Generating training data for image denoising using photon number division

Abstract

To provide systems / techniques that facilitate generation of training data for image denoising via photon-count splitting.SOLUTION: A system can access a set of sinograms generated by a photon-counting computed tomography scanner. The system can split the set of sinograms into a first reduced-photon-count set of sinograms and a second reduced-photon-count set of sinograms. The system can convert, via image reconstruction, the first reduced-photon-count set of sinograms into at least one training input image, and the second reduced-photon-count set of sinograms into at least one training output image. The system can train a deep learning neural network based on the at least one training input image and the at least one training output image.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] FIELD OF THE DISCLOSURE This disclosure relates generally to image denoising, and more particularly to generating training data for denoising images by photon number division. [Background technology]

[0002] Deep learning neural networks can be trained in a supervised manner to perform denoising on input computed tomography images. Unfortunately, annotated training datasets for training deep learning neural networks are difficult to obtain. Summary of the Invention

[0003] The following presents a summary to provide a basic understanding of one or more embodiments of the invention. This summary is not intended to identify key or critical elements or to delineate the scope of particular embodiments or claims. Its sole purpose is to present concepts in a simplified form as a prelude to the more detailed description that is presented later. In one or more embodiments described herein, an apparatus, system, computer-implemented method, device, or computer program product is described that assists in generating training data for image denoising by photon number splitting.

[0004] According to one or more embodiments, a system is provided. The system can include a non-transitory computer-readable memory capable of storing computer-executable components. The system can further include a processor operably coupled to the non-transitory computer-readable memory, the processor capable of executing the computer-executable components stored in the non-transitory computer-readable memory. In various embodiments, the computer-executable components can include an access component that accesses a set of sinograms generated by a photon-counting computed tomography scanner. In various embodiments, the computer-executable components can include a segmentation component that segments the set of sinograms into a first set of reduced-photon-count sinograms and a second set of reduced-photon-count sinograms. In various embodiments, the computer-executable components can include a reconstruction component that converts the first set of reduced-photon-count sinograms into at least one training input image and the second set of reduced-photon-count sinograms into at least one training output image by image reconstruction. In various cases, the computer-executable components may include a training component that trains a deep learning neural network based on the at least one training input image and the at least one training output image.

[0005] According to one or more embodiments, the above-described system may be implemented as a computer-implemented method or computer program product. [Brief explanation of the drawings]

[0006] [Figure 1] FIG. 1 shows a block diagram of an exemplary, non-limiting system for assisting in generating training data for denoising images by photon number splitting, according to one or more embodiments described herein. [Figure 2]FIG. 1 shows a block diagram of an exemplary, non-limiting system including various sets of sinograms with reduced photon counts, according to one or more embodiments described herein, to assist in generating training data for denoising images by photon number partitioning. [Figure 3] FIG. 1 is an exemplary, non-limiting block diagram illustrating how various sets of reduced-photon-count sinograms may be generated in accordance with one or more embodiments described herein. [Figure 4] FIG. 1 is an exemplary, non-limiting block diagram illustrating how various sets of reduced-photon-count sinograms may be generated in accordance with one or more embodiments described herein. [Figure 5] FIG. 1 is an exemplary, non-limiting block diagram illustrating how various sets of reduced-photon-count sinograms may be generated in accordance with one or more embodiments described herein. [Figure 6] FIG. 1 is an exemplary, non-limiting block diagram illustrating how various sets of reduced-photon-count sinograms may be generated in accordance with one or more embodiments described herein. [Figure 7] FIG. 1 is an exemplary, non-limiting block diagram illustrating how various sets of reduced-photon-count sinograms may be generated in accordance with one or more embodiments described herein. [Figure 8] 1 illustrates a flow diagram of an exemplary, non-limiting computer-implemented method that facilitates photon number segmentation of raw sinograms, in accordance with one or more embodiments described herein. [Figure 9] FIG. 1 is a block diagram of an exemplary, non-limiting system including at least one training input image and at least one training output image that assists in generating training data for removing noise from images by photon number splitting, according to one or more embodiments described herein. [Figure 10]FIG. 1 is an exemplary, non-limiting block diagram illustrating how training input images and training output images can be generated based on various sets of sinograms with reduced photon counts, according to one or more embodiments described herein. [Figure 11] FIG. 1 is an exemplary, non-limiting block diagram illustrating how a deep learning neural network can be trained in accordance with one or more embodiments described herein. [Figure 12] Illustrative, non-limiting experimental results are presented that demonstrate various advantages of one or more embodiments described herein. [Figure 13] Illustrative, non-limiting experimental results are presented that demonstrate various advantages of one or more embodiments described herein. [Figure 14] 1 illustrates a flow diagram of an exemplary, non-limiting computer-implemented method for assisting in generating training data for denoising images by photon number splitting, according to one or more embodiments described herein. [Figure 15] FIG. 1 illustrates a block diagram of an exemplary non-limiting operating environment for implementing one or more embodiments described herein. [Figure 16] 1 illustrates an exemplary networking environment operable to execute various implementations described herein. DETAILED DESCRIPTION OF THE INVENTION

[0007] The following detailed description is merely illustrative and is not intended to limit the embodiments or the applications / uses of the embodiments, nor is it limited by any express or implied information presented in the Background or Summary sections above, or in the Detailed Description section.

[0008] One or more embodiments will now be described with reference to the drawings, wherein like reference numerals are used to represent like elements throughout. In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a further understanding of one or more embodiments. It will be apparent, however, that in various instances one or more embodiments may be practiced without these specific details.

[0009] Deep learning neural networks can be trained in a supervised manner to perform denoising on input computed tomography (CT) images. Unfortunately, obtaining or curating annotated training datasets for training deep learning neural networks can be challenging for a variety of reasons.

[0010] With respect to image denoising of CT images, the annotated training dataset can include a set of pairs of training CT images, each pair including a training input CT image exhibiting a certain amount of visual noise and a corresponding training output CT image exhibiting a lesser amount of visual noise than the training input CT image. In various cases, the training input CT image represents the appropriate anatomical structure of a patient, and the training output CT image represents the same anatomical structure of the same patient but with less visual noise. Thus, the training output CT image can be considered a denoised or noise-reduced ground truth (e.g., target) from the training input CT image. There are various techniques for obtaining the training input CT images and the training output CT images.

[0011] To theoretically achieve optimal denoising performance, the training output CT image would be a completely noise-free (e.g., clean) version of the training input CT image. However, achieving such completely noise-free CT images in clinical practice is unrealistic. After all, the amount of visual noise present in a scanned CT image is inversely proportional to the radiation dose of that scanned CT image. Therefore, obtaining completely noise-free CT images in clinical practice would require an infinite radiation dose, which is not feasible. In other words, scanned CT images generated with less than an infinite radiation dose in clinical practice are likely to contain at least some visual noise. While noise-free CT images can be approximated in rigorous laboratory / experimental environments, these methods rely on complex or computationally intensive techniques that are unfeasible or impractical for large-scale implementation in clinical practice.

[0012] When some existing techniques are implemented, training input CT images are obtained by performing low-dose CT scans on patients, and training output CT images are obtained by performing high-dose CT scans on patients. Because the training output CT images can be generated with a higher radiation dose than the training input CT images, the training output CT images have less visual noise than the training input CT images. Therefore, the training input CT images and the training output CT images can be used for training to remove image noise. However, such existing techniques have drawbacks. In particular, such existing techniques require repeated scans. In other words, such existing techniques require two separate scans to be performed on the patient (e.g., one scan for the training input CT images and another scan for the training output CT images). Therefore, such existing techniques may expose the patient to excessive amounts of radiation, which may increase the likelihood of the patient suffering from adverse health effects. Furthermore, such existing techniques often suffer from registration issues. That is, because such existing techniques require two separate scans to be performed on the patient, the input and output training CT images may be misaligned with respect to each other (e.g., due to a slightly different patient body position or orientation during the second CT scan compared to the first CT scan), which may adversely affect subsequent denoising training.

[0013] In other existing techniques, training output CT images are obtained by performing a high-dose CT scan on a patient, and training input CT images are obtained by copying the training output CT images and injecting noise into the copies. In such cases, the training output CT images have less visual noise than the training input CT images, so the training input CT images and the training output CT images can be used for training to remove image noise. This technique does not require repeated scans, thereby avoiding problems of excessive radiation exposure and misalignment. However, like the above techniques, such existing techniques have drawbacks. Specifically, because the training output CT images are generated with a radiation exposure below infinity, the training output CT images contain a non-zero amount of noise. Furthermore, because the training input CT images are based on copies of the training output CT images, the training input CT images may also contain non-zero noise, just like the training output CT images. In other words, this non-zero noise can be considered to indicate a correlation between the training input CT images and the training output CT images. In other words, the noise present in the training input CT images can be considered not independent of the noise present in the training output CT images. Therefore, if the training input CT images are generated by injecting additional noise, a deep learning neural network trained using the training input CT images can only learn how to remove the injected additional noise; the deep learning neural network cannot learn how to remove non-zero amounts of noise that are present (e.g., correlated) in both the training input CT images and the training output CT images.

[0014] Therefore, a system or technique that can address one or more of these technical problems is desirable.

[0015] Various embodiments described herein may address one or more of the above technical problems. One or more embodiments described herein may include a system, computer-implemented method, apparatus, or computer program product that can assist in generating training data for image denoising using photon number splitting. In other words, the inventors of various embodiments described herein have realized that various drawbacks of existing techniques for generating training data for image denoising can be improved by implementing photon number splitting. In particular, a photon-counting detector computed tomography (PCD-CT) scanner (as opposed to an energy-integrating detector computed tomography (EID-CT) scanner) performs a scan on a patient, thereby obtaining a set of sinograms. In various aspects, as described herein, the set of sinograms can be split into a first set of photon-reduced sinograms and a second set of photon-reduced sinograms using a photon-wise binary selection. In various embodiments, as further described herein, the first set of photon-reduced sinograms can be converted into training input CT images using any suitable image reconstruction technique. Similarly, as described further herein, the second set of photon-count-reduced sinograms can be converted into training output CT images by any suitable image reconstruction technique. Note that such training image pairs can be generated by scanning the patient once, thus avoiding repeated scans and their attendant problems (e.g., misalignment, excessive radiation exposure). Furthermore, by performing photon-wise binary selection photon count splitting as described herein, visual noise present in the training input CT images can be made independent (e.g., uncorrelated) from visual noise present in the training output CT images. Thus, problems with correlated noise can be avoided, and training to remove noise can be improved.Indeed, the inventors have experimentally verified that the various embodiments described herein can achieve denoising performance comparable to techniques that utilize completely clean (e.g., noise-free) training output images.

[0016] Various embodiments described herein can be considered as computer tools (e.g., any suitable combination of computer-executable hardware or computer-executable software) that can assist in generating training data for denoising images by photon number splitting. In various aspects, such computer tools can include an access component, a splitting component, a reconstruction component, a training component, or a deployment component.

[0017] In various embodiments, a PCD-CT scanner (rather than an EID-CT scanner) performs a scan on a patient, thereby obtaining a set of sinograms. In various aspects, the set of sinograms can include any suitable number of sinograms. In various implementations, a sinogram is a two-dimensional pixel array representing projection data acquired by the PCD-CT scanner. For example, one dimension of the sinogram can represent the gantry angle of the PCD-CT scanner, and the other dimension of the sinogram can represent distance along the projection direction of the PCD-CT scanner (e.g., distance perpendicular to the direction of x-ray emission). In various cases, because the set of sinograms can be generated by a PCD-CT scanner (rather than an EID-CT scanner), the pixels of the set of sinograms can represent photon counts organized by photon energy bins. In other words, a given sinogram in the set of sinograms corresponds to a given photon energy range, and a given pixel in that given sinogram can be thought of as corresponding to a given angle-distance tuple swept during the patient scan, and the value of that given pixel can be thought of as representing how many individual photons within the given photon energy range were detected by the PCD-CT scanner at that given angle-distance tuple.

[0018] In various embodiments, an access component of the computer tool can receive or access the set of sinograms electronically. In some aspects, the access component can electronically retrieve the set of sinograms from any suitable centralized or distributed data structure (e.g., a graph data structure, a relational data structure, a hybrid data structure), whether remote from the access component or local to the access component. In other aspects, the access component can electronically retrieve the set of sinograms from the PCD-CT scanner that generated the set of sinograms. In either case, the access component can obtain or access the set of sinograms electronically, and other components of the computer tool can electronically interact with (e.g., read, write, edit, copy, manipulate) the set of sinograms.

[0019] In various embodiments, a segmentation component of the computer tool can electronically segment the set of sinograms into a first set of photon-reduced sinograms and a second set of photon-reduced sinograms by photon-wise binary selection. More specifically, the first set of photon-reduced sinograms and the second set of photon-reduced sinograms are initially empty sets. In various embodiments, the segmentation component iterates through each sinogram in the sinogram set. For a given sinogram from the sinogram set, the segmentation component can insert an empty sinogram having the same dimensions as the given sinogram (e.g., the empty sinogram can have the same number of pixels or the same pixel arrangement as the given sinogram, but all pixel values ​​of the empty sinogram are initially zero) into the first set of reduced-photon sinograms. For ease of description, such an empty sinogram will be referred to as sinogram A. Similarly, the segmentation component can insert an empty sinogram having the same dimensions as the given sinogram (e.g., the empty sinogram has the same number of pixels or pixel arrangement as the given sinogram, but all pixel values ​​of the empty sinogram are initially zero) into the second set of sinograms with reduced photon counts. For convenience of explanation, such an empty sinogram will be referred to as sinogram B.

[0020] As described above, each pixel of a given sinogram can represent a photon number. In other words, each pixel of a given sinogram can be considered to contain or represent a respective number of recorded photons. Thus, in various aspects, the segmentation component can iterate through each pixel of a given sinogram, and for a given pixel of a given sinogram, the segmentation component can iterate through the recorded photons represented by the given pixel. For a given recorded photon, the segmentation component can assign the given recorded photon to a respective pixel of sinogram A, or alternatively to a respective pixel of sinogram B, by binary selection. In particular, the segmentation component can probabilistically assign the given recorded photon to a respective pixel of sinogram A according to an appropriate threshold probability, and the segmentation component can alternatively probabilistically assign the given recorded photon to a respective pixel of sinogram B according to the complement of that threshold probability. Note that in various cases, applying binary choices to single photons in this manner can be thought of or called a photon-by-photon Bernoulli trial.

[0021] In various aspects, the splitting component can repeat this probabilistic assignment for each recorded photon of a given pixel, such that each pixel of sinogram A and each pixel of sinogram B are filled with recorded photons. Note that the photon number of each pixel of sinogram A and the photon number of each pixel of sinogram B sum to the photon number of the given pixel (e.g., a given pixel can include an appropriate number of recorded photons, and these recorded photons are probabilistically split or otherwise divided between each pixel of sinogram A and each pixel of sinogram B according to a threshold probability). In other words, each pixel of sinogram A and each pixel of sinogram B can be considered complementary to each other. In yet another way, each pixel of sinogram A and each pixel of sinogram B both have a photon number that is smaller than the photon number of the given pixel, hence the term "reduced photon number."

[0022] Note that while the above paragraph describes the segmentation component as performing a series of Bernoulli trials (e.g., performing a binary selection for each photon (one photon at a time)), this is merely a non-limiting example for ease of explanation. Alternatively, in various aspects, the segmentation component can perform photon-by-photon binary selection according to an appropriate batch size (e.g., for multiple photons at a time). Indeed, independent Bernoulli trials can be thought of as summing to form a binomial distribution. Thus, rather than probabilistically assigning individual photons of a given pixel to respective pixels of sinogram A by Bernoulli trials (e.g., one photon at a time), or alternatively, to respective pixels of sinogram B, the segmentation component can probabilistically assign photons recorded at a given pixel to respective pixels of sinogram A by sampling the binomial distribution using binary selection. In such a case, recorded photons of a given pixel that are not assigned to respective pixels of sinogram A by such sampling can be assigned to respective pixels of sinogram B.

[0023] In various aspects, the segmentation component can repeat the above operations for each pixel of a given sinogram and for each sinogram of the sinogram set. In this way, pixels of the first set of reduced-photon-count sinograms are filled with recorded photons from the set of sinograms at a rate that approximates a threshold probability, and pixels of the second set of reduced-photon-count sinograms are filled with recorded photons from the set of sinograms at a rate that approximates the complement of the threshold probability. In other words, the first set of reduced-photon-count sinograms and the second set of reduced-photon-count sinograms can be thought of as visually representing the same (or nearly the same) pattern or arrangement of projection data as the set of sinograms, and the first and second sets can be thought of as visually representing the same (or nearly the same) pattern or arrangement of projection data as each other, but at different dose levels.

[0024] In various embodiments, a reconstruction component of the computer tool can electronically generate a pair of images based on a first set of photon-reduced sinograms and a second set of photon-reduced sinograms. More specifically, the reconstruction component can apply an appropriate image reconstruction technique (e.g., material decomposition, filtered back projection) to the first set of photon-reduced sinograms, thereby obtaining a first CT image. In various aspects, the first CT image can have an appropriate format, size, or dimensionality (e.g., a two-dimensional array of pixels, a three-dimensional array of voxels). In some cases, the first CT image is a plurality of first reference material images (e.g., a plurality of first CT images, each representing a patient's anatomy according to a respective reference material (e.g., water or calcium)). Similarly, the reconstruction component can apply an appropriate image reconstruction technique to the second set of photon-reduced sinograms, thereby obtaining a second CT image. As above, the second CT image can have an appropriate format, size, or dimensionality. Further, similar to above, the second CT image can be a plurality of second reference images in various embodiments.

[0025] In various aspects, the segmentation component can assign the recorded photons to a first set of sinograms with a reduced number of photons according to a threshold probability, and the first CT image can be generated based on the first set of sinograms with a reduced number of photons, such that the threshold probability corresponds to or functions as a measure of the radiation dose of the first CT image. Similarly, the segmentation component can assign the recorded photons to a second set of sinograms with a reduced number of photons according to the complement of the threshold probability, and the second CT image can be generated based on the second set of sinograms with a reduced number of photons, such that the complement of the threshold probability corresponds to or functions as a measure of the radiation dose of the second CT image. Thus, if the threshold probability is less than 50%, both the first CT image and the second CT image can be considered to represent the patient's anatomy (e.g., both images represent the same anatomy of the same patient), but the first CT image can represent the patient's anatomy with a lower radiation dose than the second CT image. In such a case, the first CT image can be considered a noisier second CT image. Conversely, if the threshold probability is greater than 50%, both the first CT image and the second CT image can be considered to represent the patient's anatomy, but the first CT image represents the patient's anatomy with a higher radiation dose than the second CT image. In such a case, the first CT image can be considered a less noisy second CT image. In any case, by applying photon-wise binary selection photon number splitting, any visual noise present in the first CT image can be made independent of (e.g., uncorrelated with) the visual noise present in the second CT image.

[0026] In various embodiments, the computational training component can electronically train the deep learning neural network with the first CT image and the second CT image in a supervised manner.

[0027] In various embodiments, the deep learning neural network may have a suitable internal architecture. For example, the deep learning neural network may include a suitable number of layers of a suitable type (e.g., an input layer, one or more hidden layers, and an output layer, any of which may be a convolutional layer, a dense layer, a nonlinear layer, a pooling layer, a batch normalization layer, or a padding layer). As another example, the deep learning neural network may include a suitable number of neurons in various layers (e.g., different layers may have the same or different numbers of neurons). As yet another example, the deep learning neural network may include a suitable activation function (e.g., softmax, sigmoid, hyperbolic tangent, rectified linear unit) for various neurons (e.g., different neurons may have the same or different activation functions). As yet another example, the deep learning neural network may include a suitable inter-neuron or inter-layer connection (e.g., forward connection, skip connection, recurrent connection).

[0028] Regardless of the internal architecture of the deep learning neural network, the training component can train the deep learning neural network to perform image denoising using a first CT image and a second CT image. In fact, even if both the first CT image and the second CT image are noisy, the deep learning neural network can be trained to perform image denoising because the noise present in the first CT image is independent of the noise present in the second CT image. In any case, before such training begins, the internal parameters of the deep learning neural network (e.g., weights, biases, convolution kernels) can be randomly initialized.

[0029] In various embodiments, the training component can provide the first CT image to a deep learning neural network, causing the deep learning neural network to generate an output. For example, the first CT image can be received by an input layer of the deep learning neural network, the first CT image can be forward passed through one or more hidden layers of the deep learning neural network, and the output layer of the deep learning neural network can calculate an output based on the activations generated by the one or more hidden layers.

[0030] In various embodiments, the output can be considered a prediction or inference of what the deep learning neural network believes should correspond to the first CT image. In contrast, the second CT image can be considered a correct, accurate, ground-truth result that is known or believed to correspond to the first CT image. Note that if the deep learning neural network has not previously undergone any or very little training, the output can be highly inaccurate (e.g., the output may differ significantly from the second CT image).

[0031] In either case, the training component can calculate one or more errors or losses (e.g., mean absolute error (MAE), mean squared error (MSE), cross entropy) between the output and the second CT image. In various aspects, the training component can update the internal parameters of the deep learning neural network by performing backpropagation (e.g., stochastic gradient descent) driven by the calculated errors or losses.

[0032] This is a non-limiting example of how the training component can train a deep learning neural network with a first CT image and a second CT image. In other words, the first CT image can be considered a training input, and the second CT image can be considered a ground truth annotation or target corresponding to the training input. In yet another way, the first CT image and the second CT image can be considered an input-annotation pair. In various aspects, the computer tools described herein can generate a respective input-annotation pair for each set of sinograms generated by the PCD-CT scanner described above (or any other PCD-CT scanner). Thus, if multiple sets of sinograms generated by the PCD-CT scanner are available, the access component, segmentation component, and reconstruction component can operate to generate multiple input-annotation pairs, as described above. Such multiple input-annotation pairs can be considered a complete training dataset for the training component to train the deep learning neural network. In such cases, the training component can repeat the above-described run-update procedure for each input-annotation pair in the training dataset, thereby iteratively optimizing the internal parameters of the deep learning neural network to accurately denoise the input CT images. In various aspects, the training component can implement an appropriate batch size for training, an appropriate termination criterion for training, or an appropriate error function, loss function, or objective function.

[0033] After the deep learning neural network is trained, a deployment component of the computer tool can electronically deploy the deep learning neural network in an appropriate operational context, i.e., the deployment component can run the deep learning neural network on a given CT image (e.g., a CT image for which noise removal is desired) acquired in the field.

[0034] It should be noted that both the first CT image and the second CT image can be generated from a set of sinograms acquired by a PCD-CT scanner, where the set of sinograms is the result of a single scan performed by the PCD-CT scanner. In other words, various embodiments described herein can avoid repeated scans and avoid problems associated with repeated scans (e.g., misalignment, excessive radiation exposure). Furthermore, because the first CT image and the second CT image can be generated by photon-by-photon binary selection, noise present in the first CT image is independent of (e.g., uncorrelated with) noise present in the second CT image. In other words, various embodiments described herein can avoid problems with correlated noise. Furthermore, as described herein, the inventors have experimentally confirmed that by training a deep learning neural network to perform image denoising on training data generated by photon number splitting (e.g., by photon-wise binary choice), the deep learning neural network exhibits denoising accuracy comparable to training the deep learning neural network on completely clean (e.g., noise-free) ground truth.

[0035] Various embodiments described herein can be employed to solve practically technically sophisticated problems (e.g., assisting in the generation of training data for photon number division image denoising) using hardware or software, problems that are not abstract and cannot be performed as a series of mental activities by a human. Furthermore, some of the processing performed can be performed by a computer (e.g., a deep learning neural network with internal parameters such as convolution kernels) dedicated to performing the defined tasks associated with generating training data for photon number division image denoising. For example, such defined tasks may include accessing, by a device operatively coupled to a processor, a set of sinograms generated by a photon-counting computed tomography scanner; dividing, by the device, the set of sinograms into a first set of reduced-photon-count sinograms and a second set of reduced-photon-count sinograms; converting, by the apparatus and image reconstruction, the first set of reduced-photon-count sinograms into at least one training input image and the second set of reduced-photon-count sinograms into at least one training output image; and training, by the device, a deep learning neural network based on the at least one training input image and the at least one training output image.

[0036] This defined task is not something that can be performed manually by a human. Indeed, neither the human mind nor a pen-and-paper human can electronically access sinograms acquired by a PCD-CT scanner, electronically split the sinograms into two sets of photon-reduced sinograms using photon-wise binary selection, electronically generate input-annotation pairs by applying image reconstruction (e.g., material decomposition, filtered backprojection) to the two sets of photon-reduced sinograms, and electronically train a deep learning neural network (e.g., by backpropagation) with the input-annotation pairs. In fact, a PCD-CT scanner is an inherently computerized device and cannot be implemented in any way by the human mind without a computer. Similarly, a deep learning neural network is an inherently computerized construct that cannot be implemented in any way by the human mind without a computer. Thus, the computational tools capable of training deep learning neural networks on sinograms generated by PCD-CT scanners are likewise inherently computerized and cannot be implemented in a sensible, practical, and rational way without a computer.

[0037] Furthermore, various embodiments described herein can integrate various teachings related to generating training data for image denoising using photon number division into practical applications. As explained above, some existing techniques train deep learning neural networks to perform image denoising on CT images by utilizing completely clean (e.g., completely noise-free) ground truth. However, such completely clean ground truth is not feasible or practical at the scale (e.g., large enough volume) of clinical settings. As mentioned above, other existing techniques train deep learning neural networks to perform image denoising on CT images using training inputs generated by low-dose CT scans and corresponding training targets generated by high-dose CT scans. However, such techniques involve repeated scans of the patient, exposing the patient to excessive radiation and introducing registration errors. As mentioned above, various other existing techniques train deep learning neural networks to perform image denoising on CT images using training targets generated by high-dose CT scans and corresponding training inputs generated by noise injection. This technique avoids repeated scanning, but introduces the problem of correlated noise.

[0038] Various embodiments described herein can address these technical issues. Specifically, various embodiments described herein include acquiring a set of sinograms from a PCD-CT scanner, splitting the set of sinograms into two sets of sinograms with reduced photon counts using photon-wise binary selection, and generating training inputs and corresponding training targets by applying image reconstruction (e.g., material decomposition) to the two sets of reduced photon count sinograms. Because the set of sinograms is obtained from a single scan of the patient, the problem of repeated scans can be avoided. Furthermore, because photon number splitting using photon-wise binary selection can be implemented, noise in the training inputs can be uncorrelated with (e.g., independent of) any noise in the training targets. Furthermore, experimental results demonstrate that training a deep learning neural network in this manner can achieve denoising performance comparable to training a deep learning neural network using completely clean (e.g., completely noise-free) ground truth. That is, the various embodiments described herein can improve various shortcomings of existing techniques. Therefore, the various embodiments described herein certainly realize clear and concrete technical improvements in the field of image denoising. Therefore, the various embodiments described herein are clearly suitable for useful and practical applications of computers.

[0039] Additionally, various embodiments described herein may control real-world tangible devices based on the disclosed teachings. For example, various embodiments described herein may electronically train or run real-world deep learning neural networks on real-world CT images generated from real-world sinograms acquired by a real-world PCD-CT scanner.

[0040] It should be understood that the illustrations and descriptions herein provide non-limiting examples of various embodiments and are not necessarily drawn to scale.

[0041] 1 shows a block diagram of an exemplary, non-limiting system 100 that can assist in generating training data for denoising images by photon number partitioning, according to one or more embodiments described herein. As shown, the denoising training system 102 can be electronically integrated with a deep learning neural network 104 or a photon-counting detector computed tomography scanner 106 (hereinafter, "PCD-CT scanner 106") via suitable wired or wireless electronic connections.

[0042] In various embodiments, the deep learning neural network 104 may have or exhibit a suitable internal architecture. For example, the deep learning neural network 104 may have an input layer, one or more hidden layers, and an output layer. In various implementations, such layers may be coupled by suitable inter-neuron or inter-layer connections (such as forward connections, skip connections, or recurrent connections). Furthermore, in various implementations, such layers may be suitable types of neural network layers having suitable learnable or trainable internal parameters. For example, any of the input layer, one or more hidden layers, and output layer may be a convolutional layer, and the learnable or trainable parameters of the convolutional layer may be a convolution kernel. As another example, any of the input layer, one or more hidden layers, and output layer may be a dense layer, and the learnable or trainable parameters of the dense layer may be a weight matrix or a bias value. As yet another example, any of the input layer, one or more hidden layers, and output layer may be a batch normalization layer, and the learnable or trainable parameter of the batch normalization layer may be a shift factor or a scale factor. Furthermore, in various cases, any of the layers may be a suitable type of neural network layer having suitable fixed or non-trainable internal parameters. For example, any of the input layer, one or more hidden layers, and output layer may be a nonlinearity layer, a padding layer, a pooling layer, or a concatenation layer.

[0043] In various embodiments, the PCD-CT scanner 106 may be a suitable computed tomography scanner that utilizes photon-counting detector technology as opposed to energy-integrating detector technology. Thus, when performing a CT scan on a patient, the PCD-CT scanner 106 may count the number of individual photons within an appropriate energy band that are detected at any given point in the CT scan (e.g., a given gantry angle and projection distance).

[0044] As shown, the PCD-CT scanner 106 can, in various aspects, electronically generate or acquire the set of sinograms 108. In various embodiments, the set of sinograms 108 can be considered as raw projection data collected by the PCD-CT scanner 106 during a single scan of a single patient. In other words, the PCD-CT scanner 106 can scan any suitable anatomical structure of a patient, and the set of sinograms 108 can be considered as raw projection data obtained from this scan. As some non-limiting examples, the anatomical structure can be any suitable tissue of a patient (e.g., bone tissue, lung tissue, muscle tissue), any suitable organ of a patient (e.g., heart, liver, lung, brain), any suitable bodily fluid of a patient (e.g., blood, amniotic fluid), any other suitable body part of a patient, or any suitable portion thereof. In some cases, the PCD-CT scanner 106 can generate or acquire the set of sinograms 108 using a maximum (e.g., 100%) radiation dose level. In other embodiments, the PCD-CT scanner 106 may generate or acquire a set of sinograms 108 using a radiation dose level that is less than the maximum (eg, less than 100%).

[0045] In various embodiments, it may be desirable to train a deep learning neural network 104 to perform image denoising. In various implementations, the denoising training system 102 can generate input-target training pairs of CT images based on a set of sinograms 108 as described herein, and the denoising training system 102 can train the deep learning neural network 104 with such input-target training pairs.

[0046] In various embodiments, the noise reduction training system 102 may comprise a processor 110 (e.g., a central processing unit, microprocessor) and a non-transitory computer-readable memory 112 operably, operatively, or communicatively connected or coupled to the processor 110. The non-transitory computer-readable memory 112 may store computer-executable instructions that, when executed by the processor 110, cause the processor 110 or other components (e.g., the access component 114, the segmentation component 116, the reconstruction component 118, the training component 120, or the deployment component 122) of the noise reduction training system 102 to perform one or more operations. In various embodiments, the non-transitory computer-readable memory 112 may store, and the processor 110 may execute, the computer-executable components (e.g., the access component 114, the segmentation component 116, the reconstruction component 118, the training component 120, or the deployment component 122).

[0047] In various embodiments, the denoising training system 102 can include an access component 114. In various aspects, the access component 114 can electronically receive or access the deep learning neural network 104 or the set of sinograms 108. In various aspects, the access component 114 can electronically retrieve the deep learning neural network 104 from any suitable centralized or distributed data structure (not shown) or from any suitable centralized or distributed computing device (not shown). Similarly, in various cases, the access component 114 can electronically retrieve the set of sinograms 108 from any suitable centralized or distributed data structure (not shown) or from any suitable centralized or distributed computing device (not shown). As a non-limiting example, the PCD-CT scanner 106 can transmit the set of sinograms 108 to the access component 114. In either case, the access component 114 electronically obtains or accesses the deep learning neural network 104 or the set of sinograms 108, and other components of the denoising training system 102 can electronically interact with the deep learning neural network 104 or the set of sinograms 108.

[0048] In various embodiments, the denoising training system 102 can include a splitting component 116. In various aspects, as described herein, the splitting component 116 can split the set of sinograms on a photon-by-photon basis into a first set of sinograms with a reduced number of photons and a second set of sinograms with a reduced number of photons.

[0049] In various embodiments, the denoising training system 102 can include a reconstruction component 118. In various implementations, the reconstruction component 118 can generate training input images based on a first set of sinograms with reduced photon counts and generate training output images based on a second set of sinograms with reduced photon counts, as described herein.

[0050] In various embodiments, the denoising training system 102 can include a training component 120. In various cases, the training component 120 can train a deep learning neural network in a supervised manner based on training input images and training output images, as described herein.

[0051] In various embodiments, the denoising training system 102 can include a deployment component 122. In various aspects, the deployment component 122 can deploy or run the deep learning neural network after training on suitable desired images, as described herein.

[0052] 2 shows a block diagram of an exemplary, non-limiting system 200 including various sets of reduced-photon-count sinograms, which can assist in generating training data for denoising images through photon number partitioning, according to one or more embodiments described herein. As shown, system 200 can optionally include the same components as system 100, and can further include a set of reduced-photon-count sinograms 202 and a set of reduced-photon-count sinograms 204.

[0053] In various embodiments, segmentation component 116 can electronically segment the set of sinograms 108 into a set of photon-count reduced sinograms 202 and a set of photon-count reduced sinograms 204. More specifically, segmentation component 116 can, in various aspects, generate the set of photon-count reduced sinograms 202 and the set of photon-count reduced sinograms 204 by applying a photon-wise binary selection to the set of sinograms 108. Various non-limiting aspects are illustrated in FIGS.

[0054] 3-7 are exemplary, non-limiting block diagrams 300, 400, 500, 600, and 700 illustrating how a set of reduced photon count sinograms 202 and a set of reduced photon count sinograms 204 may be generated in accordance with one or more embodiments described herein.

[0055] Referring first to Figure 3, sinogram set 108, in various aspects, can comprise n sinograms (where n is any suitable positive integer), i.e., sinograms 108(1) through 108(n). In various embodiments, the sinograms in sinogram set 108 can exhibit any suitable format, size, or dimensionality. As a non-limiting example, a sinogram in sinogram set 108 can be an a x b array of pixels (where a and b are any suitable positive integers). In various cases, different sinograms in sinogram set 108 can have the same format or dimensionality (e.g., each sinogram in sinogram set 108 can be an a x b array of pixels).

[0056] In various embodiments, the set of sinograms 108 can be generated by the PCD-CT scanner 106, so that the set of sinograms 108 can be organized by or according to photon energy bins. Indeed, the PCD-CT scanner 106 can implement or utilize photon-counting detector technology (rather than energy-integrating detector technology), so that the PCD-CT scanner 106 can count how many individual photons within the appropriate photon energy range impinge on the appropriate photon-counting detector at the appropriate gantry angle. In particular, if a given sinogram in the set of sinograms 108 is an array of a x b pixels corresponding to a given photon energy bin, then: One major dimension or axis of such a given sinogram can be considered to represent different gantry angles of the PCD-CT scanner 106. Another major dimension or axis of such a given sinogram can be considered to represent distance along the projection direction of the PCD-CT scanner 106. A given pixel of such a given sinogram can be considered to be positioned or located at a given tuple of gantry angle and projection distance. The value of a given pixel can be considered to be the number of individual photons that fall into a given photon energy bin detected when the PCD-CT scanner 106 is set to a given tuple of gantry angle and projection distance. Because the set of sinograms 108 can have a cardinality of n, the set of sinograms 108 can be considered to be organized by or according to n distinct or unique photon energy bins. For example, sinogram 108(1) can be considered to correspond to a first photon energy bin (e.g., a first range of photon energy values ​​measured in electron volts), while sinogram 108(n) can be considered to correspond to an nth photon energy bin (e.g., an nth range of photon energy values ​​measured in electron volts). In various cases, these n photon energy bins can be disjoint (e.g., non-overlapping).

[0057] In various aspects, as shown, each of the photon-count-reduced sinogram sets 202 can correspond (e.g., one-to-one) to a corresponding sinogram set 108. Thus, since sinogram set 108 can have n sinograms, photon-count-reduced sinogram set 202 can similarly have n sinograms, i.e., photon-count-reduced sinogram 202(1) through photon-count-reduced sinogram 202(n). In other words, photon-count-reduced sinogram set 202 can be considered to have the same sinogram cardinality (e.g., the same number of sinograms) as sinogram set 108. In various aspects, each sinogram in photon-count-reduced sinogram set 202 can have the same format or dimensionality as any sinogram in sinogram set 108. For example, if each sinogram in sinogram set 108 is an a x b array of pixels, then each sinogram in reduced-photon-count sinogram set 202 can also be an a x b array of pixels. Furthermore, since sinogram set 108 can be organized by or according to photon energy bins, reduced-photon-count sinogram set 202 can also be organized by or according to photon energy bins. For example, reduced-photon-count sinogram 202(1), like sinogram 108(1), can correspond to the first photon energy bin. Similarly, reduced-photon-count sinogram 202(n), like sinogram 108(n), can correspond to the nth photon energy bin.

[0058] Similarly, as shown, each of photon-count reduced sinogram sets 204 can correspond (e.g., one-to-one) to sinogram set 108. Thus, because sinogram set 108 can have n sinograms, photon-count reduced sinogram set 204 can similarly have n sinograms, i.e., photon-count reduced sinogram 204(1) through photon-count reduced sinogram 204(n). In other words, photon-count reduced sinogram set 204 can be considered to have the same sinogram cardinality (e.g., the same number of sinograms) as sinogram set 108. As noted above, each sinogram in photon-count reduced sinogram set 204 can have the same format or dimensionality as any sinogram in sinogram set 108. For example, if each sinogram in sinogram set 108 is an a x b array of pixels, then each sinogram in reduced-photon-count sinogram set 204 can also be an a x b array of pixels. Furthermore, since sinogram set 108 can be organized by or according to photon energy bins, reduced-photon-count sinogram set 204 can also be organized by or according to photon energy bins. For example, reduced-photon-count sinogram 204(1), like sinogram 108(1), can correspond to the first photon energy bin. Similarly, reduced-photon-count sinogram 204(n), like sinogram 108(n), can correspond to the nth photon energy bin.

[0059] In various aspects, each sinogram in photon-count reduced sinogram set 202 and each sinogram in photon-count reduced sinogram set 204 can be generated by applying photon-wise binary selection photon number partitioning to each sinogram in sinogram set 108. For example, partitioning component 116 can generate photon-count reduced sinogram 202(1) and photon-count reduced sinogram 204(1) by applying photon-wise binary selection to sinogram 108(1). Thus, sinogram 108(1), photon-count reduced sinogram 202(1), and photon-count reduced sinogram 204(1) can all be considered to correspond to one another. Similarly, as another example, segmentation component 116 can generate photon-reduced sinogram 202(n) and photon-reduced sinogram 204(n) by applying a photon-wise binary selection to sinogram 108(n). Thus, sinogram 108(n), photon-reduced sinogram 202(n), and photon-reduced sinogram 204(n) can all be considered to correspond to one another. Various non-limiting aspects of such photon-wise binary selection are described in more detail in FIGS. 4-7.

[0060] As shown, FIG. 4 illustrates an exemplary, non-limiting embodiment of a sinogram 108(1). As illustrated, sinogram 108(1) can have a set of pixels 402. In various aspects, set of pixels 402 can include p pixels (where p is any suitable positive integer), i.e., pixel 402(1) through pixel 402(p). In various aspects, p can be thought of as representing the sum of a given tuple of gantry angles and projection distances swept by PCD-CT scanner 106 when generating set of sinograms 108 (e.g., if sinogram 108(1) is an a×b array of pixels, then p=ab). Furthermore, because sinogram 108(1) can be produced by PCD-CT scanner 106 and sinogram 108(1) corresponds to a first photon energy bin, each pixel of sinogram 108(1) can be thought of as recording or representing how many individual photons in the first photon energy bin are detected by PCD-CT scanner 106 at each tuple of gantry angle and projection distance. For example, pixel 402(1) can correspond to a first tuple of gantry angle and projection distance of PCD-CT scanner 106, and pixel 402(1) can be thought of as representing how many individual photons in the first photon energy bin are detected by PCD-CT scanner 106 at the first tuple of gantry angle and projection distance. As another example, pixel 402(p) can correspond to the pth tuple of gantry angle and projection distance of the PCD-CT scanner 106, and therefore pixel 402(p) can be thought of as representing how many individual photons in the first photon energy bin were detected by the PCD-CT scanner 106 at the pth tuple of gantry angle and projection distance.

[0061] In various embodiments, pixel 402(1) is C 1_Total Let C 1_Total is any suitable positive integer. In this case, pixel 402(1) has an energy in the first photon energy bin, C 1_TotalPixel 402(1) can be thought of as having recorded individual or distinct photons. Accordingly, pixel 402(1) can be thought of as containing or representing a set of recorded photons 404(1), where the cardinality of set of recorded photons 404(1) is C 1_Total That is, the set of recorded photons 404(1) is equal to C 1_Total photons, i.e., recorded photon 404(1)(1) to recorded photon 404(1)(C 1_Total ), can be included.

[0062] Similarly, pixel 402(p) is C p_Total Let C p_Total is any suitable positive integer. In this case, pixel 402(p) has an energy in the first photon energy bin, C p_Total Thus, pixel 402(p) can be thought of as containing or representing a set of recorded photons 404(p), where the cardinality of the set of recorded photons 404(p) is C p_Total That is, the set of recorded photons 404(p) is equal to C p_Total photons, i.e., recorded photon 404(p)(1) to recorded photon 404(p)(C p_Total ), can be included.

[0063] In various aspects, as shown, photon-reduced sinogram 202(1) includes a set of pixels 406. As described above, each sinogram in sinogram set 108 and each sinogram in photon-reduced sinogram set 202 can have the same number of pixels or the same arrangement of pixels as each other. Thus, because pixel set 402 in sinogram 108(1) can have a cardinality of p, pixel set 406 in photon-reduced sinogram 202(1) can also have a cardinality of p. That is, pixel set 406 can include p pixels, i.e., pixel 406(1) through pixel 406(p). In various embodiments, as shown, each pixel in pixel set 406 is initially empty. In other words, each pixel in pixel set 406 begins with a pixel value of zero. For example, pixel 406(1) may start with a pixel value of zero (e.g., before applying photon-wise binary selection by splitting component 116), indicating that no recorded photons have yet been represented by pixel 406(1). As another example, pixel 406(p) may start with a pixel value of zero (e.g., before applying photon-wise binary selection by splitting component 116), indicating that no recorded photons have yet been represented by pixel 406(p). However, this is merely a non-limiting embodiment. In various other cases, each pixel of pixel set 406 may be initialized in any other suitable manner.

[0064] In any case, the sets of pixels 406 may each correspond (e.g., one-to-one) to the sets of pixels 402. For example, pixel 406(1) may correspond to pixel 402(1) (e.g., both pixel 402(1) and pixel 406(1) may correspond to a first tuple of the gantry angle and projection distance of the PCD-CT scanner 106). Similarly, pixel 406(p) may correspond to pixel 402(p) (e.g., both pixel 402(p) and pixel 406(p) may correspond to a pth tuple of the gantry angle and projection distance of the PCD-CT scanner 106).

[0065] In various aspects, as shown, photon-count reduced sinogram 204(1) can have a set of pixels 408. As described above, each sinogram in sinogram set 108 and each sinogram in photon-count reduced sinogram set 204 can have the same number of pixels or the same arrangement of pixels as each other. Thus, because pixel set 402 in sinogram 108(1) can have a cardinality of p, pixel set 408 in photon-count reduced sinogram 204(1) can also have a cardinality of p. That is, pixel set 408 can include p pixels, i.e., pixel 408(1) through pixel 408(p). Like pixel set 406, each pixel in pixel set 408 is initially empty. That is, each pixel in pixel set 408 begins with a pixel value of zero. For example, pixel 408(1) begins with a pixel value of zero (e.g., before applying photon-wise binary selection by splitting component 116), indicating that no recorded photons have yet been represented by pixel 408(1). As another example, pixel 408(p) begins with a pixel value of zero (e.g., before applying photon-wise binary selection by splitting component 116), indicating that no recorded photons have yet been represented by pixel 408(p). However, this is merely a non-limiting embodiment. In various other cases, each pixel in set of pixels 408 can be initialized in any other suitable manner.

[0066] In any case, the sets of pixels 408 may each correspond (e.g., one-to-one) to the sets of pixels 402. For example, pixel 408(1) may correspond to pixel 402(1) (e.g., both pixel 402(1) and pixel 408(1) may correspond to the first tuple of the gantry angle and projection distance of the PCD-CT scanner 106). Similarly, pixel 408(p) may correspond to pixel 402(p) (e.g., both pixel 402(p) and pixel 408(p) may correspond to the pth tuple of the gantry angle and projection distance of the PCD-CT scanner 106).

[0067] In various embodiments, the splitting component 116 can iterate through each pixel in the set of pixels 402. Furthermore, for a given pixel in the set of pixels 402, the splitting component 116 can iterate through each recorded photon contained within or represented by the given pixel. For any recorded photon within the given pixel, the splitting component 116 can probabilistically assign the recorded photon to a pixel in the set of pixels 406 that corresponds to the given pixel or to a pixel in the set of pixels 408 that corresponds to the given pixel. Such probabilistic assignment of recorded photons can be thought of as, or referred to as, photon number splitting by photon-wise binary selection. Non-limiting examples of such photon-wise binary selection are shown in FIGS. 5-6.

[0068] FIG. 5 illustrates, in a non-limiting, exemplary manner, how splitting component 116 can iteratively apply photon-by-photon binary choice photon number splitting to each recorded photon of pixel 402(1) (e.g., to each photon of set of recorded photons 404(1)).

[0069] In various aspects, the split component 116 can consider the recorded photons 404(1)(1) of pixel 402(1). The recorded photons 404(1)(1) can be represented by pixel 402(1), and pixel 402(1) can correspond to pixels of both pixel 406(1) of the photon-number-reduced sinogram 202(1) and pixel 408(1) of the photon-number-reduced sinogram 204(1). So, the split component 116 can apply a binary selection in photon units to assign or allocate the recorded photons 404(1)(1) to either pixel 406(1) or pixel 408(1). In other words, the split component 116 can add the recorded photons 404(1)(1) to pixel 406(1) with a probability of d (where d is a suitable real number and 0 < d < 1), or the split component 116 can add the recorded photons 404(1)(1) to pixel 408(1) with a probability of 1 - d (for example, the complement of d). In still other words, the recorded photons 404(1)(1) are placed within pixel 406(1) with a 100d percent probability (for example, thereby incrementing the value of pixel 406(1) by 1), or are placed within pixel 408(1) with a 100(1 - d) percent probability (for example, thereby incrementing the value of pixel 408(1) by 1).

[0070] Similarly, the split component 116 can consider the recorded photons 404(1)(C 1_Total ) of pixel 402(1). The recorded photons 404(1)(C 1_Total ) can be represented by pixel 402(1), and pixel 402(1) can correspond to pixels of both pixel 406(1) of the photon-number-reduced sinogram 202(1) and pixel 408(1) of the photon-number-reduced sinogram 204(1). So, the split component 116 can apply a binary selection in photon units to assign or allocate the recorded photons 404(1)(C 1_Total ) to either pixel 406(1) or pixel 408(1). In other words, the split component 116 can, with a probability of d, the recorded photons 404(1)(C1_Total ) to pixel 406(1), or splitting component 116 can instead split recorded photon 404(1) (C 1_Total ) can be applied to pixel 408(1). In other words, the recorded photon 404(1) (C 1_Total ) has a 100d percent chance of being placed in pixel 406(1) (e.g., it increments the value of pixel 406(1) by 1) or a 100(1-d) percent chance of being placed in pixel 408(1) (e.g., it increments the value of pixel 408(1) by 1).

[0071] Consider now Figure 6, which illustrates, in a non-limiting, exemplary manner, how splitting component 116 can iteratively apply a photon-by-photon binary selection to each recorded photon of pixel 402(p) (e.g., to each photon of set 404(p) of recorded photons).

[0072] In various aspects, splitting component 116 can consider recorded photon 404(p)(1) of pixel 402(p). Because recorded photon 404(p)(1) can be represented by pixel 402(p), which can correspond to both pixel 406(p) of reduced-photon-count sinogram 202(1) and pixel 408(p) of reduced-photon-count sinogram 204(1), splitting component 116 can apply a photon-wise binary choice to assign or allocate recorded photon 404(p)(1) to either pixel 406(p) or pixel 408(p). In other words, splitting component 116 can add recorded photon 404(p)(1) to pixel 406(p) with probability d, or splitting component 116 can add recorded photon 404(p)(1) to pixel 408(p) with probability 1-d. In still other words, recorded photon 404(p)(1) has a 100d percent chance of being placed in pixel 406(p) (e.g., causing the value of pixel 406(p) to be incremented by 1), or a 100(1-d) percent chance of being placed in pixel 408(p) (e.g., causing the value of pixel 408(p) to be incremented by 1).

[0073] Similarly, the splitting component 116 splits the recorded photon 404(p) (C p_Total ) can be considered. The recorded photon 404(p)(C p_Total ) can be represented by pixel 402(p), which can correspond to both pixel 406(p) in reduced-photon-count sinogram 202(1) and pixel 408(p) in reduced-photon-count sinogram 204(1), so splitting component 116 applies a photon-wise binary selection to split recorded photon 404(p) (C p_Total ) to either pixel 406(p) or pixel 408(p). In other words, splitting component 116 can allocate or allocate recorded photon 404(p) (C p_Total) to pixel 406(p), or splitting component 116 can instead split the recorded photon 404(p) (C p_Total ) can be applied to pixel 408(1). In other words, the recorded photon 404(p)(C p_Total ) has a 100d percent chance of being placed in pixel 406(p) (e.g., it increments the value of pixel 406(p) by 1) or a 100(1-d) percent chance of being placed in pixel 408(p) (e.g., it increments the value of pixel 408(p) by 1).

[0074] In this manner, splitting component 116 can use photon-wise binary selection to probabilistically split the photons recorded in sinogram 108(1) between reduced-photon-count sinogram 202(1) and reduced-photon-count sinogram 204(1), as further described in FIG.

[0075] As shown, FIG. 7 illustrates a non-limiting, exemplary embodiment of a reduced-photon-count sinogram 202(1) after application of photon-wise binary selection by the splitting component 116.

[0076] In various embodiments, as shown, the set of pixels 406 of the reduced photon count sinogram 202(1) is not empty. For example, pixel 406(1) is C 1_Partial and C 1_Partial is any suitable positive integer, where C 1_Partial <C 1_Total Hence the term "reduced photon count." In other words, pixel 406(1) can be thought of as containing or representing a set of recorded photons 702(1), where the cardinality of set of recorded photons 702(1) is C 1_Partial That is, the set of recorded photons 702(1) is equal to the set of recorded photons 702(1)(C 1_Partial) can be included.

[0077] As another example, pixel 406(p) is C p_Partial and C p_Partial is any suitable positive integer, where C p_Partial <C p_Total Hence, the term "reduced photon count" is used again. That is, pixel 406(p) can be thought of as containing or representing a set of recorded photons 702(p), where the cardinality of the set of recorded photons 702(p) is C p_Partial In other words, the set of recorded photons 702(p) is equal to the set of recorded photons 702(p)(C p_Partial ) can be included.

[0078] Also as shown, FIG. 7 illustrates a non-limiting, exemplary embodiment of a reduced-photon-count sinogram 204(1) after application of photon-wise binary selection by the splitting component 116.

[0079] In various embodiments, as shown, the set of pixels 408 of the reduced photon count sinogram 204(1) is not empty. For example, pixel 408(1) is C 1_Remainder and C 1_Remainder is any suitable positive integer, where C 1_Remainder +C 1_Partial =C 1_Total In other words, pixel 408(1) can be thought of as containing or representing a set of recorded photons 704(1), where the cardinality of the set of recorded photons 704(1) is C 1_Remainder That is, the set of recorded photons 704(1) is equal to the set of recorded photons 704(1)(1) from the set of recorded photons 704(1)(C 1_Remainder ) can be included.

[0080] As another example, pixel 408(p) is C p_Remainder and Cp_Remainder is any suitable positive integer, where C p_Remainder +C p_Partial =C p_Total , hence the term "reduced photon count" being used again. That is, pixel 408(p) can be thought of as containing or representing a set of recorded photons 704(p), where the cardinality of the set of recorded photons 704(p) is C p_Remainder In other words, the set of recorded photons 704(p) is equal to the set of recorded photons 704(p)(C p_Remainder ) can be included.

[0081] In addition, C 1_Remainder +C 1_Partial =C 1_Total This is because, when applying photon-wise binary selection, any photon in the set of recorded photons 404(1) that is not allocated or assigned to pixel 406(1) is allocated or assigned to pixel 408(1).

[0082] Similarly, C p_Remainder +C p_Partial =C p_Total This is because, when applying photon-wise binary selection, photons from the set of recorded photons 404(p) that are not allocated or assigned to pixel 406(p) are allocated or assigned to pixel 408(p).

[0083] Thus, photon-count reduced sinogram 202(1) and photon-count reduced sinogram 204(1) are complementary to one another, and the pixel-wise sum of photon-count reduced sinogram 202(1) and photon-count reduced sinogram 204(1) is equal to sinogram 108(1). For at least this reason, application of photon-wise binary selection to sinogram 108(1) by splitting component 116 can be thought of as splitting sinogram 108(1) into photon-count reduced sinogram 202(1) and photon-count reduced sinogram 204(1).

[0084] Furthermore, when applying the binary selection in photon units, it becomes as follows.

Number

Number

[0085] In any case, the splitting component 116 can perform a binary selection in photon units for each of the set of sinograms 108, thereby splitting the photons recorded in the set of sinograms 108 between the set of sinograms 202 with reduced photon number and the set of sinograms 204 with reduced photon number (for example, the splitting component 116 can perform a binary selection in photon units for the sinogram 108(n) so as to generate the sinogram 202(n) with reduced photon number and the sinogram 204(n) with reduced photon number).

[0086] Although the disclosure herein primarily describes various embodiments of splitting component 116 as assigning or allocating photons one-by-one (e.g., through a series of Bernoulli trials), this is merely a non-limiting example. In various cases, splitting component 116 can perform photon-by-photon binary selection on multiple photons at a time (e.g., rather than probabilistically assigning a single photon from pixel 402(1) to pixel 406(1) with probability d or to pixel 408(1) with probability 1-d at a time, splitting component 116 can probabilistically assign multiple photons from pixel 402(1) to pixel 406(1) with probability d or to pixel 408(1) with probability 1-d at a time). In some cases, the multiple photons at a time can be the total number of recorded photons for each pixel. In various other cases, the splitting component 116 can perform photon-wise binomial selection in batches by sampling a binomial distribution (e.g., the binomial distribution associated with pixel 402(1) is the photon count (C 1_Total ) and a threshold probability d, and by sampling this binomial distribution, splitting component 116 can probabilistically determine how many of the recorded photons of pixel 402(1) should be assigned to pixel 406(1), and splitting component 116 can assign the recorded photons of pixel 402(1) that remain after this sampling to pixel 408(1).

[0087] 8 shows a flow diagram of an exemplary, non-limiting computer-implemented method 800 that can assist in photon number segmentation of raw sinograms, in accordance with one or more embodiments described herein. In various cases, the denoising training system 102 can assist in the computer-implemented method 800.

[0088] In various embodiments, operation 802 includes accessing, by a device (e.g., by access component 114) operably coupled to a processor (e.g., processor 110), a raw sinogram (e.g., sinogram 108(1)) acquired by a photon-counting computed tomography scanner (e.g., PCD-CT scanner 106).

[0089] In various aspects, operation 804 includes generating, by the device (e.g., by segmentation component 116), a first sinogram (e.g., sinogram 202(1)) and a second sinogram (e.g., sinogram 204(1)), each sinogram having the same pixel dimensions (e.g., the same number of pixels or the same arrangement of pixels) as the raw sinogram, and each sinogram being initially empty (e.g., the pixel values ​​of each sinogram are zero).

[0090] In various embodiments, operation 806 includes determining whether all pixels in the raw sinogram have been photon number segmented by the device (e.g., by segmentation component 116). If so (e.g., all pixels in the raw sinogram have already been photon number segmented), computer-implemented method 800 may end at operation 808. If not (e.g., at least one pixel in the raw sinogram has not yet been segmented based on a photon count value), computer-implemented method 800 may proceed to operation 810.

[0091] In various cases, operation 810 includes selecting, by the device (eg, by the splitting component 116), a pixel from the raw sinogram that has not yet been photon-number split (eg, pixel 402(p)).

[0092] In various embodiments, operation 812 includes determining, by the device (e.g., by splitting component 116), whether all recorded photons (e.g., 404(p)) of the selected pixel have been reallocated. If so (e.g., if all recorded photons of the selected pixel have been reallocated), computer-implemented method 800 returns to operation 806. If not yet reallocated (e.g., if at least one recorded photon of the selected pixel has not yet been reallocated), computer-implemented method 800 can proceed to operation 814.

[0093] In various embodiments, operation 814 may involve splitting by the device (e.g., by splitting component 116) any recorded photons (e.g., 404(p)(C)) from the selected pixel that have not yet been reallocated. p_Total )) is selected.

[0094] In various cases, operation 816 includes generating a random number between 0 and 1 by the device (eg, by the division component 116).

[0095] In various aspects, operation 818 includes determining, by the device (e.g., by division component 116), whether the random number is greater than a threshold probability value (e.g., d). If the random number is not greater than the threshold probability value (e.g., if the random number is less than or equal to the threshold probability value and has a likelihood of occurrence equal to the threshold probability value), computer-implemented method 800 can proceed to operation 820. If the random number is greater than the threshold probability value (e.g., if the random number is greater than the threshold probability value and has a likelihood of occurrence equal to the complement of the threshold probability value), computer-implemented method 800 can proceed to operation 822.

[0096] In various embodiments, operation 820 includes assigning, by the device (e.g., by segmentation component 116), the selected photons to corresponding pixels (e.g., 406(p)) of the first sinogram. The computer-implemented method 800 returns to operation 812.

[0097] In various cases, operation 822 includes assigning the selected photons by the device (e.g., by splitting component 116) to corresponding pixels (e.g., 408(p)) of the second sinogram. The computer-implemented method 800 returns to operation 812.

[0098] It should be noted that in various aspects, operations 814, 816, 818, 820, and 822 can be considered collectively as performing a binary choice on a photon-by-photon basis.

[0099] 9 shows a block diagram of an exemplary, non-limiting system 900 including at least one training input image and at least one training output image that can assist in generating training data for denoising images by photon number splitting, in accordance with one or more embodiments described herein. As shown, system 900 can, in some cases, include the same components as system 200 and can further include training input image 902 and training output image 904.

[0100] In various embodiments, the reconstruction component 118 can electronically generate the training input images 902 based on the set of reduced photon count sinograms 202. Similarly, the reconstruction component 118 can electronically generate the training output images 904 based on the set of reduced photon count sinograms 204. Various non-limiting aspects are illustrated in FIG.

[0101] FIG. 10 shows an exemplary, non-limiting block diagram 1000 illustrating how training input images 902 and training output images 904 can be generated based on reduced photon count sinogram set 202 and reduced photon count sinogram set 204 according to one or more embodiments described herein.

[0102] In various embodiments, as shown, the reconstruction component 118 can electronically perform any suitable image reconstruction technique on the photon-count reduced sinogram set 202, thereby generating the training input images 902. As some non-limiting examples, image reconstruction techniques can include material decomposition or filtered back projection. In any event, the training input images 902 can be suitable CT images having any suitable format, size, or dimensions. For example, in some cases, the training input images 902 can be two-dimensional pixel arrays of Hounsfield unit values ​​generated from the photon-count reduced sinogram set 202. As another example, in other cases, the training input images 902 can be three-dimensional voxel arrays of Hounsfield unit values ​​generated from the photon-count reduced sinogram set 202.

[0103] In situations where the reconstruction component 118 applies material decomposition, the training input images 902 can include a set of reference material input images 1002. In various embodiments, the set of reference material input images 1002 can include x reference material images (where x is any suitable positive integer), i.e., reference material input image 1002(1) through reference material input image 1002(x). In various embodiments, the reference material input image 1002(1) can be any suitable CT image having any suitable format or dimensions (e.g., a two-dimensional pixel array of Hounsfield unit values, a three-dimensional voxel array of Hounsfield unit values) and depicting any suitable reference material (e.g., calcium). Similarly, in various embodiments, the reference material input image 1002(x) can be any suitable CT image having the same format or dimensions as the reference material input image 1002(1), but depicting a different or unique reference material (e.g., water) than the reference material input image 1002(1).

[0104] In various embodiments, similar to the above, reconstruction component 118 can electronically perform any suitable image reconstruction technique on reduced-photon-count sinogram set 204, thereby generating training output images 904. Again, as some non-limiting examples, image reconstruction techniques can include material decomposition or filtered back projection. In any event, training output images 904 can be any suitable CT images having the same format, size, or dimensions as training input images 902. For example, if training input images 902 are two-dimensional pixel arrays of Hounsfield unit values ​​generated from reduced-photon-count sinogram set 202, training output images 904 can be two-dimensional pixel arrays of Hounsfield unit values ​​generated from reduced-photon-count sinogram set 204. As another example, if the training input images 902 are three-dimensional voxel arrays of Hounsfield unit values ​​generated from the set of reduced-photon sinograms 202, the training output images 904 can be three-dimensional voxel arrays of Hounsfield unit values ​​generated from the set of reduced-photon sinograms 204.

[0105] Similar to the above, in situations where the reconstruction component 118 applies material decomposition, the training output images 904 can include a set of reference material output images 1004. In various embodiments, the set of reference material output images 1004 can have the same cardinality as the set of reference material input images 1002. Thus, because the set of reference material input images 1002 includes x reference material images, the set of reference material output images 1004 can similarly include x reference material images, i.e., reference material output image 1004(1) through reference material output image 1004(x). In various embodiments, the sets of reference material output images 1004 can each correspond (e.g., one-to-one) to the sets of reference material input images 1002. For example, the reference material output image 1004(1) can correspond to the reference material input image 1002(1), meaning that the reference material output image 1004(1) and the reference material input image 1002(1) can have the same format or dimensions as each other and can represent the same reference material as each other. Similarly, the reference material output image 1004(x) can correspond to the reference material input image 1002(x), which means that the reference material output image 1004(x) and the reference material input image 1002(x) can have the same format or dimensions as each other and can represent the same reference material as each other.

[0106] 10 illustrates the training input images 902 and training output images 904 as being reference material images (e.g., as being material decomposition images), this is merely a non-limiting example. In various aspects, the training input images 902 and training output images 904 can be any suitable type of images (e.g., can be virtual monoenergetic images).

[0107] It should be noted that because sinogram set 108 can be considered to represent a projection of the patient's anatomy scanned by PCD-CT scanner 106, both photon count reduced sinogram set 202 and photon count reduced sinogram set 204 can also be considered to represent a projection of the patient's anatomy. Thus, both training input images 902 and training output images 904 can be considered to visually represent or visually depict the patient's anatomy. However, as discussed above, probability d can be considered a measure of the radiation dose level of photon count reduced sinogram set 202, and probability (1-d) can be considered a measure of the radiation dose level of photon count reduced sinogram set 204. Thus, because training input image 902 can be generated from reduced-photon-count sinogram set 202 and training output image 904 can be generated from reduced-photon-count sinogram set 204, probability d can be thought of as a measure of the radiation dose level of training input image 902, and probability (1-d) can be thought of as a measure of the radiation dose level of training output image 904. Thus, training input image 902 and training output image 904 can be thought of as representing the same anatomical structure of the same patient at different dose levels (e.g., reference material input image 1002(1) and reference material output image 1004(1) can be thought of as representing the same anatomical structure of the same patient at different dose levels through the same reference material, and reference material input image 1002(x) and reference material output image 1004(x) can be thought of as representing the same anatomical structure of the same patient at different dose levels through the same reference material).

[0108] When the probability d is less than 50% (for example, when 0 < d < 0.5), the probability (1 - d) is greater than 50% (for example, 1 > (1 - d) > 0.5). In this case, the radiation dose level of the training input image 902 is lower than the radiation dose level of the training output image 904. In this case, it can be considered that the training input image 902 exhibits or has more visual noise than the training output image 904 (for example, the reference material input image 1002(1) has more visual noise than the reference material output image 1004(1), and the reference material input image 1002(x) has more visual noise than the reference material output image 1004(x)).

[0109] When the probability d is equal to 50% (for example, when 0 < d = 0.5), the probability (1 - d) is also equal to 50% (for example, 1 > (1 - d) = 0.5). In this case, the radiation dose level of the training input image 902 is equal to the radiation dose level of the training output image 904. In this case, it can be considered that the training input image 902 exhibits or has the same level of visual noise as the training output image 904 (for example, the reference material input image 1002(1) has the same level of visual noise as the reference material output image 1004(1), and the reference material input image 1002(x) has the same level of visual noise as the reference material output image 1004(x)).

[0110] When the probability d is greater than 50% (for example, when 0.5 < d < 1), the probability (1 - d) is less than 50% (for example, 0.5 > (1 - d) > 0). In this case, the radiation dose level of the training input image 902 is higher than the radiation dose level of the training output image 904. In this case, it can be considered that the training input image 902 exhibits or has less visual noise than the training output image 904 (for example, the reference material input image 1002(1) has less visual noise than the reference material output image 1004(1), and the reference material input image 1002(x) has less visual noise than the reference material output image 1004(x)).

[0111] In either case (e.g., for a particular value of probability d), the noise present in the training input images 902 is uncorrelated (e.g., independent) from the noise present in the training output images 904. This noise independence arises from applying photon number splitting with photon-wise binary selection.

[0112] In various embodiments, the training component 120 of the denoising training system 102 can perform supervised training on the deep learning neural network 104 based on the training input images 902 and the training output images 904. More specifically, the training output images 904 can be considered ground truth annotations or targets corresponding to the training input images 902. Various non-limiting aspects are illustrated in FIG.

[0113] FIG. 11 shows an exemplary, non-limiting block diagram 1100 illustrating how a deep learning neural network 104 can be trained in accordance with one or more embodiments described herein.

[0114] In various embodiments, before training begins, the training component 120 can initialize any trainable internal parameters (e.g., convolution kernels, weight matrices, bias values) of the deep learning neural network 104 in any suitable manner (e.g., random initialization).

[0115] In various embodiments, the training component 120 can execute the deep learning neural network 104 on the training input images 902. In various embodiments, this execution can cause the deep learning neural network 104 to generate outputs 1102. More specifically, in some cases, an input layer of the deep learning neural network 104 can receive the training input images 902 (e.g., the set of reference material input images 1002), the training input images 902 (e.g., the set of reference material input images 1002) can complete a forward pass through one or more hidden layers of the deep learning neural network 104, and the output layer of the deep learning neural network 104 can calculate outputs 1102 based on the activations provided by the one or more hidden layers. In some cases, as shown, the outputs 1102 can include a set of reference material outputs 1104 each corresponding to the set of reference material input images 1002. That is, since the set of reference material input images 1002 has a cardinality of x, the set of reference material outputs 1104 can also have a cardinality of x, ie, reference material output 1104(1) through reference material output 1104(x).

[0116] In various aspects, output 1102 can be considered a predicted or inferred result (e.g., a denoised predicted / inferred result) that deep learning neural network 104 believes should correspond to training input image 902. In contrast, training output image 904 can be considered or treated as a correct and accurate result or ground truth (e.g., a denoised correct / accurate result) that is believed to correspond to training input image 902. In other words, reference material output 1104(1) can be considered a predicted or inferred result that deep learning neural network 104 believes should correspond to reference material input image 1002(1), while reference material output image 1004(1) can be considered or treated as a correct and accurate result or ground truth that corresponds to reference material input image 1002(1). Similarly, reference material output 1104(x) can be considered the predicted or inferred result that deep learning neural network 104 believes should correspond to reference material input image 1002(x), while reference material output image 1004(x) can be considered or treated as the correct and accurate result, or ground truth, that corresponds to reference material input image 1002(x). Note that if deep learning neural network 104 has received no or little prior training, output 1102 may be significantly inaccurate (e.g., output 1102 may be significantly different from training output image 904, reference material output 1104(1) may be significantly different from reference material output image 1004(1), reference material output 1104(x) may be significantly different from reference output image 1004(x)).

[0117] In various embodiments, as shown, training component 120 can calculate an error or loss (e.g., MAE, MSE, cross-entropy) between output 1102 and training output images 904 (e.g., between reference material output 1104(1) and reference material output image 1004(1), between reference material output 1104(x) and reference material output image 1004(x)). In various embodiments, training component 120 can incrementally update trainable internal parameters of deep learning neural network 104 by backpropagation based on the calculated error or loss.

[0118] It should be noted that the above-described training procedure that can be performed by the training component 120 illustrates merely a non-limiting example in which the training batch size is “1.” This example is used for ease of explanation and illustration. In various aspects, the access component 114 can access multiple sets of sinograms generated by photon-counting detector technology, the segmentation component 116 and the reconstruction component 118 can generate multiple training input images and corresponding multiple training output images based on the multiple sinogram sets, and the multiple training input images and corresponding multiple training output images can be considered an entire training dataset, and the training component 120 can train the deep learning neural network 104 with such training dataset. In other words, the training component 120 can implement any suitable training batch size to train the deep learning neural network 104. Furthermore, the training component 120 can implement any suitable error function / loss function or any suitable training termination criterion.

[0119] Since the training input image 902 and the training output image 904 can have independent noise, it should be noted that the deep learning neural network 104 can be considered to learn a method for removing noise from the input CT image regardless of d (for example, 0 < d ≤ 0.5 or 0.5 < d < 1). That is, even when the training input image 902 has less noise than the training output image 904 (for example, when 0.5 < d < 1), due to the fact that the noise present in the training input image 902 is independent (for example, uncorrelated) from the noise present in the training output image 904, the deep learning neural network 104 can be considered to learn a noise removal method.

[0120] In various embodiments, after the deep learning neural network 104 is trained by the training component 120, the deployment component 122 can electronically deploy the deep learning neural network 104. In other words, each time a given CT image for which noise removal is desired is encountered, the deployment component 122 can electronically execute the deep learning neural network 104 on such a given CT image, whereby the deep learning neural network 104 can generate, as a result, a CT image (for example, an inferred or predicted image from which noise has been removed from a given CT image). In various cases, the deployment component 122 can electronically transmit the resulting CT image to any suitable computing device. In various aspects, the deployment component 122 can electronically display the resulting CT image on any suitable computer screen, computer display, computer monitor, or graphical user interface.

[0121] Various non - limiting aspects are further described below.

[0122] In the diagnostic energy range, the mass attenuation coefficient can be discriminated as a linear combination of a plurality of reference materials as follows.

Equation

[0123] Using the material discrimination method, the line integral of the discrimination coefficient can be estimated from the spectrum measurement. In the case of PCD, the measured value can be formulated as follows.

Equation

[0124] In various embodiments, using a binary selection of photons, the full-dose photon count value can be divided into two independent low-dose count values that follow the Poisson distribution. That is, it can be divided as follows.

Equation

[0125] 12-13 show exemplary, non-limiting experimental results demonstrating various advantages of one or more embodiments described herein.

[0126] The inventors conducted various experiments. In these experiments, a first deep learning neural network was trained using low-dose training inputs (e.g., some inputs were acquired at 10% dose, others at 30% dose, and still others at 50% dose) and completely noise-free training outputs to perform image denoising. This training strategy was called Ld2Clean (e.g., "low-dose to clean"). This can be considered as an existing training method that can theoretically achieve optimal performance.

[0127] In this experiment, a second deep learning neural network was trained to perform image denoising using half-dose training inputs (e.g., inputs acquired at 50% dose) and completely noise-free training outputs. This training strategy was called Half2Clean (e.g., "half-dose to clean"). It can be considered an existing training method that can theoretically achieve optimal performance (when training on half-dose images).

[0128] In this experiment, we trained a third deep learning neural network to perform image denoising using half-dose training inputs simulated (e.g., simulated at 50% dose) from full-dose training outputs (e.g., outputs obtained at 100% dose). We termed this training strategy Half2Fd (e.g., "half-dose to full-dose"). This can be considered a conventional training method (note that there is correlated noise between the training inputs and the training outputs).

[0129] In this experiment, a fourth deep learning neural network was trained to perform image denoising using low-dose training inputs and remaining-dose training outputs generated by photon-wise binary selection as described herein. (For example, some inputs had a 10% dose level d, meaning that their corresponding output had a 90% remaining dose level (1-d). Other inputs had a 30% dose level d, meaning that their corresponding output had a 70% dose level (1-d). Yet other inputs had a 50% dose level d, meaning that their corresponding output had a 50% remaining dose level (1-d).) This training strategy was referred to as Ld2Rd (e.g., "low-dose to remaining-dose"). This may be considered a non-limiting embodiment of the method described herein.

[0130] In this experiment, a fifth deep learning neural network was trained to perform image denoising using half-dose training inputs and half-dose training outputs generated by photon-wise binary selection, as described herein (e.g., the input has a dose level d of 50%, which means that its corresponding output has a remaining dose level (1-d) of 50%). Such a training strategy was called Half2Half (e.g., "half-dose to half-dose"). This can be considered another non-limiting embodiment of the method described herein.

[0131] After training, these five deep learning neural networks were tested and validated. The test and validation results are shown in Figures 12 and 13.

[0132] In particular, FIG. 12 is a bar graph 1200 illustrating the performance results of the second, third, and fifth deep learning neural networks. That is, the bar graph 1200 illustrates how Half2Clean, Half2Fd, and Half2Half perform relative to one another. Note that the x-axis of the bar graph 1200 indicates the dose levels of the test inputs provided to the second, third, and fifth deep learning neural networks for validation. Note also that the y-axis of the bar graph 1200 represents the measured root mean square error in Hounsfield units, which is calculated for the second, third, and fifth deep learning neural networks based on the test / validation results of the second, third, and fifth deep learning neural networks. As shown, Half2Fd (the existing technology) significantly underperformed Half2Clean (the theoretical optimum) at all dose levels. In contrast, Half2Half (embodiments described herein) performs comparable (e.g., nearly comparable, with no statistically significant differences) to Half2Clean at all dosage levels. These experimental results serve to demonstrate various advantages of training a deep learning neural network to perform image denoising using photon-wise binary choice photon number partitioning.

[0133] Similarly, FIG. 13 is a bar graph 1300 illustrating the performance results of the first, second, fourth, and fifth deep learning neural networks. That is, the bar graph 1300 illustrates how Ld2Clean, Half2Clean, Ld2Rd, and Half2Half perform relative to one another. Note that the x-axis of the bar graph 1300 indicates the dose levels of the test inputs provided to the first, second, fourth, and fifth deep learning neural networks for validation. Note also that the y-axis of the bar graph 1300 represents the measured root mean square error in Hounsfield units, which is calculated for the first, second, fourth, and fifth deep learning neural networks based on the test / validation results of the first, second, fourth, and fifth deep learning neural networks. As shown, Half2Half (an embodiment described herein) performs comparable (e.g., approximately equivalent, with no statistically significant difference) to Half2Clean (the theoretical optimum when trained at half dose) at all dose levels. Similarly, as shown, Ld2Rd (another embodiment described herein) performs comparable (e.g., approximately equivalent, with no statistically significant difference) to Ld2Clean (the theoretical optimum when trained at low dose) at all dose levels. Again, these experimental results help demonstrate various advantages of training a deep learning neural network to perform image denoising using photon-wise binary choice photon number partitioning (e.g., being able to match denoising performance associated with using completely noise-free ground truth without the problems associated with obtaining completely noise-free ground truth). Furthermore, Ld2Rd and Ld2Clean outperformed Half2Half and Half2Clean (by a statistically significant margin) at lower test dose levels (e.g., 10% dose), demonstrating the various benefits of training at different low dose levels.

[0134] 14 illustrates a flow diagram of an exemplary, non-limiting computer-implemented method 1400 that can assist in generating training data for denoising images using photon number splitting, in accordance with one or more embodiments described herein. In various cases, the denoising training system 102 can assist in the computer-implemented method 1400.

[0135] In various embodiments, operation 1402 includes accessing (e.g., by access component 114) a set of sinograms (e.g., set of sinograms 108) generated by a photon-counting computed tomography scanner (e.g., PCD-CT scanner 106) by a device operably coupled to a processor (e.g., processor 110).

[0136] In various aspects, operation 1404 includes splitting, by the device (e.g., by splitting component 116), the set of sinograms into a first set of sinograms with reduced photon counts (e.g., 202) and a second set of sinograms with reduced photon counts (e.g., 204).

[0137] In various embodiments, operation 1406 includes converting the first set of sinograms with reduced photon counts into at least one training input image (e.g., 902) by the device (e.g., by reconstruction component 118) and by image reconstruction, and converting the second set of sinograms with reduced photon counts into at least one training output image (e.g., 904).

[0138] In various aspects, operation 1408 includes training, by the device (e.g., by training component 120), a deep learning neural network (e.g., 104) based on the at least one training input image and the at least one training output image.

[0139] Although not explicitly shown in FIG. 14, the first set of reduced photon numbers and the second set of reduced photon numbers are complementary, and adding the first set of reduced photon numbers and the second set of reduced photon numbers together results in a set of sinograms, pixel by pixel (e.g., as described in FIG. 7).

[0140] Although not explicitly shown in FIG. 14 , dividing the set of sinograms into a first set with a reduced number of photons and a second set with a reduced number of photons may include photon-wise binary selection, where the photon-wise binary selection can probabilistically assign photons recorded in the set of sinograms to the first set with a reduced number of photons according to a defined probability value (e.g., d) and the photon-wise binary selection can probabilistically assign photons recorded in the set of sinograms to the second set with a reduced number of photons according to the complement of the defined probability value (e.g., 1-d), and the photon-wise binary selection can make the training input images and the training output images noise-independent. In various cases, the defined probability value can be less than 50% to make at least one training input image noisier than at least one training output image. In other various cases, the defined probability value can be 50% to make at least one training input image as noisy as at least one training output image. In still other cases, the defined probability value may be greater than 50% such that at least one training input image is less noisy than at least one training output image.

[0141] Although not explicitly shown in FIG. 14, at least one training input image and at least one training output image may be a material decomposition image or a virtual mono-energetic image.

[0142] Although not explicitly shown in FIG. 14, the set of sinograms can be organized by photon energy bins.

[0143] In this disclosure, various embodiments of sinograms are primarily described as being two-dimensional arrays of pixels, but this is merely a non-limiting example for ease of illustration and explanation. In various aspects, the sinogram can exhibit any suitable dimensions (e.g., in some examples, the sinogram can be a three-dimensional array of voxels).

[0144] Although the disclosure herein primarily describes various embodiments as applied to deep learning neural networks (e.g., 104), this is merely a non-limiting example. In various aspects, the teachings described herein may be applied to any suitable machine learning model exhibiting any suitable artificial intelligence architecture (e.g., support vector machines, naive Bayes, linear regression, logistic regression, decision trees, random forests).

[0145] In various embodiments, machine learning algorithms or models may be implemented in any suitable manner to support any suitable aspect described herein. To support some of the above-described machine learning aspects of various embodiments, consider artificial intelligence (AI) below. Various embodiments described herein may employ artificial intelligence to assist in automating one or more features or functions. Components may employ various AI-based schemes to perform various embodiments / examples disclosed herein. To perform or support the numerous decisions described herein (e.g., determining, ascertaining, inferring, calculating, predicting, foreseeing, estimating, deriving, forecasting, detecting, computing), the components described herein may examine all or a subset of the data to which they have access and infer or determine the state of a system or environment from a set of information obtained by events or data. The decisions may be employed, for example, to identify a particular situation or action, or may generate a probability distribution over multiple states. The decisions may be probabilistic, i.e., calculating a probability distribution over states of interest based on a consideration of data and events. A decision may also represent a technique employed to construct a higher level event from a set of events or data.

[0146] Such determinations can construct new events or actions from sets of observed events or stored event data, regardless of whether the events are closely correlated in time, and whether the events and data come from one or more event and data sources. The components disclosed herein can employ a variety of classification (explicitly trained (e.g., by training data) and implicitly trained (e.g., by observing behavior, preferences, historical information, receiving external information, etc.)) schemes or systems (e.g., support vector machines, neural networks, expert systems, Bayesian belief networks, fuzzy logic, data fusion engines, etc.) for performing automated or determined actions related to the claimed subject matter. Thus, classification schemes or systems can be used to automatically learn and perform numerous functions, actions, or decisions.

[0147] A classifier can map an input attribute vector z = (z1, z2, z3, z4, zn) to a confidence that the input belongs to a class, such as f(z) = confidence(class). Such classification can employ probabilistic or statistical analysis (e.g., considering analytical utility and cost) to determine the action to be taken automatically. A support vector machine (SVM) is one example of a classifier that can be used. SVMs operate by finding a hypersurface in the space of possible inputs, where the hypersurface attempts to separate triggering criteria from non-triggering events. Intuitively, the hypersurface correctly classifies test data that is close to, but not identical to, the training data. Other directed and undirected model classification techniques can be used, such as naive Bayes, Bayesian networks, decision trees, neural networks, fuzzy logic models, or probabilistic classification models that provide patterns of independence. As used herein, classification also includes statistical regression, which is used to develop priority models.

[0148] The present disclosure describes non-limiting examples. For ease of description or explanation, various parts of the present disclosure utilize the terms "each," "every," or "all" when describing various embodiments. Such use of the terms "each," "every," or "all" is intended to be non-limiting. In other words, when the present disclosure provides a description that applies to "each," "every," or "all" of a particular object or component, it should be understood that this is a non-limiting example, and it should be further understood that in various other instances, the description may apply to fewer than "each," "every," or "all" of the particular objects or components.

[0149] To provide further explanation for the various embodiments described herein, Figure 15 and the following discussion are intended to provide a brief, general description of a suitable computing environment 1500 in which various embodiments of the embodiments described herein may be implemented. While the embodiments have been described in the general context of computer-executable instructions that may be executed by one or more computers, those skilled in the art will recognize that the embodiments may also be implemented in combination with other program modules, or as a combination of hardware and software.

[0150] Generally, program modules include routines, programs, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Moreover, those skilled in the art will appreciate that the methods of the present invention can be practiced with other computer system configurations, including single-processor or multi-processor computer systems, minicomputers, mainframe computers, Internet of Things (IoT) devices, distributed computing systems, as well as personal computers, handheld computing devices, microprocessor-based or programmable consumer electronics, and the like, each of which can be operatively coupled to one or more associated devices.

[0151] The illustrated embodiments herein may also be practiced in distributed computing environments where certain tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.

[0152] A computing device typically includes a variety of media, which may include computer-readable storage media, machine-readable storage media, or communication media, and these two terms are used in different contexts herein as follows: A computer-readable storage medium or machine-readable storage medium is an available storage medium that can be accessed by a computer, and includes both volatile and non-volatile media, and both removable and non-removable media. By way of example, but not limitation, a computer-readable storage medium or machine-readable storage medium may be implemented in connection with any method or technology for storing information (such as computer-readable or machine-readable instructions, program modules, structured or unstructured data, etc.).

[0153] A computer-readable storage medium may include, but is not limited to, random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD), Blu-ray disc (BD) or other optical disc storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device, solid-state drive or other solid-state storage device, or other tangible or non-transitory medium that can be used to store the desired information. In this regard, the terms "tangible" or "non-transitory" as applied to storage, memory, or computer-readable medium in this specification are understood as modifiers that exclude only the transitory propagating signal itself, and do not disclaim any right to all standard storage, memory, or computer-readable medium that is not only the transitory propagating signal itself.

[0154] A computer-readable storage medium can be accessed by one or more local or remote computing devices, for example, by access requests, queries, or other data retrieval protocols, to enable various operations to be performed on the information stored by the medium.

[0155] Communication media typically embodies computer-readable instructions, data structures, program modules, or other structured or unstructured data in a data signal such as a modulated data signal (e.g., carrier wave or other transport technology) and includes any information delivery or transmission media. The term "modulated data signal" refers to a signal that has one or more of its characteristics set or changed in such a manner as to encode one or more information in the signal. By way of example, and not limitation, communication media include wired media (such as a wired network or direct-wired connection) and wireless media (such as acoustic, RF, infrared and other wireless media).

[0156] 15 , an exemplary environment 1500 for implementing various embodiments of the aspects described herein includes a computer 1502, which includes a processing unit 1504, a system memory 1506, and a system bus 1508. The system bus 1508 couples system components, including but not limited to the system memory 1506, to the processing unit 1504. The processing unit 1504 can be any of a variety of commercially available processors. Dual microprocessors and other multi-processor architectures can also be used as the processing unit 1504.

[0157] The system bus 1508 may be any of several types of bus structures that may be further interconnected to a memory bus (with or without a memory controller), a peripheral bus, and a local bus using any of a variety of commercially available bus architectures. The system memory 1506 includes ROM 1510 and RAM 1512. A basic input / output system (BIOS) may be stored in non-volatile memory (such as ROM, erasable programmable read-only memory (EPROM), EEPROM, etc.), and contains the basic routines that help to transfer information between components within the computer 1502, such as during start-up. The RAM 1512 may also include high-speed RAM (such as static RAM for caching data).

[0158] The computer 1502 further includes an internal hard disk drive (HDD) 1514 (e.g., EIDE, SATA), one or more external storage devices 1516 (e.g., a magnetic floppy disk drive (FDD) 1516, a memory stick or flash drive reader, a memory card reader, etc.), and a drive 1520 (e.g., a solid-state drive, an optical disk drive, etc.) capable of reading from or writing to a disk 1522 (a CD-ROM disk, a DVD, a BD, etc.). Alternatively, if a solid-state drive is involved, the disk 1522 is not included unless it is separate. While the internal HDD 1514 is illustrated as being located within the computer 1502, the internal HDD 1514 can also be configured for external use within a suitable enclosure (not shown). Additionally, although not shown in the environment 1500, a solid-state drive (SSD) can also be used in addition to or in place of the HDD 1514. HDD 1514, external storage device 1516, and drive 1520 can be connected to system bus 1508 by HDD interface 1524, external storage device interface 1526, and drive interface 1528, respectively. Interface 1524 for external drive implementations can include at least one or both of Universal Serial Bus (USB) and Institute of Electrical and Electronics Engineers (IEEE) 1394 interface technologies. Other external drive connection technologies are within the contemplated scope of the embodiments described herein.

[0159] The drives and their associated computer-readable storage media are capable of non-volatile storage of data, data structures, computer-executable instructions, etc. In computer 1502, the drives and storage media store any data in a suitable digital format. While the above description of computer-readable storage media refers to each type of storage device, those skilled in the art will understand that other types of computer-readable storage media, whether existing or developed in the future, can be used in the illustrated operating environment, and further that such storage media include computer-executable instructions for performing the methods described herein.

[0160] The drives and RAM 1512 can store a number of program modules, including an operating system 1530, one or more application programs 1532, other program modules 1534, and program data 1536. All or portions of the operating system, applications, modules, or data can also be cached in RAM 1512. The systems and methods described herein can be implemented using various commercially available operating systems or combinations of operating systems.

[0161] The computer 1502 may optionally include emulation technology. For example, a hypervisor (not shown) or other intermediary may emulate a hardware environment for the operating system 1530, and the emulated hardware may optionally differ from the hardware depicted in FIG. 15 . In such an embodiment, the operating system 1530 may include one virtual machine (VM) among multiple virtual machines (VMs) hosted on the computer 1502. Additionally, the operating system 1530 may provide a runtime environment (such as the Java Runtime Environment or the .NET Framework) for the application 1532. The runtime environment is a consistent execution environment that allows the application 1532 to run on any operating system that includes the runtime environment. Similarly, the operating system 1530 may support containers, and the application 1532 may be in the form of a container, which is a lightweight, standalone, executable software package that includes, for example, system libraries and system settings, system tools, runtime, and code for the application.

[0162] Additionally, computer 1502 can use a security module, such as a Trusted Processing Module (TPM). For example, using a TPM, a boot component can hash the next boot component in time and wait until the result matches a secure value before loading the next boot component. This process can be performed at any layer within the code execution stack of computer 1502 (e.g., at the application execution level or the operating system (OS) kernel level), thereby enabling security at any level of code execution.

[0163] A user can enter commands and information into the computer 1502 by one or more wired / wireless input devices, such as a keyboard 1538, a touch screen 1540, and a pointing device such as a mouse 1542. Other input devices (not shown) can include a microphone, an infrared (IR) remote control, a radio frequency (RF) remote control, or other remote control, a joystick, a virtual reality controller or headset, a game pad, a stylus pen, an image input device (e.g., a camera), a gesture sensor input device, a visual motion sensor input device, an emotion or face detection device, a biometric input device (e.g., a fingerprint or iris scanner, etc.). These and other input devices are often connected to the processing unit 1504 through an input device interface 1544, which can be coupled to the system bus 1508, but can also be connected by other interfaces, such as a parallel port, an IEEE 1394 serial port, a game port, a USB port, an IR interface, a BLUETOOTH interface, etc.

[0164] A monitor 1546 or other type of display device can also be connected to the system bus 1508 via an interface, such as a video adapter 1548. In addition to the monitor 1546, computers typically include other peripheral output devices (not shown), such as speakers and printers.

[0165] The computer 1502 can operate in a networked environment using logical connections to one or more remote computers (e.g., remote computer 1550) through wired or wireless communications. The remote computer 1550 can be a workstation, a server computer, a router, a personal computer, a portable computer, a microprocessor-based entertainment appliance, a peer device, or other common network node, and typically includes many or all of the elements described relative to the computer 1502, although for simplicity, only the memory / storage device 1552 is illustrated. The logical connections shown include wired and wireless connections to a local area network (LAN) 1554 or a larger network (e.g., a wide area network (WAN) 1556). Such LAN and WAN networking environments are commonplace in offices and enterprises, supporting enterprise-wide computer networks (e.g., intranets), and all of the LAN and WAN networking environments can be connected to a global communications network (e.g., the Internet).

[0166] When used in a LAN networking environment, the computer 1502 can be connected to the local network 1554 through a wired or wireless communication network interface or adapter 1558. The adapter 1558 can support wired or wireless communication to the LAN 1554, and the LAN 1554 can also include a wireless access point (AP) disposed in the LAN to communicate with the adapter 1558 in a wireless mode.

[0167] When used in a WAN networking environment, the computer 1502 may include a modem 1560 or may be connected to a communications server on the WAN 1556 through other means for establishing communications over the WAN 1556 (e.g., via the Internet). The modem 1560, which may be internal or external, and wired or wireless, may be connected to the system bus 1508 through the input device interface 1544. In a networked environment, program modules depicted relative to the computer 1502, or portions thereof, may be stored in the remote memory / storage device 1552. It will be appreciated that the network connections shown are exemplary and other means of establishing a communications link between computers may be used.

[0168] When used in either a LAN or WAN networking environment, the computer 1502 can access a cloud storage system or other network-based storage systems, such as, but not limited to, a networked virtual machine, that provide one or more aspects of information storage or processing, in addition to or instead of the external storage device 1516 described above. Generally, a connection between the computer 1502 and the cloud storage system can be established over the LAN 1554 or WAN 1556, for example, by an adapter 1558 or a modem 1560, respectively. Upon connecting the computer 1502 to an associated cloud storage system, the external storage interface 1526, with the aid of the adapter 1558 or the modem 1560, can manage the storage provided by the cloud storage system in the same way as other types of external storage. For example, the external storage interface 1526 can provide access to the cloud storage sources as if they were physically connected to the computer 1502.

[0169] The computer 1502 is operable to communicate with any wireless device or entity operatively positioned for wireless communication (e.g., a printer, a scanner, a desktop or portable computer, a portable data assistant, a communications satellite, an appliance or location associated with a wirelessly detectable tag (e.g., a kiosk, a newsstand, a store shelf, etc.), and a telephone). This includes Wireless Fidelity (Wi-Fi) and BLUETOOTH® wireless technologies. In this manner, communication can be in a predefined structure, such as a traditional network, or simply ad hoc communication between at least two devices.

[0170] FIG. 16 is a schematic block diagram of a sample computing environment 1600 with which the disclosed subject matter can interact. The sample computing environment 1600 includes one or more client(s) 1610. The client(s) 1610 can be hardware or software (e.g., threads, processes, computing devices). The sample computing environment 1600 also includes one or more server(s) 1630. The server(s) 1630 can also be hardware or software (e.g., threads, processes, computing devices). The server(s) 1630 can house threads for performing transformations, for example, by employing one or more embodiments described herein. One possible communication between the client(s) 1610 and the server(s) 1630 can be in the form of a data packet adapted to be transmitted between two or more computer processes. The sample computing environment 1600 includes a communication framework 1650 that can be employed to facilitate communication between the client(s) 1610 and the server(s) 1630. The client(s) 1610 are operably connected to one or more client data store(s) 1620 that can be employed to store information local to the client(s) 1610. Similarly, the server(s) 1630 are operatively connected to one or more server data store(s) 1640 that can be employed to store information local to the servers 1630 .

[0171] The present invention is a system, method, apparatus, or computer program product that can be integrated at a technically detailed level. The computer program product can include a computer-readable storage medium having computer-readable program instructions for causing a processor to perform aspects of the present invention. The computer-readable storage medium can be a tangible device capable of holding and storing instructions for use by an instruction execution device. The computer-readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples of computer-readable storage media include, but are not limited to, portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory sticks, floppy disks, mechanically encoded devices (such as punch cards or raised structures with instructions recorded in grooves), and any suitable combination of the foregoing, although this list is not exhaustive. As used herein, computer-readable storage media is not intended to be understood as a transitory signal per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., light pulses traveling through a fiber optic cable), or electrical signals traveling through a wire.

[0172] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device over a network (e.g., the Internet, a local area network, a wide area network, or a wireless network). The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, or edge servers. A network adapter card or network interface of each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions, which are stored in a computer-readable storage medium within each computing / processing device. The computer-readable program instructions for carrying out operations of the present invention can be either assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, configuration data for integrated circuits, or source code or object code written in any combination of one or more programming languages ​​(e.g., object-oriented programming languages ​​such as Smalltalk, C++, and procedural programming languages ​​such as the “C” programming language or similar programming languages). The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server, as a stand-alone software package. In the latter case, the remote computer may be connected to the user's computer through any type of network (such as a local area network (LAN) or a wide area network (WAN)) or may be connected to an external computer (e.g., through the Internet using an Internet Service Provider).In some embodiments, to carry out aspects of the present invention, an electronic circuit (e.g., a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA)) can utilize state information of the computer-readable program instructions to execute the computer-readable program instructions and personalize the electronic circuit.

[0173] Aspects of the present invention are described herein with reference to flowchart illustrations or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations or block diagrams, and combinations of blocks in the flowchart illustrations or block diagrams, can be embodied by computer-readable program instructions. These computer-readable program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, and can generate a machine-readable program such that the instructions, executed by the processor of the computer or other programmable data processing apparatus, create means for performing the function(s) / act(s) specified in the block(s) of the flowchart illustrations or block diagrams. These computer-readable program instructions can also be stored on a computer-readable storage medium that can direct a computer, programmable data processing apparatus, or other device to function in a particular manner, and a computer-readable storage medium having instructions stored thereon includes an article of manufacture containing instructions that implement aspects of the function(s) / act(s) specified in the block(s) of the flowchart illustrations or block diagrams. Furthermore, the computer-readable program instructions can be loaded into a computer, other programmable data processing device, or other device and cause a series of operations to be executed on the computer, other programmable device, or other device to generate a computer-executed process, and the instructions executed on the computer, other programmable device, or other device can perform the functions / operations specified in one or more blocks of the flowchart or block diagram.

[0174] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowcharts or block diagrams represents a module, segment, or portion of instructions, which may include one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions noted in the blocks may be executed out of the order noted in the figures. For example, two blocks shown as successive may in fact be executed substantially concurrently, or blocks may sometimes be executed in the reverse order, depending on the functionality involved. It should also be noted that each block in the block diagrams or flowchart diagrams, and combinations of blocks in the block diagrams or flowchart diagrams, may be implemented by a system using dedicated hardware that performs the specified functions or operations or a combination of dedicated hardware and computer instructions.

[0175] Although the subject matter has been described in the general context of computer-executable instructions for a computer program product executed on one computer or multiple computers, those skilled in the art will recognize that the present disclosure can also be implemented in combination with other program modules. Generally, program modules include routines, programs, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Moreover, those skilled in the art will appreciate that the computer-implemented methods of the present invention can be practiced with other computer system configurations, including single-processor or multiprocessor computer systems, minicomputer devices, mainframe computers, as well as computers, handheld computing devices (e.g., PDAs, phones), microprocessor-based or programmable consumer or industrial electronic devices, etc. The illustrated aspects can also be practiced in distributed computing environments where tasks are performed by remote processing devices linked through a communications network. However, some, if not all, aspects of the present disclosure can be practiced on stand-alone computers. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.

[0176] In this application, the terms “component,” “system,” “platform,” “interface,” etc., may represent or include a computer-related entity having one or more specific functions or an entity associated with an operating machine having one or more specific functions. The entities disclosed herein may be hardware, a combination of hardware and software, software, or software in execution. For example, a component may be, but is not limited to, a process running on a processor, a processor, an object, an executable file, a thread of execution, a program, or a computer. By way of example, a component may be both an application running on a server and the server. One or more components may reside within an executing process or thread of execution, and the components may reside on one computer or may be distributed among two or more computers. In another example, each component may execute from various computer-readable media having various data structures stored thereon. Components may communicate, for example, through local or remote processes, according to signals comprising one or more data packets (e.g., data from one component communicating with another component in a local system, a distributed system, or other systems mediated by signals over a network such as the Internet). As another example, a component may be a device in which certain functionality is provided by mechanical parts operated by electrical or electronic circuitry driven by a software or firmware application executed by a processor, in which case the processor may be internal or external to the device and may execute at least a portion of the software or firmware application.As yet another example, a component may be a device that provides a particular function through electronic components without mechanical components, where the electronic components may include a processor or other means for executing software or firmware that provides at least a portion of the functionality of the electronic component. In some aspects, a component may emulate an electronic component, for example, by a virtual machine in a cloud computing system.

[0177] Furthermore, the term "or" is intended to mean an inclusive "or," not an exclusive "or." That is, unless specifically stated otherwise or clear from the context, "X uses A or B" is intended to mean any inclusive, natural combination. That is, if X uses A, X uses B, or X uses both A and B, any of these instances would satisfy "X uses A or B." Furthermore, as used herein, the term "and / or" is intended to have the same meaning as "or." Furthermore, the articles "a" and "an," as used in this specification and the accompanying drawings, should generally be construed to mean "one or more," unless otherwise stated or clearly indicated in the singular from the context. As used herein, the words "example" and "exemplary" are used to mean serving as an example, instance, or illustration. For the avoidance of doubt, the subject matter disclosed herein is not limited to such examples. Additionally, any aspect or design described herein as "example" and / or "exemplary" is not necessarily to be construed as preferred or advantageous over other aspects or designs, and is not intended to exclude exemplary equivalent structures and techniques known to those skilled in the art.

[0178] As used herein, the term "processor" can refer to virtually any computing processing unit or device, including, but not limited to, a single-core processor, a single processor with software multithreaded execution capabilities, a multi-core processor, a multi-core processor with software multithreaded execution capabilities, a multi-core processor using hardware multithreading techniques, a parallel platform, and a parallel platform with distributed shared memory. Furthermore, a processor can refer to an integrated circuit, an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic controller (PLC), a complex programmable logic device (CPLD), discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. Furthermore, a processor can utilize nanoscale architectures, such as, but not limited to, molecular and quantum dot-based transistors, switches, and gates, to optimize space usage or improve performance of user equipment. A processor can also be implemented as a combination of multiple computing processing units. This disclosure uses terms such as "store," "storage," "data store," "data storage," "database," and substantially any other information storage element related to the operation and functionality of a component to refer to a "memory component," an entity embodied in a "memory," or an element that includes a memory. It should be understood that the memory or memory component described herein can be volatile or non-volatile memory, or may include both volatile and non-volatile memory.By way of example, and not limitation, non-volatile memory may be read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM), flash memory, or non-volatile random access memory (RAM) (e.g., ferroelectric RAM (FeRAM)). Volatile memory may include, for example, RAM, which may act as external cache memory. By way of example, and not limitation, RAM comes in many forms, including synchronous RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), Direct Rambus RAM (DRRAM), Direct Rambus Dynamic RAM (DRDRAM), and Rambus Dynamic RAM (RDRAM). Additionally, the memory components of the systems or computer-implemented methods disclosed herein are intended to comprise, but not be limited to, these and any other suitable types of memory.

[0179] The foregoing description includes merely exemplary systems and computer-implemented methods. Of course, for purposes of describing this disclosure, it is not possible to describe every conceivable combination of components or computer-implemented methods, although many more combinations and permutations are possible. Furthermore, when the terms "comprises," "having," "possessing," and the like are used in the detailed description, claims, appendices, and drawings, such terms are intended to be inclusive, similar to the term "comprises," as interpreted when "comprises" is used as a transitional term in the claims.

[0180] The descriptions of various embodiments are provided for illustrative purposes, but are not intended to be all inclusive of all possible embodiments, nor are they intended to limit the disclosed embodiments. Many modifications and variations will be apparent without departing from the scope and spirit of the described embodiments. The terms used herein have been selected to best explain the principles of the embodiments, their practical applications that are superior to technology found in the market, or their technical improvements over technology found in the market, or to allow those skilled in the art to understand the embodiments disclosed herein. [Explanation of symbols]

[0181] 102 Noise Reduction Training System

Claims

1. A processor executing computer-executable components stored in non-transitory computer-readable memory, the computer-executable components comprising: an access component for accessing a set of sinograms generated by the photon counting computed tomography scanner; a segmentation component that segments the set of sinograms into a first set of reduced photon count sinograms and a second set of reduced photon count sinograms; a reconstruction component that converts the first set of reduced photon count sinograms into at least one training input image and the second set of reduced photon count sinograms into at least one training output image by image reconstruction; Including, the splitting component splits the set of sinograms into a first set of reduced photon counts and a second set of reduced photon counts by photon-wise binary selection, the photon-wise binary selection probabilistically assigning photons recorded in the set of sinograms to the first set of reduced photon count sinograms and the second set of reduced photon count sinograms; The computer-executable components include: a training component that trains a deep learning neural network based on the at least one training input image and the at least one training output image. A system including:

2. 2. The system of claim 1, wherein the first reduced photon number set and the second reduced photon number set are complementary, and the first reduced photon number set and the second reduced photon number set sum to the sinogram set, on a pixel-by-pixel basis.

3. 2. The system of claim 1, wherein the photon-wise binary selection probabilistically assigns photons recorded in the set of sinograms to a first set with a reduced number of photons according to a defined probability value, and the photon-wise binary selection probabilistically assigns photons recorded in the set of sinograms to a second set with a reduced number of photons according to a complement of the defined probability value, and wherein the photon-wise binary selection makes the at least one training input image and the at least one training output image independent of each other with respect to noise.

4. The system of claim 3 , wherein the defined probability value is less than 50% and the at least one training input image is noisier than the at least one training output image.

5. The system of claim 3 , wherein the defined probability value is equal to 50 percent and the at least one training input image is as noisy as the at least one training output image.

6. The system described in claim 3, wherein the defined probability value is greater than 50 percent and the at least one training input image has less noise than the at least one training output image.

7. The system of claim 1 , wherein the at least one training input image and the at least one training output image are material decomposed images or virtual monoenergetic images.

8. The system of claim 1 , wherein the set of sinograms is organized by photon energy bin.

9. a device operably coupled to the processor accessing a set of sinograms generated by a photon counting computed tomography scanner; the device partitioning the set of sinograms into a first set of reduced-photon-count sinograms and a second set of reduced-photon-count sinograms by photon-wise binary selection, the photon-wise binary selection probabilistically assigning photons recorded in the set of sinograms to the first set of reduced-photon-count sinograms and the second set of reduced-photon-count sinograms; the device converting the first set of reduced photon count sinograms into at least one training input image and the second set of reduced photon count sinograms into at least one training output image by image reconstruction; and the device training a deep learning neural network based on the at least one training input image and the at least one training output image. A computer-implemented method comprising:

10. 10. The computer-implemented method of claim 9, wherein the first reduced photon number set and the second reduced photon number set are complementary, and the first reduced photon number set and the second reduced photon number set sum to the sinogram set, on a pixel-by-pixel basis.

11. 10. The computer-implemented method of claim 9, wherein the photon-wise binary selection probabilistically assigns photons recorded in the set of sinograms to a first set with a reduced number of photons according to a defined probability value, the photon-wise binary selection probabilistically assigns photons recorded in the set of sinograms to a second set with a reduced number of photons according to the complement of the defined probability value, and the photon-wise binary selection makes the at least one training input image and the at least one training output image independent of each other with respect to noise.

12. 12. The computer-implemented method of claim 11, wherein the defined probability value is less than 50% and the at least one training input image is noisier than the at least one training output image.

13. 12. The computer-implemented method of claim 11, wherein the defined probability value is equal to 50 percent and the at least one training input image is as noisy as the at least one training output image.

14. A computer-implemented method as described in claim 11, wherein the defined probability value is greater than 50 percent and the at least one training input image has less noise than the at least one training output image.

15. 10. The computer-implemented method of claim 9, wherein the at least one training input image and the at least one training output image are material decomposed images or virtual monoenergetic images.

16. 10. The computer-implemented method of claim 9, wherein the set of sinograms is organized by photon energy bin.

17. 1. A computer program product for assisting in generating training data for image denoising by photon number splitting, the computer program product including a non-transitory computer-readable memory having program instructions embodied therein, the program instructions, when executed by a processor, causing the processor to: accessing a set of sinograms generated by a photon-counting computed tomography scanner; partitioning the set of sinograms into a first set of reduced-photon-count sinograms and a second set of reduced-photon-count sinograms by photon-wise binary selection, wherein the photon-wise binary selection probabilistically assigns photons recorded in the set of sinograms to the first set of reduced-photon-count sinograms and the second set of reduced-photon-count sinograms; converting the first set of reduced photon count sinograms into at least one training input image and the second set of reduced photon count sinograms into at least one training output image by image reconstruction; training a deep learning neural network based on the at least one training input image and the at least one training output image; A computer program product that causes the

18. 18. The computer program product of claim 17, wherein the first reduced photon number set and the second reduced photon number set are complementary, and wherein the first reduced photon number set and the second reduced photon number set sum to a set of sinograms, on a pixel-by-pixel basis.

19. 18. The computer program product of claim 17, wherein the photon-wise binary selection probabilistically assigns photons recorded in the set of sinograms to a first set with a reduced number of photons according to a defined probability value, and wherein the photon-wise binary selection probabilistically assigns photons recorded in the set of sinograms to a second set with a reduced number of photons according to a complement of the defined probability value, and wherein the photon-wise binary selection makes the at least one training input image and the at least one training output image independent of each other with respect to noise.

20. 20. The computer program product of claim 19, wherein the defined probability value is less than 50% and the at least one training input image is noisier than the at least one training output image.

Images