Therefore, alternative approaches are preferred.

[0069] Another approach is to use dual focus (DFS) acquisition. In the DFS acquisition technique, the focal spot of the radiation source can be deflected along the source path. By appropriately selecting the deflection distance, it is possible to shift the sampling pattern by 1 / 2 pixel interval in the radial direction for every other sample acquired by the detector array. Thus, two sample sets with different focal spots are acquired for each rotation of the rotating gantry. Parallel rebinning allows the data from the two focal positions to be interleaved. This forms, for a single view, a projection data portion from two consecutive samples acquired by the detector array. This effectively increases the radial sampling density.

[0070] A more complete description of the parallel rebinning process, and the DFS acquisition approach more generally, is provided in US Pat. No. 4,637,040.

[0071] FIG. 4 conceptually illustrates raw samples contained in the projection data of a CT scanner operating using a DFS acquisition process.

[0072] As explained above, under DFS acquisition, two sample sets are acquired at successive rotation angles of the rotating gantry. The first sample set 401 is shown as a circle and the second sample set 402 is shown as a triangle. Thus, the first sample set is interleaved with the second sample set. In other words, a pair of sample sets is acquired.

[0073] The first and second sample sets are then subjected to parallel rebinning. This approach may effectively move each sample in a pair of first and second sample sets to have the same angle Φ or to lie on a single tilted line in rΦ space instead of two tilted lines. In this way, within each pair of first and second sample sets, the first and second sample sets are combined to generate a combined sampling set that represents a single projection view. This approach effectively reduces the angular sampling density of the projection data while increasing the effective radial sampling of any given sampling set. This is because the number of samples per sampling set increases while the total number of sampling sets of samples that lie along the same line decreases.

[0074] A comparison of Figures 3 and 4 demonstrates how the effective radial sampling can be increased when DFS is implemented.

[0075] Of course, any combination of these approaches described above may be implemented.

[0076] The proposed approach provides an alternative mechanism for increasing the effective sampling or resolution of the projection data that can be used instead of or in parallel with existing approaches such as those mentioned above. In particular, embodiments of the present invention can be configured to effectively provide supplemental samples to those contained in the projection data, for example to interpolate between existing samples. This provides an upsampling or super-resolution process.

[0077] The present disclosure utilizes machine learning algorithms to perform super-resolution or upsampling on projection data. For example, upsampling can be performed to increase the apparent radial and / or angular sampling density. Approaches for generating or defining training data used to configure / train such machine learning algorithms are also provided.

[0078] FIG. 5 conceptually illustrates one approach 500 for using a machine learning algorithm 550 according to one embodiment.

[0079] The machine learning algorithm 550 is configured to receive as input projection data 510 generated by a CT scanner. The machine learning algorithm provides as output upsampled projection data 520. The upsampled projection data 520 has a higher resolution than the input projection data.

[0080] The machine learning algorithm may be configured or trained to perform upsampling in a single direction on the projection data, for example, in the radial or angular direction, respectively. In other examples, the machine learning algorithm is configured or trained to perform upsampling in multiple directions, for example, in both the radial and angular directions.

[0081] More specifically, the machine learning algorithm may be configured to generate a plurality of new samples for the projection data. For example, referring to FIG. 3 or FIG. 4, the machine learning algorithm may be configured to generate a plurality of new samples that are positioned (in rΦ space) between existing or captured samples of the original projection data. Thus, the machine learning algorithm is used to improve the resolution of the projection data. This improves the resolution of image data generated from the projection data.

[0082] A machine learning algorithm is any self-training algorithm that processes input data to generate or predict output data, where the input data includes projection data and the output data includes upsampled or higher resolution projection data.

[0083] Machine learning algorithms suitable for use in the present invention will be apparent to those skilled in the art. Examples of suitable machine learning algorithms include neural networks, logistic regression, support vector machines, or naive Bayes models. Neural networks are a particularly effective approach for processing projection data.

[0084] The structure of a neural network (or simply neural network) is inspired by the human brain. A neural network is composed of layers, each layer containing multiple neurons. Each neuron contains a mathematical operation. In particular, each neuron may comprise a different weighted combination of a single type of transformation (e.g., the same type of transformation, such as sigmoid, but with different weightings). In the process of processing input data, the mathematical operation of each neuron is performed on the input data to generate a numerical output, and the output of each layer in the neural network is sequentially fed to the next layer. The final layer provides the output.

[0085] Methods for training machine learning algorithms are well known. Typically, such methods include obtaining a training data set that includes input training data entries and corresponding output training data entries. In other words, the training data set includes a number of pairs of input training data inputs and output training data inputs.

[0086] The initialized machine learning algorithm is then applied to each input data entry to generate a predicted output data entry. The error between the predicted output data entry and the corresponding training output data entry is used to modify the machine learning algorithm. This process can be repeated until the error converges and the predicted output data entry is sufficiently similar to the training output data entry (e.g., ±1%). This is commonly known as a supervised learning technique.

[0087] For example, if a machine learning algorithm is formed from a neural network, the mathematical operations (weights) of each neuron can be modified until the error converges. Known methods for modifying neural networks include gradient descent, backpropagation algorithms, etc.

[0088] Each input training data entry includes exemplary projection data or lower resolution projection data. Each output training data entry includes high resolution projection data that is an upsampled or high resolution version of the exemplary projection data or lower resolution projection data. Each input training data entry corresponds to an output training data entry and describes the same scene or region.

[0089] To define the output training data entries, different instances of projection data are acquired, with each acquired instance of projection data serving as an output training data entry, i.e., higher resolution projection data.

[0090] Each instance of the projection data was generated by a training CT scanner using the dual focus acquisition technique described above, and therefore each instance of the projection data may be acquired from a training computed tomography scanner or from a memory that stores data generated from a training computed tomography scanner.

[0091] Thus, each instance of projection data to be used as an output training data entry is generated by the training CT scanner by using a dual focus acquisition technique to generate intermediate data, the intermediate data comprising interleaved first and second sample sets, each sample set acquired using a different focus, and parallel binning is performed on the interleaved first and second sample sets to generate an instance of projection data, i.e., an instance of high resolution projection data that serves as the output training data entry. This process is repeated multiple times to generate multiple output training data entries.

[0092] The process for generating the higher resolution projection data (representing the output training data entry) is modified to generate the corresponding input training data entry. As described above, the method for generating projection data using a dual focus acquisition technique uses a parallel rebinning technique to form, for a particular view, a projection data portion from two successive sample sets taken at the same view but with different focus positions. To generate lower resolution projection data, this parallel binning technique can be omitted, and instead, only one of the two successive sample sets is used as the projection data portion for that particular view. This effectively reduces the radial sampling density for the lower resolution projection data compared to the higher resolution projection data.

[0093] As further explanation, with reference to FIG. 4 , the low resolution projection data may include only the first sample set 401 or only the second sample set 402, and the high resolution projection data may include the first sample set 401 and the second sample set 402.

[0094] This approach means that the machine learning algorithm effectively simulates or acts like a bifocal acquisition technique: by using such a machine learning algorithm to increase the resolution of the projection data (in the inference stage), it can replace the performance of the bifocal acquisition.

[0095] The use of machine learning algorithms trained using a training data set is particularly advantageous when the CT scanner generating the projection data operates in kVp switching mode, as DFS is technically difficult to implement on fast kVp switching systems, making the use of such machine learning algorithms particularly advantageous.

[0096] In some embodiments, the machine learning algorithm may be further trained using a second training data set, the second training data set including a plurality of second input training data entries and a plurality of second output training data entries.

[0097] In first and second examples of such an embodiment, each second output training data entry (as well as the output training data entries of the training data set) includes an instance of projection data generated by the training computed tomography scanner. In a third example of such an embodiment, each pair of input training data input and output training data input includes image data of a particular scene, and for any given pair, the input training data entry may include lower resolution image data of the scene and the output training data entry may include higher resolution image data of the scene.

[0098] In a first example, different instances of projection data are generated by a training CT scanner. Thus, each instance of projection data may be acquired / received from a training computed tomography scanner or from a memory that stores data generated from the training computed tomography scanner. Each acquired instance of projection data may serve as a second output training data entry. Each acquired instance of projection data may be subjected to a resolution reduction process, for example, by applying a blur kernel to each acquired instance of projection data. This generates a corresponding instance of reduced resolution projection data that acts as a corresponding second input training data entry.

[0099] This first example can effectively simulate projection data having reduced resolution through the use of a larger focal spot (to serve as the second input training data entry).

[0100] In a second example, different instances of projection data are similarly generated by the training CT scanner. Thus, each instance of projection data may be acquired / received from a training computed tomography scanner or from a memory that stores data generated from the training computed tomography scanner. Each acquired instance of projection data may serve as a second output training data entry. The process for generating the projection data (representing the second output training data entry) may be modified to generate the corresponding second input training data entry. In particular, the projection data (for the second output training data entry) may be generated by the training CT scanner by combining projection data portions acquired from multiple different views. A similar generation process may be used to generate low-resolution projection data for the corresponding second input training data entry, but the combination of projection data portions may be modified to omit, delete, or remove projection data portions associated with one or more views, e.g., selection of an alternative view. For example, only projection data portions from every other view in the sequence of views may be used to generate the projection data of the second input training data entry.

[0101] Thus, the projection data for each second output training data entry may be generated by a generation process that includes generating intermediate data for multiple views of the subject. The projection data for each second input training data entry may be generated by removing portions of the intermediate data that correspond to selections of the multiple views.

[0102] The approach taken by the second example effectively reduces the angular sampling of the projection data to produce lower resolution projection data, and a machine learning algorithm further trained on such an approach correspondingly increases the angular sampling of the projection data.

[0103] In a third example, each pair of second input training data entries and second output training data entries includes image data for a particular scene, and for any given pair, the second input training data entry may comprise lower resolution image data of the scene and the second output training data entry may comprise higher resolution image data of the scene.

[0104] The image data may be non-projected image data, ie, image data not generated by a CT scanner.

[0105] In an advantageous version of this third example, the low and high resolution image data may be grayscale and the values ​​of different pixels in the image data may be constrained or scaled to a range of values ​​that may be included in the projection data produced by the CT scanner. A softmax function is one suitable function for modifying the values ​​of the image data to fall within the range of values ​​of the projection data output by the CT scanner.

[0106] This third example recognizes that a machine learning algorithm trained to upsample or improve the resolution of image data can be directly adapted to upsample or improve the resolution of projection data. This approach facilitates the use of more accurate upsampling techniques, e.g., those trained using extremely large databases of images. Such large databases containing comparable amounts of projection data can be difficult and time consuming to obtain.

[0107] Also, the image data need not be projection data, but may be medical image data, rather it is recognized that any trained super-resolution machine learning method can be readily applied to projection data to improve the apparent sampling of the projection data without significant loss of accuracy or resolution.

[0108] In any of the above examples, noise can be added to the lower resolution projection data of the input training data entry (or the second input training data entry). This effectively trains or configures a machine learning algorithm trained using such modified training data set to also perform the denoising technique. Thus, the machine learning algorithm can be trained to perform joint resolution restoration and denoising.

[0109] The machine learning method may be configured such that (in generating the upsampled projection data) each new sample generated by the machine learning method depends on a subset (i.e., not all) of the original samples of the upsampled projection data, thus intentionally limiting the receptive field for a particular sample.

[0110] For example, in some cases, each new sample may depend only on a group of samples that are in a nearby or proximate region of rΦ-space.

[0111] In some examples, the machine learning method is designed such that for any new sample generated by the machine learning method, the receptive field includes only those samples of projection data in the closest two or closest four sample sets of projection data in rΦ-space. Thus, if a new sample is to have a value Φ1 for angle Φ, only the two or four sample sets closest to this value are used to generate the new sample.

[0112] This is conceptually illustrated in FIG. 6, which shows how the receiving field 612 can include the four closest sample sets for the first new sample 611. Other suitable sizes for the receiving field 612 are envisioned, e.g., the two closest sample sets, the six closest sample sets, etc. In such embodiments, the number of sample sets in the receiving field can be an even positive number to improve the balance of sample sets contributing to the new sample.

[0113] As another example, the machine learning method is designed such that for any new sample generated by the machine learning method, the receiving field includes only samples of projection data generated by a subset of detectors of the detector array in the two or four closest sample sets of projection data in rΦ space. Thus, if a new sample should have a value r2 for parameter r and a value Φ2 for angle Φ, only samples obtained by detectors that generated the K neighboring values ​​closest to r2 for parameter r will be used to generate the new sample for the L set closest to value Φ2. The values ​​of K and L may be equal or different and are positive integer values ​​(e.g., 2 or 4). The values ​​of K and / or L may be even positive numbers to improve the balance of the sample sets contributing to the new sample.

[0114] This is also conceptually shown in FIG. 6, which demonstrates how, for a second new sample 621, the receiving field 622 can contain only nearby samples.

[0115] Accordingly, the machine learning algorithm may be configured to generate a plurality of new samples for the projection data, each new sample generated from a different subset of the plurality of samples, the subset comprising only samples for which an absolute difference between a value of the parameter r of the sample and a value of the parameter r of the new sample is below a first predetermined threshold and / or an absolute difference between a value of the angle Φ of the sample and a value of the angle Φ of the new sample is below a second predetermined threshold.

[0116] The first predetermined threshold effectively limits the number of detectors that contribute to the generation of a new sample. In particular, each new sample may be mapped to a valid detector location in the detector array, and the first predetermined threshold may define the number of neighbors to a valid detector location that contribute to the generation of the new sample.

[0117] The second predetermined threshold effectively limits the number of sample sets, i.e., samples in different projection views, that contribute to the generation of the new sample. In particular, each sample may be mapped to a valid sample set, and the second predetermined threshold may define the number of neighboring sample sets relative to the valid sample set that contribute to the generation of the new sample.

[0118] FIG. 7 illustrates the effectiveness of a machine learning algorithm trained using a training dataset generated according to the fourth example above.

[0119] The first CT image 710 represents image data generated from projection data that has not been processed by such machine learning algorithms, and the second CT image 720 represents image data generated from projection data that has been processed by such machine learning algorithms.

[0120] The second CT image can be clearly identified as being of better resolution than the first CT image, providing potentially valuable information to assist the clinician in diagnosing or assessing the condition of the imaged subject.

[0121] For example, the identified portion 725 of the second CT image exhibits less aliasing than the equivalent portion 715 (showing the same features / elements) of the first CT image, clearly illustrating how the resolution and image quality of the image data is improved by the approaches disclosed herein.

[0122] FIG. 8 illustrates a method 800 according to one embodiment.

[0123] The method 800 includes acquiring 810 projection data generated by a CT scanner.

[0124] Step 810 may include generating the projection data using a CT scanner. In other examples, step 810 may include obtaining the projection data from, for example, a memory or storage unit that stores projection data previously generated by the CT scanner.

[0125] The method 800 also includes a step 820 of processing the projection data using a machine learning (ML) algorithm configured to perform a super-resolution imaging technique on the projection data to increase the apparent sampling of the projection data in at least one dimension.

[0126] The method 800 further includes a step 830 of outputting the processed projection data.

[0127] In some examples, the method 800 may further include generating 840 image data from the output and processed projection data. Approaches for generating image data from projection data are well established in the art, and backpropagation techniques may be used.

[0128] The method 800 may further include a step 850 of providing a visual representation of the generated image data, which may include controlling a user interface, such as a display, to provide the visual representation of the image data.

[0129] Various approaches to training machine learning methods, i.e., using training data sets to train machine learning methods, have been described herein. Approaches to generating training data sets have been described. Thus, one skilled in the art will be able to readily configure a computer-implemented method for generating training data sets for use in the present invention.

[0130] Therefore, a computer-implemented method for generating a training data set for training a machine learning algorithm that performs a super-resolution imaging technique on projection data generated by a CT scanner is also proposed.

[0131] The computer-implemented method includes generating an output training data set formed from a plurality of output training data entries, each including high-resolution projection data of an imaged subject for the training data set, by receiving intermediate data generated by a training CT scanner that uses a dual focus acquisition technique to generate the intermediate data; performing parallel binning on the interleaved first and second sample sets to generate the high-resolution projection data; and generating an input training data set formed from a plurality of input training data entries, each input training data entry corresponding to a respective output training data entry and including low-resolution image projection data of the same subject of the respective output training data entry.

[0132] Those skilled in the art will be readily able to configure a processing system to perform the methods or approaches described herein, and thus each block in any illustrated flowchart may represent a module of the processing system or a process performed by the processing system.

[0133] In some examples, such a processing system is integrated into a reconstruction unit of an imaging system such as that shown in Figure 1. The imaging system may further include a CT scanner.

[0134] 9 shows an example of a processing system 900 in which one or more parts of the embodiments may be used. The processing system 900 may form part of a reconstruction device and / or an imaging system.

[0135] The various operations described above may utilize the capabilities of the processing system 900. For example, one or more portions of the system for performing super-resolution techniques on projection data may be incorporated into any of the elements, modules, applications, and / or components described herein. In this regard, it should be understood that the system functional blocks may be executed on a single computer or may be distributed across several computers and locations, for example, connected via the Internet.

[0136] The processing system 900 may include, but is not limited to, a PC, a workstation, a laptop, a PDA, a palm device, a server, storage, and the like. In general, with respect to a hardware architecture, the processing system 900 may include one or more processors 901, a memory 902, and one or more I / O devices 907 communicatively coupled via a local interface (not shown). The local interface may be, for example, but not limited to, one or more buses or other wired or wireless connections, as known in the art. The local interface may have additional elements, such as controllers, buffers (caches), drivers, repeaters, and receivers, to enable communication. Additionally, the local interface may include address, control, and / or data connections to enable appropriate communication between the aforementioned components.

[0137] The processor 901 is a hardware device for executing software that may be stored in the memory 902. The processor 901 may be virtually any custom or commercially available processor, a central processing unit (CPU), a digital signal processor (DSP), or coprocessor of any number of processors associated with the processing system 900, and the processor 901 may be a semiconductor-based microprocessor (in the form of a microchip) or microprocessor.

[0138] The memory 902 may include any one or combination of volatile memory elements, random access memory (RAM), such as dynamic random access memory (DRAM), static random access memory (SRAM), and non-volatile memory elements, ROM, erasable programmable read only memory (EPROM), electronically erasable programmable read only memory (EEPROM), programmable read only memory (PROM), tape, compact hard disk read only memory (CDROM), hard disk, diskette, cartridge, cassette, etc. Additionally, the memory 902 may incorporate electronic, magnetic, optical, and / or other types of storage media. It should be noted that the memory 902 may have a distributed architecture in which various components are located remotely from each other, but may be accessed by the processor 901.

[0139] The software in memory 902 may include one or more separate programs, each of which comprises an ordered list of executable instructions for implementing logical functions. The software in memory 902 includes a suitable operating system (O / S) 905, a compiler 904, source code 903, and one or more applications 906, according to an exemplary embodiment. As shown, the applications 906 comprise a number of functional components for implementing the features and operations of the exemplary embodiments. The applications 906 of the processing system 900 may represent various applications, computational units, logic, functional units, processes, operations, virtual entities, and / or modules, according to an exemplary embodiment, although the applications 906 are not meant to be limiting.

[0140] Operating system 905 controls the execution of other computer programs and provides scheduling, input / output control, file and data management, memory management, and communication control and related services. It is contemplated by the inventors that applications 906 for implementing exemplary embodiments may be applicable to all commercially available operating systems.

[0141] The application 906 may be a source program, an executable program (object code), a script, or any other entity comprising a set of instructions to be executed. In the case of a source program, the program is typically translated via a compiler (such as compiler 904), assembler, interpreter, etc., which may or may not be included in memory 902, so that the program operates appropriately in conjunction with the O / S 905. Furthermore, the application 906 may be written as an object-oriented programming language with classes of data and methods, with routines, subroutines, and / or functions, or a procedural programming language, such as, but not limited to, C, C++, C#, Pascal, BASIC, API calls, HTML, XHTML, XML, ASP script, JavaScript, FORTRAN, COBOL, Perl, Java, ADA, NET, etc.

[0142] The I / O devices 907 may include input devices such as, but not limited to, a mouse, keyboard, scanner, microphone, camera, etc. Additionally, the I / O devices 907 may also include output devices such as, but not limited to, a printer, display, etc. Finally, the I / O devices 907 may further include devices that communicate both input and output, such as, but not limited to, network interface cards or modulators / demodulators (for accessing remote devices, other files, devices, systems, or networks), radio frequency (RF) or other transceivers, telephone interfaces, bridges, routers, etc. The I / O devices 907 also include components for communicating over various networks, such as the Internet or an intranet.

[0143] If the processing system 900 is a PC, workstation, or other intelligent device, the software in memory 902 may further include a Basic Input / Output System (BIOS) (omitted for simplicity). The BIOS is a set of essential software routines that initializes and tests the hardware at power-on, starts the O / S 905, and supports data transfers between hardware devices. The BIOS may be stored in some type of read-only memory, such as a ROM, PROM, EPROM, EEPROM, etc., so that the BIOS may be executed when the processing system 900 is powered up.

[0144] When the processing system 900 is in operation, the processor 901 is configured to execute software stored in the memory 902, to communicate data to and from the memory 902, and to generally control the operation of the processing system 900 in accordance with the software. The applications 906 and O / S 905 are read, in whole or in part, by the processor 901, possibly buffered within the processor 901, and then executed.

[0145] It should be noted that when the application 906 is implemented in software, the application 906 may be stored on virtually any computer-readable medium for use by or in connection with any computer-related system or method. In the context of this document, a computer-readable medium may be an electronic, magnetic, optical, or other physical device or means that can contain or store a computer program for use by or in connection with a computer-related system or method.

[0146] The application 906 may be embodied on any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, a processor-containing system, or other system that can fetch instructions from and execute instructions from an instruction execution system, apparatus, or device. In the context of this specification, a "computer-readable medium" may be any means that can store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. A computer-readable medium may be, for example, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, device, or propagation medium.

[0147] It will be understood that the disclosed methods are preferably computer-implemented methods. Thus, the concept of a computer program comprising code means for implementing any described method when said program is executed on a processing system such as a computer is also proposed. Thus, different parts, lines or blocks of code of a computer program according to an embodiment may be executed by a processing system or a computer to perform the methods described herein. In some alternative implementations, the functions depicted in the block diagrams or flow charts may occur out of the order depicted in the figures. For example, two blocks shown in succession may in fact be executed substantially simultaneously or may be executed in the reverse order depending on the functionality involved in the blocks.

[0148] Variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain means are recited in mutually different dependent claims does not indicate that a combination of these means cannot be used to advantage. Where a computer program is described above, it may be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium, provided together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. It should be noted that when the term "adapted to" is used in the claims or description, the term "adapted to" is intended to be equivalent to the term "configured to". When the term "one or more" is used, this is intended to be equivalent to the term "at least one" and vice versa. Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. 1. A computer-implemented method for processing projection data generated by a computed tomography scanner, the computer-implemented method comprising: acquiring projection data generated by the computed tomography scanner; processing the projection data using a machine learning algorithm configured to perform a super-resolution imaging technique on the projection data to increase the apparent sampling of the projection data in at least one dimension; outputting the processed projection data; and The machine learning algorithm is trained using a training dataset, the training dataset comprising: an input training data set formed from a plurality of input training data entries each having low resolution projection data of an imaged subject; an output training data set formed from a plurality of output training data entries, each output training data entry corresponding to a respective input training data entry and having high resolution projection data of the same imaged subject as the respective input training data entry; and The high resolution projection data for each output training data entry is generating intermediate data using a dual focus acquisition technique, the intermediate data having interleaved first and second sample sets, each sample set acquired using a different focus; performing parallel binning on the interleaved first and second sample sets to generate the high-resolution projection data; generated by a training computed tomography scanner that generates high resolution projection data by performing the low-resolution projection data is generated by discarding a first sample set or a second sample set of the intermediate data. Computer-implemented methods.

2. The computer-implemented method of claim 1 , wherein the projection data is projection data generated by the computed tomography scanner using a dual-focus acquisition approach.

3. The machine learning algorithm is further trained using a modified training dataset, the modified training dataset comprising: a modified input training data set in which noise is added to the low-resolution projection data of each input training data entry of said training data set; an output training dataset of said training dataset; 2. The computer-implemented method of claim 1, comprising:

4. The machine learning algorithm is further trained using a second training data set, the second training data set comprising: a second input training data set formed from a plurality of second input training data entries, each having low resolution image data of a scene; a second output training data set formed from a plurality of second output training data entries, each second output training data entry corresponding to a respective second input training data entry and having high resolution image data of the same scene as said respective second input training data entry; 2. The computer-implemented method of claim 1, comprising:

5. 5. The computer-implemented method of claim 4, wherein values ​​of the low-resolution image data and the high-resolution image data are scaled to correspond to a range of possible values ​​of projection data produced by the computed tomography scanner.

6. the projection data comprises a plurality of samples, each associated with a different combination of values ​​for the parameter r and the angle Φ; the parameter r represents the minimum distance between the isocenter of the computed tomography scanner and the radiation used by the computed tomography scanner to generate the sample; Angle Φ represents the angle between a ray used by the computed tomography scanner to generate the sample and a predefined plane; the machine learning algorithm is configured to generate a plurality of new samples for the projection data; Each new sample is generated from a different subset of the plurality of samples, the subsets comprising: the absolute difference between the value of parameter r of the sample and the value of parameter r of the new sample is below a first predetermined threshold, and / or the absolute difference between the value of the angle Φ of the sample and the value of the angle Φ of the new sample is below a second predetermined threshold having only the sample, 10. The computer-implemented method of claim 1.

7. 1. A computer-implemented method for training a machine learning algorithm to generate a training data set for performing a super-resolution imaging technique on projection data generated by a computed tomography scanner, the computer-implemented method comprising: receiving intermediate data generated by a training computed tomography scanner that generates intermediate data using a dual focus acquisition technique, the intermediate data having interleaved first and second sample sets, each sample set acquired using a different focus; performing parallel binning on the interleaved first and second sample sets to generate the high-resolution projection data; By running generating an output training data set for the training data set formed from a plurality of output training data entries each having high resolution projection data of an imaged subject; generating an input training data set formed from a plurality of input training data entries, each input training data entry corresponding to a respective output training data entry, the input training data entry having low resolution image projection data of the same image subject as the respective output training data entry by discarding the first sample set or the second sample set of the intermediate data; 10. A computer-implemented method comprising:

8. 8. The computer-implemented method of claim 7, wherein generating the output training data set further comprises controlling the training computed tomography to generate the intermediate data using a dual focus acquisition technique.

9. 9. A non-transitory computer program product comprising computer program code that, when executed on a computing device having a processing system, causes the processing system to perform all of the steps of the method of any one of claims 1 to 8.

10. 1. A processing system for processing projection data produced by a computed tomography scanner, the processing system comprising: acquiring projection data generated by the computed tomography scanner; processing the projection data using a machine learning algorithm configured to perform a super-resolution imaging technique on the projection data to increase the apparent sampling of the projection data in at least one dimension; outputting the processed projection data; configured to run The machine learning algorithm is trained using a training dataset, the training dataset comprising: an input training data set formed from a plurality of input training data entries each having low resolution projection data of an imaged subject; an output training data set formed from a plurality of output training data entries, each output training data entry corresponding to a respective input training data entry and having high resolution projection data of the same imaged subject as the respective input training data entry; and The high resolution projection data for each output training data entry is generating intermediate data using a dual focus acquisition technique, the intermediate data having interleaved first and second sample sets, each sample set acquired using a different focus; performing parallel binning on the interleaved first and second sample sets to generate the high-resolution projection data; a training computed tomography scanner that generates the high-resolution projection data by performing the low-resolution projection data is generated by discarding a first sample set or a second sample set of the intermediate data. Processing system.

11. The processing system of claim 10 , wherein the projection data is projection data generated by the computed tomography scanner using a dual-focus acquisition approach.

12. The machine learning algorithm is further trained using a modified training dataset, the modified training dataset comprising: a modified input training data set in which noise is added to the low-resolution projection data of each input training data entry of said training data set; an output training dataset of said training dataset; The processing system of claim 10 , comprising:

13. The machine learning algorithm is further trained using a second training data set, the second training data set comprising: a second input training data set formed from a plurality of second input training data entries, each having low resolution image data of a scene; a second output training data set formed from a plurality of second output training data entries, each second output training data entry corresponding to a respective second input training data entry and having high resolution image data of the same scene as said respective second input training data entry; The processing system of claim 10 , comprising:

14. 14. The processing system of claim 13, wherein values ​​of the low-resolution image data and the high-resolution image data are scaled to correspond to a range of possible values ​​of projection data produced by the computed tomography scanner.

15. A processing system according to any one of claims 10 to 14; said computed tomography scanner; An imaging system having:

Images