Processing projection data

Abstract

A mechanism for producing image data, from projection data, with reduced noise. The projection data is filtered, before undergoing a denoising procedure using one or more machine-learning algorithms. The filtered and denoised projection data is then processed using a reconstruction algorithm to produce the image data.

Description

[0001] PROCESSING PROJECTION DATA

[0002] FIELD OF THE INVENTION

[0003] The present disclosure relates to the field of computed tomography imaging, and in particular to the processing of projection data.

[0004] BACKGROUND OF THE INVENTION

[0005] There is an ongoing interest in the performance of (non-invasive) imaging of a subject or patient. In modern medicine, images produced using such techniques are important for aiding in the performance of diagnosis and analysis of a condition of the subject or patient. A computed tomography (CT) scan provides one approach for performing imaging of the subject.

[0006] In particular, computed tomography (CT) scanners are well-established medical imaging devices that use a detector to detect an interaction between X-ray radiation and irradiated material in order to generate projection data of an imaged area. The projection data is then processed using a reconstruction algorithm, such as the filtered back-projection (FBP) algorithm, to produce image data.

[0007] Other forms of medical imaging devices may be designed to produce raw acquisition data during a scanning procedure, including magnetic resonance imaging (MRI) devices and positron emission tomography (PET) imaging devices.

[0008] There is an ongoing desire to reduce the noise present in image data produced using any such medical devices.

[0009] In particular, for a CT scanner, noise can be introduced into an image data when performing the imaging scan using a relatively low dose of X-ray radiation. Reliable and accurate noise reduction techniques thereby allow for lower dosages of X-ray radiation, i.e., ionizing radiation, to be applied to a subject during a CT scanning procedure.

[0010] SUMMARY OF THE INVENTION

[0011] The invention is defined by the claims.

[0012] According to examples in accordance with an aspect of the invention, there is provided a computer-implemented method for processing projection data for a medical imaging system.

[0013] The computer-implemented method comprises: receiving projection data of a subject, wherein the projection data is generated using the medical imaging system; processing the projection data using at least one filter to produce filtered projection data; processing the filtered projection data, using at least one machine-learning algorithm trained to perform a denoising procedure on filtered projection data, to produce denoised projection data; and processing the denoised projection data using a reconstruction algorithm to produce image data.

[0014] The present disclosure provides a technique for improved generation of image data from projection data with less noise . The proposed approach is to effectively split a filtered reconstruction (e.g., filtered back projection) procedure into a filtering step and a reconstruction step, and performing denoising between the two steps. The denoising is performed using at least one machine-learning algorithm.

[0015] The proposed approach achieves the reduction of noise (at least compared to a procedure in which no denoising is performed) with a reduced risk of blurring, thereby reducing a likelihood that potentially important elements and / or pathologies are rendered imperceptible and / or less clearly delineated from one another as a result of inappropriate or poorly performing denoising of image data.

[0016] When employed for processing projection data for a CT scanner or imaging system, improved denoising when producing image data facilitates the use of lower radiation dosages when generating the projection data for the CT scanner. This advantageously reduces exposure of a subject to potentially harmful radiation, e.g., ionizing radiation.

[0017] In some examples, the at least one filter comprises a high-pass filter. For instance, the at least one filter may comprise a ramp filter. The use of a high-pass filter, particularly a ramp filter, has been identified as contributing to the provision of high-quality and low noise image data.

[0018] The at least one machine-learning algorithm may comprise a neural network. This provides a reliable and robust approach for performing denoising of filtered projection data.

[0019] In some examples, the step of processing the filtered projection data comprises processing the projection data or the filtered projection data to identify anatomical data that indicates one or more anatomical elements and / or anatomical regions represented by the filtered projection data; and processing the anatomical data and the filtered projection data using the at least one machine-learning algorithm to produce the denoised projection data.

[0020] This approach recognizes that the shape and structure of different anatomical elements and / or anatomies means that improved denoising can be achieved using anatomyspecific denoising procedures or by otherwise taking the anatomical view or representation in account. Improved denoising is thereby achieved by the use of anatomical information in the denoising procedure.

[0021] In some examples, processing the anatomical data and the filtered projection data comprises: selecting one or more of a plurality of candidate machine-learning algorithms using the anatomical data, wherein different candidate machine-learning algorithms are associated with different anatomical elements and / or anatomical regions; and processing the filtered projection data using the selected one or more candidate machine-learning algorithms.

[0022] This provides anatomy-specific sets of one or more machine-learning algorithms. This approach may have improved reliability in the performance of denoising, as there is a reduced risk of crosstalk or conflict between different types of anatomy during the training of the machine-learning algorithms, e.g., if the noise characteristics for different anatomies differ.

[0023] In some examples, the anatomical data and the filtered projection data comprises inputting the anatomical data and the filtered projection data into the one or more machinelearning algorithms to produce the denoised projection data. This provides a simpler and more resource efficient approach to performing denoising of the filtered projection data, e.g., without a need to store and / or identify any particular machine-learning algorithms for producing the denoised projection data.

[0024] In some examples, the filtered projection data comprises a plurality of different filtered projection data portions and the anatomical data comprises, for each filtered projection data portion, a respective anatomical data portion that indicates one or more anatomical elements and / or anatomical regions represented by the filtered projection data portion.

[0025] In such examples, processing the anatomical data and the filtered projection data may comprise, for each filtered projection data portion, processing the filtered projection data portion and the respective anatomical data portion using a subset (e.g., not all) of the at least one machine-learning algorithm to produce the denoised projection data.

[0026] In this way, different parts of the projection data, representing different anatomical features / regions, can be individually denoised. This advantageously achieves anatomyspecific denoising of the projection data even when the projection data spans or represents multiple different anatomical elements and / or regions. The filtered projection data may comprise a plurality of different filtered sets of projection data elements, each filtered set being associated with a different projection angle at which the projection data elements were captured by the medical imaging system.

[0027] In such examples, processing the filtered projection data comprises, for each filtered set of projection data elements, processing the filtered set and the associated projection angle using at least one machine-learning algorithm to produce denoised projection data. This approach recognizes that similar projection angles (i.e., angles of scanning) are likely to capture or scan similar object, e.g., anatomical objects / elements. By denoising the projection data using the projection angle(s), then differing anatomical elements can be taken into account during the denoising procedure. This leads to improved denoising.

[0028] The method may further comprise obtaining initial projection data of the subject, wherein the initial projection data is generated using the medical imaging system; and performing one or more pre-processing and / or rebinning functions on the initial projection data to produce the projection data.

[0029] The computer-implemented method may comprise processing the image data using an image denoising algorithm to produce denoised image data. This approach provides additional denoising in producing the image data, to thereby improve the overall denoising performed by the method.

[0030] In some examples, the method further comprises processing the denoised image data using image-domain filtering to produce output image data, wherein the image-domain filtering is configured to performing filtering on the denoised image data in accordance with a predefined protocol. This achieves modification or adaptation of the image data to have one or more desired image characteristics.

[0031] In some examples, the least one filter is designed for use in a filtered back-projection technique usable to reconstruct image data from projection data.

[0032] There is also provided a computer program product comprising computer program code means which, when executed on a device having processing circuitry, cause the processing circuitry to perform all of the steps of any herein proposed computer-implemented method.

[0033] There is also proposed a device for processing projection data for a medical imaging system. The device comprises: processing circuitry; and a memory containing instructions that, when executed by the processing circuitry, configure the processing circuitry to: obtain projection data of a subject, wherein the projection data is generated using the medical imaging system; process the projection data using at least one filter to produce filtered projection data; process the filtered projection data, using at least one machine-learning algorithm trained to perform a denoising procedure on filtered projection data, to produce denoised projection data; and process the denoised projection data using a reconstruction algorithm to produce image data.

[0034] The instructions contained by the memory may be configured to, when executed by the processing circuitry, configure the processing circuit to carry out any herein proposed method. The skilled person would be capable of appropriately adapting the instructions contained by the memory to configure the processing circuit to carry out any herein proposed method.

[0035] There is also proposed a system comprising: the herein proposed device; and the medical imaging system configured to generate the projection data of the subject.

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

[0037] BRIEF DESCRIPTION OF THE DRAWINGS

[0038] For a better understanding of the 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:

[0039] Figure 1 illustrates a system in which embodiments may be employed;

[0040] Figure 2 illustrates an existing method for generating image data;

[0041] Figure 3 illustrates a proposed method for generating image data;

[0042] Figure 4 illustrates images produced using different methods of processing projection data;

[0043] Figure 5 illustrates a variant proposed method for generating image data;

[0044] Figure 6 illustrates another variant proposed method for generating image data; and Figure 7 illustrates optional steps for use in proposed methods.

[0045] DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The invention will be described with reference to the Figures.

[0047] It should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the apparatus, systems and methods, are intended 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 apparatus, 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 indicate the same or similar parts.

[0048] The invention provides a mechanism for producing image data, from projection data, with reduced noise. The projection data is filtered, before undergoing a denoising procedure using one or more machine-learning algorithms. The filtered and denoised projection data is then processed using a reconstruction algorithm to produce the image data.

[0049] Embodiments are based on the realization that existing projection domain denoising (such as machine-learning-based denoising) is performed before filtering, as a result of the near-ubiquitous use of filtered back projection to perform image reconstruction. It has been identified that such existing projection-domain denoising does result in noise reduction but at the costs of blurred reconstructed images.

[0050] The present disclosure recognizes that that one possible explanation for the phenomenon is the discrepancy between optima in projection and image spaces. More particularly, because the filtering filters projection data, training a machine-learning algorithm to perform denoising before the filter can make the denoising algorithm suboptimal with respect to the filtered frequencies.

[0051] Unlike previously known approaches, it is herein proposed to train and / or use a denoising machine-learning algorithm after performing filtering of projection data and before performing reconstruction of the projection data into image data. In this way, denoising is still performed in the projection domain, but without reduced risk of blurring image data constructed from the denoised projection data.

[0052] Figure 1 illustrates a system 100 in which embodiments may be employed, for improved contextual understanding. The system 100 comprises a CT scanning system 110 and a device 120.

[0053] The CT scanning system 110 is configured to capture and / or generate (CT) projection data 150 of a subject 190 during a single scanning procedure. In this context, the projection data comprises the raw data captured by the CT scanning system 110 during the single scanning procedure, before reconstruction has been performed.

[0054] By way of explanation, the projection data may be captured by iteratively sampling a set of projection data elements using a detection system 111, as the subject 190 and detection system are moved with respect to one another. Each set therefore represents data captured during a single, respective capture period (i.e., during one iteration of the sampling). In this way, the projection data comprises a sequence of sets of projection data elements. In producing a projection data element, it is known to apply one or more weightings (e.g., a weighting function) to different values in the projection data element.

[0055] In general, during a scanning procedure, the detection system rotates around the subject. The rotation of the detection system with respect to the subject is definable using a rotation angle, also known as a projection angle. In particular, each set of projection data elements may be associated with a respective rotation / proj ection angle, representing the rotation angle of the detection system with respect to the subject during capture of the projection data element(s) of the set.

[0056] Known types of scanning procedures include circular CT scanning procedures and helical CT scanning procedures. The function and design of CT scanning systems for carrying out such CT scanning procedures are well known and established in the art.

[0057] The device 120 comprises processing circuitry 121 and a memory 122. The memory contains instructions that, when executed by the processing circuitry, configure the processing circuitry to perform one or more tasks or functions. The device 120 may, for instance, be replaced by any other form of processing system.

[0058] The device 120 may be communicatively coupled to the CT scanning system 110 so as to receive at least projection data 150 from the CT scanning system. The communicative coupling may be wired or wireless, and approaches are known in the art.

[0059] In other approaches, the CT scanning system 110 may store the projection data 150 in a memory or storage unit 130 (which may form part of the system 100). The device 120 may be communicatively coupled to the memory or storage unit, e.g., to receive or obtain projection from the memory storage unit 130. The communicative coupling may be wired or wireless, and approaches are known in the art.

[0060] The processing circuitry 121 may include, but is not limited to, one or more of the following: conventional microprocessors, application specific integrated circuits (ASICs), and / or field-programmable gate arrays (FPGAs). The memory 122 may comprise any volatile and / or non-volatile computer memory such as RAM, PROM, EPROM, and EEPROM. The instructions contained in the memory may effectively define one or more programs that, when executed on the processing circuitry, cause the processing circuitry to perform encoded functions.

[0061] As a typical use-case scenario, the processing system 120 is often configured to obtain or receive the projection data 150, produced by the CT scanning system 110, and perform reconstruction on the projection data 150 to thereby produce CT image data. The processing system 120 may then control a user interface 140 (e.g., a screen) to provide a visual representation of the CT image data. Approaches for controlling a user interface in this way are well established in the art.

[0062] The present disclosure proposed an improved mechanism for performing reconstruction of the projection data. In particular, the proposed approach facilitates the generation of image data with reduced noise and with a reduced risk of blurring in the image data.

[0063] Although the above described system makes use of a CT scanning system 110 to produce projection data, the skilled person will readily appreciate how any other suitable form of medical imaging device or system could be employed to produce projection data to be processed according to the herein proposed method, mechanism and / or processing circuitry. For instance, the CT scanning system 110 may be replaced by an MRI system or a PET system. Other suitable medical imaging systems will be readily apparent to the skilled person.

[0064] Bearing this in mind, for the sake of explanative clarity, the following description will assume that the projection data is produced using a CT scanning system.

[0065] Figure 2 is a flowchart illustrating an existing method 200 for processing projection data to generate or reconstruct image data from the projection data. The projection data is generated using a medical imaging system.

[0066] The method 200 comprises a step 210 of receiving / obtaining projection data of the subject. The method also comprises a step 220 of denoising the projection data using a projection-domain denoising algorithm or function, which step may be labelled projection domain denoising. The method 200 then performs a step 230 of processing the denoised projection data using a filtered back-projection technique, which effectively comprises performing a sub-step 231 of filtering the denoised projection data before performing a substep 232 of processing the filtered projection data using back-projection algorithm.

[0067] The output of step 230 is image data, e.g., CT image data. The image data may, in some known examples, undergo further denoising in a step 240, known as image-domain denoising.

[0068] By way of example only, the method 200 outlined above is exploited by Zhang, Haiyan, et al. "Projection domain denoising method based on dictionary learning for low- dose CT image reconstruction." Journal of X-ray Science and Technology 23.5 (2015): 567- 578.

[0069] The present disclosure proposes an alternative technique to performing denoising within a method for processing projection data to generate or reconstruct image data. More specifically, the proposed approach applies a denoising step after performing filtering of the projection data. The proposed approach effectively splits the filtered back-projection algorithm such that denoising, using a machine-learning algorithm, is performed between filtering and performing back-projection.

[0070] Any herein proposed computer-implemented method may, for instance, be performed by the device 120 previously described. More particularly, the computer-implemented method may be performed by the processing circuitry 121 of the device 120 when executing instructions contained in the memory 122.

[0071] Figure 3 is a flowchart illustrating a proposed computer-implemented method 300 for processing projection data, generated using a medical imaging system, to generate or reconstruct image data.

[0072] The method 300 comprises a step 310 of receiving projection data of a subject. As previously explained, the projection data is generated using the medical imaging system. The projection data may be received / obtained directly from the medical imaging system itself. Alternatively, the projection data may be received / obtained from a memory communicatively connected to the medical imaging system. Other approaches for storing and providing projection data will be apparent to the skilled person.

[0073] The method 300 also comprises a step 320 of processing the projection data using at least one filter to produce filtered projection data. The at least one filter may comprise one or more filters typically used in the application of a filtered back-projection (e.g., reconstruction) algorithm for processing projection data. Examples include high-pass filters, such as a ramp filter, although other examples are known in the art.

[0074] Specific examples of suitable filters usable for filtered back projection include the Ram-Lak filter (i.e. a ramp filter), the Shepp-Logan filter, a cosine filter, a Hamming filter and so on. More specific examples of suitable filters usable for filtered back projection are suggested by, amongst others, Horbelt, Stefan, Michael Liebling, and Michael A. Unser. "Filter design for filtered back-projection guided by the interpolation model." Medical Imaging 2002: Image Processing. Vol. 4684. SPIE, 2002,

[0075] The step 320 of processing the projection data using at least one filter may be performed without using a machine-learning method. This improves a processing efficiency of the computer-implemented method 300, as it has been recognized that the proposed approach avoids or mitigates a need to perform any such filtering using a machine-learning method whilst still achieving a high-quality denoising procedure. The method 300 further comprises a step 330 of processing the filtered projection data, using at least one machine-learning algorithm trained to perform a denoising procedure on filtered projection data, to produce denoised projection data. In general, step 330 comprises inputting at least the filtered projection data into at least one machine-learning algorithm which outputs the denoised projection data.

[0076] Further detail on examples of suitable machine-learning algorithms and / or approaches for training the machine-learning algorithms is provided later.

[0077] In some examples, step 330 is the only step that performs a denoising procedure using one or more machine-learning algorithms in the projection domain, i.e., before any reconstruction is performed.

[0078] The method 300 further comprises a step 340 of processing the denoised projection data using a reconstruction algorithm to produce image data. A suitable example of a reconstruction algorithm is a back projection algorithm.

[0079] Appropriate approaches for performing reconstruction of (e.g., filtered and denoised) projection data are widely known in the art, including those that are based upon the inverse of the Radon transform (i.e., back projection), as explained by, inter alia, Beckmann, Matthias, and Armin Iske. "Error estimates and convergence rates for filtered back projection." Mathematics of Computation 88.316 (2019): 801-835.

[0080] In this way, projection domain denoising is performed between the steps of filtering the projection data and reconstructing the projection data, e.g., using a back projection technique.

[0081] In preferred examples, no further filtering of the projection data is performed between the denoising procedure and the reconstruction procedure. In particular, in preferred examples, no further high-pass and / or ramp filter is applied to the denoised projection data before it is processed using the at least one machine-learning algorithm. This reduces a risk of blurring in the image data subsequently produced.

[0082] Step 330 makes use of at least one machine-learning algorithm to perform denoising of filtered projection data, i.e., projection data that has undergone a filtered procedure.

[0083] A machine-learning algorithm is any self-training / self-learning algorithm that processes input data in order to produce or predict output data. Here, the input data comprises filtered projection data and the output data comprises denoised projection data.

[0084] Suitable machine-learning algorithms for being employed in the present invention will be apparent to the skilled person. Examples of suitable machine-learning algorithms include decision tree algorithms and artificial neural networks. Other machine-learning algorithms such as logistic regression, support vector machines or Naive Bayesian models are suitable alternatives.

[0085] The structure of an artificial neural network (or, simply, neural network) is inspired by the human brain. Neural networks are comprised of layers, each layer comprising a plurality of neurons. Each neuron comprises 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, sigmoid etc. but with different weightings). In the process of processing input data, the mathematical operation of each neuron is performed on the input data to produce a numerical output, and the outputs of each layer in the neural network are fed into the next layer sequentially. The final layer provides the output.

[0086] A decision tree algorithm processes input data through a tree of nodes. In the tree of nodes, each successive node splits into two or more further nodes until reaching a terminal or end node. When performing the decision tree algorithm using the tree of nodes, at each node, a decision is made as to which further node to move to next based on the input data. The end node defines the outcome of the decision tree algorithm, and therefore the machine-learning algorithm.

[0087] Methods of training a machine-learning algorithm are well known. Typically, such methods comprise obtaining a training dataset, comprising training input data entries and corresponding training output data entries. The training input data entries provide example inputs for the machine-learning algorithm and the training output data entries provide target outputs for the machine-learning algorithm when processing the corresponding example inputs.

[0088] For many machine-learning algorithms, such as a neural network or decision tree, training is performed by applying an initialized machine-learning algorithm to each input data entry to generate predicted output data entries. An error between the predicted output data entries and corresponding training output data entries is used to modify the machinelearning algorithm. This process can be repeated until the error converges, and the predicted output data entries are sufficiently similar (e.g. ±1%) to the training output data entries. This is commonly known as a supervised learning technique.

[0089] For example, where the machine-learning algorithm is formed from a neural network, (weightings of) the mathematical operation of each neuron may be modified until the error converges. Known methods of modifying a neural network include gradient descent, backpropagation algorithms and so on. Other approaches for training machine-learning algorithms (e.g., decision trees) are known in the art. For instance, decision trees are often trained using a decision tree builder or learning techniques, such as those set out by Suthaharan, Shan, and Shan Suthaharan. "Decision tree learning." Machine Learning Models and Algorithms for Big Data Classification: Thinking with Examples for Effective Learning (2016): 237-269 or Ruggieri, Salvatore. "Yadt: Yet another decision tree builder." 16th IEEE International Conference on Tools with Artificial Intelligence. IEEE, 2004.

[0090] For the purposes of performing step 330, the training input data entries correspond to example instances of filtered projection data. The training output data entries correspond to instances of denoised projection data.

[0091] One approach to generating the training input data entries and the training output data entries can be as follows. Firstly, collect a plurality of instances of projection data. Then, add noise (e.g., pseudo-randomly generated noise) to each of the plurality of instances of projection data. This produces two sets of projection data, the first set including the originally collected instances of projection data and the second set including the noise-added instances of projection data. Each instance of projection data is then processed using the same one or more filters to produce respective filtered instances of projection data. The first set of filtered instances of projection data act as the training output data entries. The second set of filtered instances of projection data act as the training input data entries.

[0092] Another approach to generating the training input data entries and the training output data entries is hereafter described. This approach is designed for use where the projection data is produced using a CT scanner, i.e., is CT projection data. In this approach, a plurality of pairs of instances of projection data are obtained. Each pair of instances of projection data comprises a first instance of projection data, obtained by a CT scanner using a first dosage of radiation, and a second instance of projection data, obtained by the CT scanner using a second dosage of radiation, the second dosage being greater than the first dosage. The first and the second instance of projection data represent a same imaged region. Each instance of projection data is then processed using the same one or more filters to produce respective pairs of filtered instances of projection data. Each filtered first instance of projection data acts as a respective training input data entry. Each filtered second instance of projection data acts as a respective training output data entry.

[0093] Figure 4 illustrates the effect of employing the proposed computer-implemented method 300 to process projection data to produce image data. In particular, Figure 4 illustrates four separate images, being slices of respective image data, each representing a same anatomical region. A first image 410 is produced by performing a filtered back propagation (i.e., filtering and reconstruction only) on first projection data without denoising. A second image 420 is produced by performing a denoising procedure on the first projection data before performing a filtered back propagation (i.e., filtering and reconstruction), i.e., a procedure as previously described with reference to Figure 2. A third image 430 is produced by performing a herein proposed computer- implemented method using the first projection data. A fourth image 440 is produced by performing a filtered back propagation (i.e., filtering and reconstruction only) on second projection data without denoising, wherein the second projection data is captured using a higher dosage, i.e., a higher radiation dosage, than the first projection data.

[0094] As clearly evident in Figure 4, the proposed approach significantly reduces the noise visible in the image (e.g., compare the first image 410 to the third image 430) without causing significant blurring in the image (e.g., compare the second image 420 to the third image 430). Moreover, the proposed approach more closely aligns the denoised image (third image 430) with an image produced using projection data captured a higher dosage (fourth image 440), thereby more closely resembling a true or ground-truth representation of the imaged subject, i.e., without blurring of potentially important elements.

[0095] Figure 5 illustrates a variant computer-implemented method 500 for processing projection data, generated using a medical imaging system, to generate or reconstruct image data. The method 500 comprises the steps 310, 320, 330 and 340 previously described, with any variations being hereafter detailed.

[0096] The computer-implemented method further comprises a step 510 of processing the projection data or the filtered projection data to identify anatomical data that indicates one or more anatomical elements and / or anatomical regions represented by the filtered projection data.

[0097] Step 510 may, for instance, be performed using an appropriately trained machinelearning method. This machine-learning method may receive, as input, the projection data or the filtered projection data and provide, as output, anatomical data.

[0098] The general principle and possible training techniques for machine-learning methods or algorithms has been previously described. In particular, a machine-learning method is trained using a training data comprising training input data entries and training output data entries. For the purposes of step 510, the machine-learning method may be trained using, as training input data entries, example instances of projection data and, as training output data entries, example instances of anatomical data (e.g., labels of any anatomical elements and / or regions). The training output data entries may be generated or defined by an appropriately trained clinician reviewing image data produced by reconstructing (e.g., using pre-existing techniques) the example instances of projection data. Other approaches for defining or producing suitable training input data entries and output data entries will be apparent to the person skilled in the art.

[0099] In a simple alternative, step 510 may be performed by receiving a user input indicating one or more anatomical elements and / or anatomical regions represented by the filtered projection data. This can be trivially performed if, for instance, a clinician or operator knows a region that is being or has been imaged (e.g., knows that the head or torso has been imaged).

[0100] Step 330 is modified to comprise processing the anatomical data and the filtered projection data using the at least one machine-learning algorithm to produce the denoised projection data. Thus, the anatomical data identified in step 510 is used in the denoising procedure performed in step 330.

[0101] In a simple example, step 330 may comprise inputting (e.g., at least) the anatomical data and the filtered projection data into the one or more machine-learning algorithms. The machine-learning algorithm is then used to process the input data to produce the output data.

[0102] For this simple example, the machine-learning algorithm is trained further using example instances of anatomical data, e.g., for use as the example input data entries and the example output data entries. Appropriate adaptation of a training scheme and / or training dataset will be readily apparent to the appropriately skilled person familiar with machinelearning techniques.

[0103] Figure 5 illustrates an alternative example in which processing 330 the anatomical data and the filtered projection data comprises selecting 521 one or more of a plurality of candidate machine-learning algorithms using the anatomical data and processing 522 the filtered projection data using the selected one or more candidate machine-learning algorithms.

[0104] Different candidate machine-learning algorithms are associated with different anatomical elements and / or anatomical regions. It will be appreciated that different candidate machine-learning algorithms may, for instance, have a same or similar structure, but have different settings and / or parameterizations. For instance, each candidate machine-learning algorithm may be a neural network having a same number of layers and / or neurons (in each layer), but with different weightings for mathematical operations of the neurons. This can be achieved, for instance, by training each candidate machine-learning algorithm with a different training dataset. Each training dataset may, for instance, only comprise data that includes or at least partially represents the corresponding anatomical element and / or anatomical region. For instance, a first training dataset may comprise only instances of projection data of a heart, whereas a second training dataset may comprise only instances of projection data of the brain or skull.

[0105] In this way, step 521 can be trivially performed by selecting the machine-learning algorithm(s) associated with the relevant anatomical element and / or anatomical region. For instance, each machine-learning algorithm may be labelled or associated with metadata identifying the relevant anatomical element, e.g., in a look-up table or the like. Identification of the relevant machine-learning algorithm(s) is thereby performed trivially.

[0106] In an example, steps 510 and 330 are modified to be performed on different portions or sections of the filtered projection data.

[0107] In particular, the filtered projection data may comprise a plurality of different filtered projection data portions. This can be achieved, for instance, by dividing the projection data into a plurality of portions before performing step 320 on each projection data portion to produce a plurality of different portions of filtered projection data. Alternatively, this can be achieved by dividing filtered projection data, produced by performing step 320 on projection data, to produce a plurality of different portions of filtered projection data.

[0108] Step 510 may be performed on each filtered projection data portion or, if available, each projection data portion. In this way, the anatomical data comprises, for each filtered projection data portion, a respective anatomical data portion that indicates one or more anatomical elements and / or anatomical regions represented by the filtered projection data portion.

[0109] Step 330 may be performed for each projection data portion. In particular, step 330 may comprise processing, for each filtered projection data portion, processing the filtered projection data portion and the respective anatomical data portion using a subset of the at least one machine-learning algorithm to produce the denoised projection data. As a portion of projection data can, itself, be treated as an instance of projection data, then adapting step 330 to perform this function can be readily performed by the skilled person.

[0110] This option for performing step 330 for each projection data portion is conceptually illustrated using a dotted line. The skilled person will appreciate that, in some examples, a different instance of step 330 is performed (e.g., in parallel) for each projection data portion. In some examples, step 330 may comprise scaling 523 the denoising of the filtered projection data (e.g., each filtered projection data portion) responsive to the anatomical data.

[0111] By way of example, this can be performed by, after producing denoised projection data using the at least one machine-learning algorithm, performing a weighted average of the denoised projection data and the filtered projection data. Weightings are applied to (e.g., multiplied with) each projection data element of the denoised or filtered projection data. In this approach, the value of a weighting applied to each projection data element is responsive to the anatomical data.

[0112] For instance, if the anatomical data simply indicates the one or more anatomical elements and / or anatomical regions represented by the (overall) filtered projection data, then the value of each weighting may be responsive to the indicated one or more anatomical elements and / or anatomical regions. For instance, a look-up table or similar may be used to map the indicated one or more anatomical elements and / or anatomical regions to an appropriate weighting for each projection data element.

[0113] As another example, if the anatomical data indicates the one or more anatomical elements and / or anatomical regions represented by each projection data portion, then the value of the weighting for each projection data element may be responsive to the one or more anatomical elements and / or anatomical regions represented by the projection data portion to which the corresponding projection data element belongs. In a similar manner, a look-up table or similar may be used to map the indicated one or more anatomical elements and / or anatomical regions (of the corresponding projection data portion) to an appropriate weighting for each projection data element.

[0114] The step of performing scaling 523 may be performed independently of a step of selecting and using candidate machine-learning algorithms (i.e., step 521). In other words, step 521 may be omitted, and the filtered projection data simply processed in step 522 by one or more machine-learning algorithms, to perform denoising, without using the anatomical data. The anatomical data may then be subsequently employed to scale the denoising.

[0115] Figure 6 illustrates a variant computer-implemented method 600 for processing projection data, generated using a medical imaging system, to generate or reconstruct image data. The method 600 comprises the steps 310, 320, 330 and 340 previously described, with any variations being hereafter detailed.

[0116] It has previously been explained how projection data comprises a plurality of sets of projection data elements, and how each projection data element may be associated with a corresponding projection angle. Accordingly, it is possible to define the filtered projection data as comprising a plurality of different filtered sets of projection data elements, each filtered set being associated with a different projection angle. Each projection angle indicates at which angle the projection data elements were captured by the medical imaging system. The projection angle may be information output by the medical imaging system during capture of the projection data.

[0117] In this way, the method 600 may effectively comprise a step 610 of identifying a projection angle for each filtered set of projection data elements. The projection angle may, for instance, be stored in metadata for the filtered set of projection data elements or in a lookup table. Other suitable approaches will be known in the art.

[0118] Step 330 may be accordingly modified to comprise, for each filtered set of projection data elements, processing 630 the filtered set and the associated projection angle using at least one machine-learning algorithm to produce denoised projection data.

[0119] Thus, in some examples, step 330 may comprise, for each filtered set, inputting at least the filtered set and the associated projection angle into the at least one machine-learning algorithm. This produces denoised projection data.

[0120] In this approach, the machine-learning algorithm(s) used to process the filtered set and the associated projection angle will need to be trained appropriately, e.g., using appropriately configured training input data entries and training output data entries. More particularly, each training input data entry will comprise an example instance of a filtered set of projection data elements and a corresponding projection angle. Each output training data entry will comprise a corresponding example instance of a denoised set of projection data elements. Appropriate training input data entries and training output data entries can be produced using previously described approaches.

[0121] In some examples, step 330 may comprise scaling the denoising of each filtered set of projection data elements as a function of projection angle for said set.

[0122] By way of example, this can be performed by, after producing denoised projection data using the at least one machine-learning algorithm, performing a weighted average of the denoised projection data and the filtered projection data. Weightings are applied to (e.g., multiplied with) each projection data element of the denoised or filtered projection data. In this approach, the value of a weighting applied to each projection data element is a function of the projection angle to the set of projection data elements to which the projection data element belongs.

[0123] In this way, the scaling of the denoising is effectively a function of the projection angle. It is possible to combine any previously disclosed embodiment.

[0124] For instance, each projection data portion (described with reference to Figure 5) may be defined as comprising one or more sets of projection data elements.

[0125] In such examples, step 330 may comprise, for each set of projection data elements, processing the projection data element, the corresponding projection angle and the corresponding anatomical data portion (i.e., of the filtered projection data portion to which the set of projection data belongs) using at least one machine-learning algorithm.

[0126] In other such examples, step 330 may comprise, for projection data portion, processing the projection data portion and the corresponding anatomical data portion using at least one machine learning algorithm (e.g., performing steps 521 and 522 previously disclosed). Step 330 may also comprise, for each projection data portion, scaling the denoising of each set of projection data elements in the projection data portion as a function of projection angle.

[0127] Figure 7 is a flowchart that illustrates a method 700 comprising further optional steps that may be performed by some herein proposed methods.

[0128] For instance, one proposed method further comprises obtaining 710 initial projection data of the subject, wherein the initial projection data is generated using the medical imaging system; and performing 720 one or more pre-processing and / or rebinning functions on the initial projection data to produce the projection data. The rebinning function may, for instance, be a fan beam to parallel beam rebinning function, and is particularly useful if the initial projection data is captured using a cone-beam technique.

[0129] Example pre-processing and / or rebinning functions used for CT imaging are well known in the art, such as those put forward by Grass, M., Th Kohler, and R. Proksa. "3D cone-beam CT reconstruction for circular trajectories." Physics in Medicine & Biology 45.2 (2000): 329; Schaller, Stefan & Noo, Frederic & Sauer, Frank & Tam, K.C. & Lauritsch, G & Flohr, Thomas. (2000). Exact Radon rebinning algorithm for the long object problem in helical cone-beam CT. IEEE transactions on medical imaging. 19. 361-75; and / or KachelrieB, Marc, Stefan Schaller, and Willi A. Kalender. "Advanced single-slice rebinning in conebeam spiral CT." Medical Physics 27.4 (2000): 754-772.

[0130] The projection data produced by step 720 can be subsequently obtained and processed by a previously described method 300, 500, 600, to produce image data.

[0131] The method 700 may comprise a step 730 of processing the image data using an image denoising algorithm to produce denoised image data. Image denoising algorithms for use in processing image data derived from CT projection data are well known in the art. Examples include those set out by Thanh, Dang, and Prasath Surya. "A review on CT and X- ray images denoising methods." Informatica 43.2 (2019); and / or Yang, Qingsong, et al. "Low-dose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss." IEEE transactions on medical imaging 37.6 (2018): 1348-1357.

[0132] Approaches for performing denoising of image data derived from MRI raw scan data are known in the art, e.g., as discussed by Mohan, J., V. Krishnaveni, and Yanhui Guo. "A survey on the magnetic resonance image denoising methods." Biomedical signal processing and control 9 (2014): 56-69.

[0133] The method 700 may further comprise a step 740 of processing the denoised image data using image-domain filtering to produce output image data, wherein the image-domain filtering is configured to performing filtering on the denoised image data in accordance with the filter characteristics (e.g. MTF, NPS) of predefined scan protocols.

[0134] This approach recognizes that there may be a preference or desire for image data to have a particular appearance and / or one or more characteristic(s). Using such image-domain filtering thereby allows the image data to be adapted for any such preferences, e.g., to achieve a desired contrast distribution and so on.

[0135] The method 700 may further comprise a step 750 of outputting the image data and / or any further data derived from the image data, such as the denoised and / or output image data previously described. Step 750 may comprise outputting the image data to another device, e.g., separate to the device that carried out and / or performed the method 700.

[0136] For instance, step 750 may comprise controlling a user interface to provide a visual representation of the image data or, if produced, the output image data. Step 750 may thereby effectively comprise displaying the image data or, if produced, the output image data.

[0137] In some examples, step 750 comprise storing the image data or, if produced, the output image data. The image data may be stored in a memory and / or storage unit.

[0138] In some examples, step 750 comprises passing the image data and / or any further data derived therefrom to another device for further processing and / or analysis.

[0139] The skilled person would be readily capable of developing processing circuitry for carrying out any herein described method, e.g., when executing instructions contained or carried out by a memory. Thus, each step of the flow chart may represent a different action performed by processing circuitry, and may be performed by a respective module of the processing circuitry. Embodiments may therefore make use of processing circuitry. Processing circuitry can be implemented in numerous ways, with software and / or hardware, to perform the various functions required.

[0140] A processor is one example of processing circuitry which employs one or more microprocessors that may be programmed using software (e.g., microcode) to perform the required functions. Processing circuitry may however be implemented with or without employing a processor, and also may be implemented as a combination of dedicated hardware to perform some functions and a processor (e.g., one or more programmed microprocessors and associated circuitry) to perform other functions.

[0141] Examples of processing system components that may be employed in various embodiments of the present disclosure include, but are not limited to, conventional microprocessors, application specific integrated circuits (ASICs), and field-programmable gate arrays (FPGAs).

[0142] In various implementations, the processing circuitry may be associated with memory (i.e., one or more storage media) such as volatile and non-volatile computer memory such as RAM, PROM, EPROM, and EEPROM. The memory may be encoded with instructions (i.e., one or more programs) that, when executed by the processing circuitry, perform the required functions. Various storage media or medium may be fixed within a device comprising the processing circuitry or may be transportable, such that the one or more programs stored thereon can be loaded into processing circuitry.

[0143] Accordingly, there is provided a device for processing projection data for a medical imaging system. The device comprises processing circuitry; and a memory containing instructions that, when executed by the processing circuitry, configure the processing circuitry to perform any herein proposed method.

[0144] There is further proposed a system comprising the device and the medical imaging system configured to generate the projection data and / or, if relevant, the initial projection data.

[0145] There is also proposed a user interface system comprising the device and a user interface. The device may be configured to control the user interface to provide a visual representation of the image data and / or any further data / information derived from the image data (such as the denoised image data and / or output image data previously described).

[0146] The system described above, comprising the device and the medical imaging system, may comprise the user interface system, i.e., further comprise the user interface. It will be understood that disclosed methods are preferably computer-implemented methods. As such, there is also proposed the concept of a computer program comprising code means for implementing any described method when said program is run on processing circuitry, such as a computer. Thus, different portions, lines or blocks of code of a computer program according to an embodiment may be executed by processing circuitry or computer to perform any herein described method.

[0147] There is also proposed a non-transitory storage medium or memory that stores or carries instructions (e.g., a computer program or computer code) that, when executed by processing circuitry, causes the processing circuitry to carry out any herein described method.

[0148] In some alternative implementations, the functions noted in the block diagram(s) or flow chart(s) may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved.

[0149] 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. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

[0150] 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. If the term "adapted to" is used in the claims or description, it is noted the term "adapted to" is intended to be equivalent to the term "configured to". If the term "arrangement" is used in the claims or description, it is noted the term "arrangement" is intended to be equivalent to the term "system", and vice versa.

[0151] A single processor or other unit may fulfill the functions of several items recited in the claims. If a computer program is discussed above, it may be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied 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.

[0152] Any reference signs in the claims should not be construed as limiting the scope.

Claims

CLAIMS:

1. A computer-implemented method for processing projection data for a medical imaging system, the computer-implemented method comprising: receiving projection data of a subject, wherein the projection data is generated using the medical imaging system; processing the projection data using at least one filter to produce filtered projection data; processing the filtered projection data, using at least one machine-learning algorithm trained to perform a denoising procedure on filtered projection data, to produce denoised projection data; and processing the denoised projection data using a reconstruction algorithm to produce image data.

2. The computer-implemented method of claim 1, wherein the at least one filter comprises a high-pass filter.

3. The computer-implemented method of claim 1, wherein the at least one filter comprises a ramp filter.

4. The computer-implemented method of claim 1, wherein processing the filtered projection data comprises: processing the projection data or the filtered projection data to identify anatomical data that indicates one or more anatomical elements and / or anatomical regions represented by the filtered projection data; and processing the anatomical data and the filtered projection data using the at least one machine-learning algorithm to produce the denoised projection data.

5. The computer-implemented method of claim 4, wherein processing the anatomical data and the filtered projection data comprises: selecting one or more of a plurality of candidate machine-learning algorithms using the anatomical data, wherein different candidate machine-learning algorithms are associated with different anatomical elements and / or anatomical regions; andprocessing the filtered projection data using the selected one or more candidate machine-learning algorithms.

6. The computer-implemented method of claim 4, wherein processing the anatomical data and the filtered projection data comprises inputting the anatomical data and the filtered projection data into the one or more machine-learning algorithms to produce the denoised projection data.

7. The computer-implemented method of claim 4, wherein: the filtered projection data comprises a plurality of different filtered projection data portions; the anatomical data comprises, for each filtered projection data portion, a respective anatomical data portion that indicates one or more anatomical elements and / or anatomical regions represented by the filtered projection data portion; and processing the anatomical data and the filtered projection data comprises, for each filtered projection data portion, processing the filtered projection data portion and the respective anatomical data portion using a subset of the at least one machine-learning algorithm to produce the denoised projection data.

8. The computer-implemented method of claim 1, wherein: the filtered projection data comprises plurality of different filtered sets of projection data elements, each filtered set being associated with a different projection angle at which the projection data elements were captured by the medical imaging system; and processing the filtered projection data comprises, for each filtered set of projection data elements, processing the filtered set and the associated projection angle using at least one machine-learning algorithm to produce denoised projection data.

9. The computer-implemented method of claim 1, further comprising: obtaining initial projection data of the subject, wherein the initial projection data is generated using the medical imaging system; and performing one or more pre-processing and / or rebinning functions on the initial projection data to produce the projection data.

10. The computer-implemented method of claim 1, further comprising processing the image data using an image denoising algorithm to produce denoised image data.

11. The computer-implemented method of claim 10, further comprising processing the denoised image data using image-domain filtering to produce output image data, wherein the image-domain filtering is configured to performing filtering on the denoised image data in accordance with a predefined imaging protocol.

12. The computer-implemented method of claim 1, wherein the at least one filter is designed for use in a filtered back-projection technique usable to reconstruct image data from projection data.

13. A computer program product comprising computer program code means which, when executed on a device having processing circuitry, cause the processing circuitry to perform all of the steps of the method according to claim 1.

14. A device for processing projection data for a medical imaging system, the device comprising: processing circuitry; and a memory containing instructions that, when executed by the processing circuitry, configure the processing circuitry to: obtain projection data of a subject, wherein the projection data is generated using the medical imaging system; process the projection data using at least one filter to produce filtered projection data; process the filtered projection data, using at least one machine-learning algorithm trained to perform a denoising procedure on filtered projection data, to produce denoised projection data; and process the denoised projection data using a reconstruction algorithm to produce image data.

15. A system, comprising: the device of claim 14; and the medical imaging system configured to generate the projection data of the subject.

Images