Systems and methods for correcting bad pixels in a computed tomography detector with a deep learning network

Abstract

The present disclosure provides a computer-implemented method for correcting artifacts in computed tomography data. The method includes inputting a sinogram into a trained sinogram correction network, where the sinogram is missing a pixel value for at least one pixel. The method also includes processing the sinogram via one or more layers of the trained sinogram correction network, where processing the sinogram includes deriving complementary information from the sinogram and estimating the pixel value for the at least one pixel based on the complementary information. The method further includes outputting, from the trained sinogram correction network, a corrected sinogram having the estimated pixel value.

Description

BACKGROUND

[0001] The subject matter disclosed herein relates to medical imaging, and more particularly, to systems and methods for utilizing a deep learning network for correcting bad pixels in a computed tomography detector.

[0002] Non-invasive imaging techniques allow for non-invasively obtaining images of internal structures or features of a subject (a patient, manufactured goods, luggage, a package, or a passenger). In particular, such non-invasive imaging techniques rely on various physical principles, such as differential transmission of X-rays through a target volume or reflection of sound waves, to acquire data and construct images or otherwise represent internal features of a subject.

[0003] For example, in X-ray based imaging techniques, a subject of interest, such as a human patient, is irradiated with X-ray radiation, and the attenuated radiation will have an effect on a detector that collects attenuation intensity data. In a digital X-ray system, the detector produces signals that represent the amount or intensity of radiation affecting discrete pixel areas of the detector surface. The signals can then be processed to generate images that can be displayed for viewing.

[0004] In one such X-ray based technique, called computed tomography (CT), a scanner can project a fan or cone shaped beam of X-rays from an X-ray source at multiple view positions onto an object being imaged, such as a patient. The X-ray beam attenuates as it passes through the object and is detected by a set of detector elements that produce signals representing the intensity of the attenuated X-ray radiation on the detector. The signals are processed to produce data representing the line integrals of the linear attenuation coefficients of the object along the X-ray path. These signals are often referred to as “projection data” or simply “projections.” By using reconstruction techniques, such as filtered back-projection, an image can be generated that represents a volume or volume rendering of the region of interest of the patient or object being imaged. In a medical context, pathological structures or other structures of interest can then be located or identified from the reconstructed image or rendered volume.

[0005] Bad pixels in a detector can result in missing data in the sinogram domain of acquired CT data and undesirable artifacts (rings and streaks) in the reconstructed image or volume. Additionally, having bad pixels in a CT detector impacts the ability to utilize the CT system. For example, a single bad pixel in the isocenter of a CT detector renders the CT system no longer usable. A certain number of bad pixels in a CT detector also renders the CT system no longer usable. SUMMARY

[0006] A summary of certain embodiments disclosed herein is set out below. It is to be understood that these aspects are merely provided as a brief summary of these particular embodiments and are not intended to limit the scope of the present disclosure. Indeed, the present disclosure can encompass various aspects not set out below.

[0007] In one embodiment, a computer-implemented method for correcting artifacts in computed tomography data is provided. The method includes inputting a sinogram into a trained sinogram correction network, where the sinogram is missing a pixel value for at least one pixel. The method further includes processing the sinogram via one or more layers of the trained sinogram correction network, where processing the sinogram includes deriving complementary information from the sinogram and estimating the pixel value for the at least one pixel based on the complementary information. The method further includes outputting, from the trained sinogram correction network, a corrected sinogram having the estimated pixel value.

[0008] In another embodiment, a computer-implemented method for generating a trained neural network to estimate missing values in computed tomography data is provided. The method includes providing training data including sinograms and complementary information derived from the sinograms, where the sinograms include sinograms without any missing pixel values and corresponding sinograms having missing pixel values simulated from the sinograms without any missing pixel values. The method further includes training a neural network using the training data to correct sinograms having missing pixel values for at least one pixel based on a combined training loss derived from both a sinogram domain of the training data and an image reconstruction domain of images reconstructed from the training data.

[0009] In another embodiment, a deep learning based sinogram correction system is provided. The system includes a memory that encodes processor-executable routines. The system also includes a processing component configured to access the memory and execute the processor-executable routines, where the routines, when executed by the processing component, cause the processing component to perform actions. The actions include inputting a sinogram into a trained sinogram correction network, where the sinogram is missing a pixel value for at least one pixel. The actions also include processing the sinogram via one or more layers of the trained sinogram correction network, where processing the sinogram includes deriving complementary information from the sinogram and estimating the pixel value for the at least one pixel based on the complementary information, where the complementary information includes multi-channel patches, and the multi-channel patches include local neighboring patches from the sinogram corresponding to portions along a channel view direction of a row having the at least one pixel, neighboring row patches from the sinogram corresponding to neighboring rows of the row having the at least one pixel, and conjugate patches from the sinogram corresponding to a conjugate region relative to the at least one pixel. The actions also include outputting a corrected sinogram having the estimated pixel value from the trained sinogram correction network. BRIEF DESCRIPTION OF DRAWINGS

[0010] These and other features, aspects, and advantages of the present application will become better understood when the following detailed description is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, wherein:

[0011] Figure 1 FIG. 1 depicts an example of an artificial neural network for training a deep learning model in accordance with aspects of the present disclosure;

[0012] Figure 2 is a block diagram depicting components of a computed tomography (CT) imaging system in accordance with aspects of the present disclosure;

[0013] Figure 3 is a schematic diagram depicting training of a neural network to correct bad pixels in accordance with aspects of the present disclosure;

[0014] Figure 4 is a schematic diagram depicting a multi-channel sinogram correction network in accordance with aspects of the present disclosure;

[0015] Figure 5 is a schematic diagram depicting a reconstruction domain analysis for determining an image reconstruction domain loss in accordance with aspects of the present disclosure;

[0016] Figure 6 is a schematic diagram depicting utilization of a trained sinogram correction network in accordance with aspects of the present disclosure; and

[0017] Figure 7 depicts different types of sinograms and corresponding image reconstructions in accordance with aspects of the present disclosure. DETAILED DESCRIPTION

[0018] One or more specific embodiments will be described below. To provide a context for these embodiments, a brief, general description of a suitable environment in which the embodiments can be implemented will be described now. In the interest of clarity, not all features of an actual implementation are described in this specification. It will of course be appreciated that in the development of any such actual implementation, numerous implementation-specific decisions must be made to achieve the developers' specific goals, such as compliance with system-related and business-related constraints, which will vary from one implementation to another. Moreover, it will be appreciated that such a development effort might be complex and time-consuming, but would nevertheless be a routine undertaking for those of ordinary skill in the art having the benefit of this disclosure.

[0019] When introducing elements of various embodiments of the present subject matter, the articles "a," "an," "the" and "said" are intended to mean that there are one or more of the elements. The terms "comprising," "including," and "having" are intended to be inclusive and mean that there can be additional elements other than the listed elements. Additionally, any numerical examples in the following discussion are intended to be non-limiting, and thus additional numerical values, ranges, and percentages are within the scope of the disclosed embodiments.

[0020] While aspects of the following discussion are provided in the context of medical imaging, it should be understood that the present technology is not limited to such a medical context. Indeed, the examples and explanations provided in such a medical context are merely for ease of discussion by providing examples of real-world implementations and applications. However, the present methods can also be used in other contexts, such as in industrial computed tomography (CT) used in non-destructive inspection of manufactured parts or articles (i.e., quality control or quality audit applications) and / or non-invasive inspection of packages, boxes, luggage, etc. (i.e., security or screening applications).

[0021] As discussed herein, artifacts (rings, streaks, etc.) found in CT images can be a symptom of a CT component problem. For example, a bad pixel in a CT detector can cause a ring artifact (a ring centered in the rotation center) that results in structured non-uniformity and degradation of image quality. These rings make the CT images unusable for diagnostic purposes. Additionally, if the number of bad pixels is above a certain threshold or if the bad pixels are located in the central region of the CT detector (system isocenter), it can result in higher service costs and can result in replacement of the CT detector.

[0022] The methods discussed herein address these issues by applying a deep learning method (e.g., a convolutional neural network or a multilayer perceptron) that utilizes supervised learning to remove these artifacts due to one or more bad pixels. The deep learning algorithm works in both the raw sinogram domain and the reconstructed domain to learn to remove distortions caused by bad pixels. Specifically, the deep learning algorithm learns to remove artifacts created by bad pixels by learning the correlation between complementary information within the sinogram domain data. For example, in one implementation, a deep neural network (or other suitable machine learning architecture) can be employed in the process. As can be appreciated, a neural network as discussed herein can be trained for use across multiple types of configurations (e.g., axial or helical scans (different pitch), varying kV / mA ratings, detector size, or bad pixel configurations (single or multiple pixels in individual or group locations)). Moreover, in some embodiments, more than one neural network can be utilized.

[0023] In view of the foregoing, a neural network as discussed herein can encompass a deep neural network, a fully connected network, a convolutional neural network (CNN), a perceptron (e.g., a multilayer perceptron (MLP)), an autoencoder, a regression network, a wavelet filter bank, or other neural network architecture. These techniques are generally referred to herein as machine learning. As discussed herein, one implementation of machine learning can be deep learning techniques, and such deep learning terminology can also be used in particular reference to the use of deep neural networks, which are neural networks having multiple layers.

[0024] As discussed herein, deep learning techniques (which can also be referred to as deep machine learning, hierarchical learning, or deep structured learning) are a branch of machine learning that employs mathematical representations of data and artificial neural networks to learn. By way of example, deep learning methods can be characterized as using one or more algorithms to extract high-level abstractions or model a class of data of interest. This can be done using one or more processing layers, where each layer generally corresponds to a different level of abstract concepts or a different stage of a process or event, and thus can employ or utilize different aspects of the initial data or the output of a previous layer (i.e., a hierarchical or cascading structure of layers) as the target of the process or algorithm of a given layer. In the context of image processing or reconstruction, this can be characterized as different layers corresponding to different levels of features or resolutions in the data. Generally, a processing from one representation space to the next can be considered a “stage” of a process. Each stage of reconstruction can be performed by a separate neural network or by different parts of a larger neural network.

[0025] As discussed herein, as part of the initial training of a deep learning process to solve a particular problem (such as identifying a service problem based on artifacts identified in image data), a training dataset can be employed having known initial values (e.g., input images, projection data (e.g., sinograms with or without missing values for bad pixels in the detector), etc.) and known or expected values for the final output (e.g., corrected sinograms, reconstructed tomographic reconstruction such as cross-sectional images or volumetric representations). A single stage of training can have known input values corresponding to one representation space and known output values corresponding to the next level representation space. In this way, the deep learning algorithm can process the known dataset or training dataset (in a supervised manner, i.e., with all training data fully labeled) until the mathematical relationship between the initial data and the expected output is seen and / or the mathematical relationship between the input and output of each layer is seen and characterized. Similarly, a separate validation dataset can be employed in which the initial and expected target values are known, but only the initial values are provided to the trained deep learning algorithm, which output is then compared to the output of the deep learning algorithm to verify prior training and / or prevent overtraining.

[0026] With the foregoing in mind, Figure 1 An example of an artificial neural network 50 is schematically depicted, which can be trained as a deep learning model as discussed herein. In this example, the network 50 is multi-layered, with a training input 52 and a plurality of layers present in the network 50 (including an input layer 54, hidden layers 58A, 58B, etc., and an output layer 60 and training target 64). In this example, each layer is composed of a plurality of “neurons” or nodes 56. The number of neurons 56 can be constant between layers, or as shown, can vary from layer to layer. The neurons 56 of each layer generate a respective output that is used as input to the neurons 56 of the next layer. In practice, a weighted sum of the inputs is computed with added bias to “fire” or “activate” each respective neuron of the layer according to an activation function, such as a rectified linear unit (ReLU), a sigmoid function, a hyperbolic tangent function, or otherwise specified or programmed. The output of the last layer constitutes the network output 60 (e.g., one or more convolution kernel parameters, convolution kernels, etc.), which is used with the training target 64 to compute some loss or error function 62 that will be backpropagated to guide network training.

[0027] The loss or error function 62 measures the difference between the network output (e.g., convolutional kernels or kernel parameters) and the training target. In certain embodiments, the loss function can be a mean absolute error (MAE) (e.g., measuring between a sinogram and a corrected sinogram). In certain implementations, the loss function can be a mean squared error (MSE) of voxel-level values or partial line integral values (e.g., between a reconstructed image derived from a measured sinogram and a reconstructed image derived from a corrected sinogram) and / or can account for differences involving other image features, such as image gradients or other image statistics. Alternatively, the loss function 62 can be defined by other metrics associated with the particular task at issue (e.g., a structural similarity index measurement (SSIM)). As described in greater detail below, a hybrid domain loss function (e.g., loss in the sinogram domain and loss in the image reconstruction domain) can be utilized during training. The following losses can also be utilized. In the sinogram domain, a content loss (e.g., LI and / or L2 loss computed between a target and a predicted sinogram) can be utilized. A perception loss (e.g., SSIM loss computed between a sinogram and a reconstruction domain) can be utilized. A transform domain loss (e.g., loss that can be computed by a filter domain / wavelet domain) can be utilized in the sinogram domain or the image reconstruction domain. Furthermore, an adversarial loss can also be used in training.

[0028] In training the example, the neural network 50 can first be constrained to be linear (i.e., by removing all non-linear units) to ensure good initialization of the network parameters. The neural network 50 can also be pre-trained in stages using a computer simulated input-target dataset, as discussed in greater detail below. After pre-training, the neural network 50 can be trained as a whole and also incorporating the non-linear units.

[0029] To facilitate explanation of the image analysis methods of the present disclosure using deep learning techniques, the disclosure discusses these methods in the context of a CT system. However, it should be understood that the following discussion can also apply to other image modalities and systems, including but not limited to PET, CT, CBCT, PET-CT, SPECT, multi-spectral CT, and any context in which tomographic reconstruction is employed to reconstruct images, whether in a medical or non-medical context.

[0030] With this in mind, Figure 2 An example of a CT imaging system 110 (i.e., a CT scanner) is depicted in FIG. 1. In the depicted example, the imaging system 110 is designed to acquire scan data (e.g., x-ray attenuation data) around a patient (or other subject or object of interest) in various views and is suitable for performing image reconstruction using tomographic reconstruction techniques. In the depicted example, the imaging system 110 includes a gantry 112 that supports a source 114 (e.g., an x-ray tube) and a detector 116 (e.g., a flat panel detector) that are configured to rotate around the patient 118. The source 114 and the detector 116 are coupled to a gantry controller 120 that controls the rotation of the source 114 and the detector 116 around the patient 118. The source 114 and the detector 116 are configured to acquire scan data (e.g., x-ray attenuation data) from the patient 118 during rotation of the source 114 and the detector 116 around the patient 118. The acquired scan data is provided to a reconstruction controller 122 that reconstructs the acquired scan data into an image of the patient 118. The reconstructed image is provided to a display 124 for viewing by a user. Figure 2In the illustrated embodiment, the imaging system 110 includes an X-ray radiation source 112 positioned proximate to a collimator 114. The X-ray source 112 can be an X-ray tube, a distributed X-ray source such as a solid state or thermionic X-ray source, or any other X-ray radiation source suitable for acquiring medical or other images.

[0031] In the depicted example, the collimator 114 shapes or limits the X-ray beam 116 that enters the region in which the patient / subject 118 is positioned. In the depicted example, the X-rays 116 are collimated as a cone beam that passes through the imaging volume, i.e., a cone beam. A portion of the X-ray radiation 120 passes through or around the patient / subject 118 (or other subject of interest) and impinges on a detector array, generally indicated at 122. The detector elements of the array produce electrical signals representative of the intensity of the incident X-rays 120. These signals are acquired and processed to reconstruct an image of the features within the patient / subject 118.

[0032] The source 112 is controlled by a system controller 124 that provides both power and control signals for a CT examination sequence, including acquisition of two-dimensional localizers or scout images to identify anatomical structures of interest within a patient / subject for subsequent scan protocols. In the depicted embodiment, the system controller 124 controls the source 112 via an X-ray controller 126, which can be a component of the system controller 124. In such embodiments, the X-ray controller 126 can be configured to provide power and timing signals to the X-ray source 112.

[0033] Further, the detector 122 is coupled to the system controller 124, which controls acquisition of the signals generated in the detector 122. In the depicted embodiment, the system controller 124 acquires the signals generated by the detector using a data acquisition system 128. The data acquisition system 128 receives data collected by the readout electronics of the detector 122. The data acquisition system 128 can receive sampled analog signals from the detector 122 and convert the data to digital signals for subsequent processing by a processor 130 discussed below. Alternatively, in other embodiments, the digital-to-analog conversion can be performed by circuitry disposed on the detector 122 itself. The system controller 124 can also perform various signal processing and filtering functions with respect to the acquired image signals, such as initial adjustment for dynamic range, digital image data interleaving, etc.

[0034] In Figure 2In the illustrated embodiment, the system controller 124 is coupled to a rotational subsystem 132 and a linear positioning subsystem 134. The rotational subsystem 132 enables the X-ray source 112, collimator 114, and detector 122 to be rotated one or more revolutions around the patient / subject 118, such as in an x, y plane about the patient. It should be noted that the rotational subsystem 132 can include a gantry or C-arm on which the respective X-ray emitting and detecting components are disposed. Thus, in such embodiments, the system controller 124 can be used to operate the gantry or C-arm.

[0035] The linear positioning subsystem 134 can enable the patient / subject 118, or more specifically a table supporting the patient, to be displaced within the bore of the CT system 110, such as in the z direction relative to the rotation of the gantry. Thus, the table can be moved linearly (in a continuous or step-wise fashion) within the gantry to generate images of a particular region of the patient 118. In the depicted embodiment, the system controller 124 controls the movement of the rotational subsystem 132 and / or the linear positioning subsystem 134 via a motor controller 136.

[0036] In general, the system controller 124 commands the operation of the imaging system 110, such as via the operation of the source 112, detector 122, and positioning systems described above, to perform an examination protocol and process acquired data. For example, the system controller 124, via the above-described systems and controllers, can cause a gantry supporting the source 112 and detector 122 to rotate about a subject of interest so that X-ray attenuation data can be obtained in one or more views relative to the subject. In this context, the system controller 124 can also include signal processing circuitry, associated memory circuitry for storing programs and routines executed by the computer (one or more neural networks (e.g., a multi-channel sinogram correction network)) and configuration parameters, image data, etc.

[0037] In the depicted embodiment, image signals acquired and processed by the system controller 124 are provided to a processing component 130 for reconstructing images. The processing component 130 can be one or more general- or special-purpose microprocessors. Data acquired by the data acquisition system 128 can be transmitted directly to the processing component 130, or after being stored in a memory 138. Any type of memory suitable for storing data can be utilized by such exemplary system 110. For example, the memory 138 can include one or more optical, magnetic, and / or solid-state memory storage structures. Moreover, the memory 138 can be located at the acquisition system site and / or can include remote storage devices for storing data, processing parameters, and / or routines for tomographic image reconstruction and analysis, as described below.

[0038] The processing component 130 can be configured to receive commands and scan parameters from an operator via an operator workstation 140, which is typically equipped with a keyboard and / or other input devices. The operator can control the system 110 via the operator workstation 140. Thus, the operator can use the operator workstation 140 to observe the reconstructed image and / or otherwise operate the system 110. For example, a display 142 coupled to the operator workstation 140 can be used to observe the reconstructed image and control the imaging. Additionally, the images can also be printed by a printer 144, which can be coupled to the operator workstation 140.

[0039] Further, the processing component 130 and the operator workstation 140 can be coupled to other output devices, which can include standard or special computers monitors and associated processing circuitry. One or more operator workstations 140 can be further linked in the system for outputting system parameters, requesting examinations, viewing images, and the like. In general, the displays, printers, workstations, and similar devices provided within the system can be local to the data acquisition components, or can be remote therefrom, such as elsewhere within a facility or hospital, or at a completely different location linked to the image acquisition system via one or more configurable networks, such as the Internet, a virtual private network, and the like.

[0040] It is also noted that the operator workstation 140 can also be coupled to a picture archiving and communication system (PACS) 146. The PACS 146 can in turn be coupled to a remote client 148, a radiology information system (RIS), a hospital information system (HIS), or to an internal or external network, such that raw or processed image data can be accessed by others at different locations.

[0041] While the foregoing discussion has treated various exemplary components of the imaging system 110 separately, these various components can be provided within a common platform or in interconnected platforms. For example, the processing component 130, the memory 138, and the operator workstation 140 can be collectively provided as a general or special purpose computer or workstation configured to operate in accordance with aspects of the present disclosure. In such embodiments, the general or special purpose computer can be provided as a separate component relative to the data acquisition components of the system 110, or can be provided in a common platform with such components. Likewise, the system controller 124 can be provided as part of such a computer or workstation, or as part of a separate system dedicated to image acquisition.

[0042] As discussed herein, Figure 2The system 110 of FIG. 1 can be used to perform a CT scan by measuring a series of views or projections from a number of different angles around a patient 118 or subject. Each view has a lateral dimension and a longitudinal dimension that respectively correspond to the number of columns and rows of the CT detector 122. The projections acquired at different viewing angles can be combined into a sinogram, which collects multiple views into a single dataset. The sinogram represents the spatial distribution of x-ray attenuation coefficients within the patient. Typically, the sinogram represents the spatial distribution of x-ray attenuation coefficients over a full rotation of the CT gantry (e.g., at a single axial location). A reconstruction algorithm processes the sinogram to produce a spatial domain image representing the patient 118 or subject.

[0043] As noted above, the CT detector can have one or more bad pixels that can cause artifacts in the reconstructed image. Figure 3 is a schematic depicting training a neural network to correct bad pixels. To generate training data for training a neural network (multi-channel sinogram correction network) 151, a measured sinogram 150 (a good sinogram that does not lack any values of any pixels) is obtained with the CT imaging system described above. From the measured sinogram 150, a simulated sinogram 152 with bad pixels is generated that simulates bad pixels 154. For example, a bad pixel mask 156 is utilized to simulate bad pixels at random locations within the measured sinogram 150. The bad pixels are represented by one or more lines 158 in the simulated sinogram 152 that lack any pixel values for a particular detector pixel. One or more bad pixels can be simulated in the simulated sinogram 152. The simulated bad pixels can be in separate locations or grouped together in a particular location. In certain embodiments, the simulated bad pixels can correspond to a central region (isocenter) of the CT detector. The bad pixel mask 156 and the corresponding simulated sinogram 152 (e.g., as a patch) are provided as input to the multi-channel sinogram correction network 151. The network 151 learns to predict the missing pixel values in the simulated sinogram 152. In particular, the network 151 learns to predict the missing pixel values from complementary information available in the sinogram 152. The complementary information is fed to the network 151 in the form of multi-channel input data (e.g., multi-channel two-dimensional (2D) or three-dimensional (3D) patches), as described in more detail below. In the case of utilizing multiple views, 3D patches can be utilized. The complementary information can include local neighboring sinogram information and conjugate sinogram information (e.g., information from a conjugate region relative to the bad pixel location (e.g., along the CT detector 180 degrees from the bad pixel location)).

[0044] The corrected (e.g., estimated or predicted) sinogram (e.g., as patches) 160 can be output by the network 151. The corrected sinogram or sinogram patches 160 can be compared to the measured sinogram or sinogram patches 150 (which serve as ground truth) to determine a training loss in the sinogram domain 161. The sinogram domain loss 161 can be in the form of an MAE loss. The sinogram loss 161 can be in the form of a content loss (e.g., LI and / or L2 loss computed between the target and predicted sinogram). In certain embodiments, the sinogram 152 can be provided to the network 151 as the raw sinogram. In certain embodiments, the sinogram 152 can be transformed (e.g., filtered) before being provided to the network 151. The correction can then be performed in a transformed domain (e.g., wavelet domain) before being finally transformed back to the normal or native sinogram domain and output. Processing the sinogram in the transformed domain highlights the fine features that drive the training and enhance the sinogram domain loss.

[0045] A tomographic image or volume 162 is generated from the corrected sinogram 160 via reconstruction 164. Additionally, a tomographic image or volume is generated from the measured sinogram 150 via reconstruction 166. Patches of the tomographic image or volume 162 are compared to patches of the tomographic image or volume generated from the measured sinogram 150 (which serve as ground truth) to determine a training loss in the image reconstruction domain 168. The image reconstruction domain loss 168 can be in the form of an MSE loss or SSIM. Other losses can be utilized. For example, a perceptual loss (e.g., SSIM loss computed between the sinogram and the reconstructed domain) can be utilized. A transformed domain loss (e.g., a loss that can be computed by a filtering domain / wavelet domain) can be utilized in the sinogram domain or the image reconstruction domain. Furthermore, an adversarial loss can also be used in the training.

[0046] The training weights (network weights) are updated (as indicated by reference number 170) at least via the sinogram domain loss 161. In certain embodiments, the training weights 170 are updated via a dual domain loss function that utilizes both the sinogram domain loss 161 and the image reconstruction domain loss 168. In certain embodiments, a single correction network utilizes the dual domain loss function. In other embodiments, separate correction networks can be utilized in a serial fashion. For example, a first network (e.g., a sinogram domain correction network) can be utilized that corrects sinograms with poor pixel data based on a loss defined in the native (raw) domain or a transformed domain of the sinogram. Then, a second network (e.g., an image reconstruction domain correction network) can be utilized that corrects any perceptual artifacts in the reconstructed tomographic image derived from the corrected sinogram based on a reconstruction domain loss to improve the final image.

[0047] As described above, the multi-channel sinogram correction network 151 learns to predict missing pixel values from available complementary information in sinograms having missing pixel values. Figure 4 is a schematic depicting a multi-channel sinogram correction network 151 that is trained and ultimately used to correct predicted missing pixel values in sinograms having missing data due to bad pixels. As depicted, the network 151 can include multiple networks (e.g., feature networks 172, 174, 176; fusion network 184; blending network 186). Complementary information is fed to the network 151 in the form of multi-channel input data (e.g., multi-channel 2D or 3D patches). In certain embodiments, some of these networks can be 2D-CNN networks.

[0048] An input patch 178 is derived from a sinogram having missing pixel values of at least one bad pixel. The input patch 178 is a local neighboring patch from the bad sinogram (e.g., corresponding to a portion along a channel view direction of a row having a bad pixel). The local neighboring patch utilizes spatial correlation and views within neighboring channels. The input patch 178 is input into a feature network 172.

[0049] A neighboring patch 180 (e.g., a neighboring row patch) is derived from a portion of the bad sinogram corresponding to a neighboring row of the row having the bad pixel. The neighboring row patch utilizes neighboring sensor correlation in the z-direction. The neighboring patch 180 is input into a feature network 174.

[0050] A conjugate patch 182 is derived from a conjugate region of the bad sinogram relative to the bad pixel. For example, the conjugate region of the sinogram can contain data acquired at a pixel position that is 180 degrees from the bad pixel position along the CT detector. The conjugate patch 182 utilizes complementary information available due to CT geometry. The conjugate patch 182 is input into a feature network 176.

[0051] Data output from the feature network 176 and the feature network 172 (e.g., via data concatenation) are combined in a fusion network 184. It should be noted that the feature networks 172, 176 and the fusion network 184 can utilize deep residual learning in learning to estimate pixel values of missing pixel values in the bad sinogram due to the bad pixel. The output of the fusion network 184 is used in a blending network 186 along with outputs from the feature networks 172, 174 to generate an output patch 188 for a corrected sinogram. In certain embodiments, the blending network 186 can generate the output patch via mask addition (e.g., generating a mask including pixel values of the patch applied to the missing pixel values).

[0052] In addition to the complementary information provided via patches 178, 180, 182, there are other inputs that provide complementary information for training the multi-channel sinogram correction network 151. For example, in the case of dual-energy CT scans, the second energy scan is used as an input channel. In addition, alternative patches including different definitions based on user-defined similarity of neighborhoods can be utilized.

[0053] As described above, training the neural network involves determining a training loss in the image reconstruction domain. Figure 5 is a schematic diagram depicting a reconstruction domain analysis for determining an image reconstruction domain loss. A reconstructed image 190 derived from a good sinogram (i.e., not having missing pixel values due to bad pixels) is used as a ground truth. A reconstructed image 192 is derived from a corrected sinogram (e.g., an estimated sinogram having estimated or predicted pixel values for bad pixels) that is output from the sinogram correction network. A reconstructed image 194 derived from an uncorrected or bad sinogram (e.g., having missing pixel values) has ring artifacts. The reconstructed image 194 has ring artifacts. The reconstructed image 192 is compared to the ground truth (reconstructed image 190) to determine a reconstruction difference 196 (e.g., an image reconstruction domain training loss). In certain embodiments, patches of images 190, 192 can be compared to determine the image reconstruction domain training loss. In certain embodiments, the image domain can be converted (e.g., filtered) to a different domain prior to performing the reconstruction domain loss analysis. As described above, this training loss 196 is used to update the training weights used in training the sinogram correction network as part of a dual-domain or hybrid-domain loss. In certain embodiments, a separate network (e.g., separate from the sinogram correction network) can be utilized (e.g., in a serial manner) to correct any artifacts (e.g., ring artifacts) that remain in the reconstructed image generated from the corrected sinogram. The separate network can utilize the image reconstruction domain loss to learn.

[0054] Figure 6This is a schematic diagram depicting the utilization of a trained sinogram correction network. A trained multichannel sinogram correction network or model 151 is configured to receive a faulty sinogram 198 (i.e., a sinogram with missing pixel values ​​due to faulty pixels) and output a corrected sinogram 200. Pixel values ​​are inferred (e.g., predicted or estimated) for any missing pixel values ​​in the faulty sinogram 198. Specifically, network 151 uses training weights 202 to infer missing pixel values. Training weights 202 may have been adjusted at least by the sinogram domain training loss as described above. In some embodiments, both the sinogram domain training loss and the image reconstruction domain loss are used to update the training weights. Multichannel data (e.g., 2D or 3D patches) is input into the network. Correction occurs only in the sinogram domain. Network 151 relies only on the input patches in the sinogram domain. No entry into the reconstruction domain is required. Therefore, the transformation is independent of anatomical structures, display or reconstructed fields of view, and reconstruction parameters (including the reconstruction kernel). Since the process is independent of reconstruction parameters, no tuning parameters are required. As described above, network 151 can operate in the original sinogram domain or in a transformed (e.g., filtered) domain of the sinogram. With inference occurring entirely in the sinogram domain, the network achieves faster predictions.

[0055] In some implementations, as described above, separate correction networks can be utilized in a serial manner. For example, a first network (e.g., a sinogram domain correction network) that corrects a sine wave with poor pixel data can be used, followed by a second network (e.g., an image reconstruction domain correction network) that corrects any perceptual artifacts (ring artifacts or stripe artifacts) in the reconstructed tomographic image derived from the corrected sine wave to improve the final image.

[0056] Training was conducted using data from 40 tests. Figure 6 The trained network 151 resulted in approximately 500,000 multi-channel training patches (neighboring, row, conjugate). To obtain the ground truth, patches were obtained from a good sine wave (missing data with bad pixels and reconstructed images based on the good sine wave). Training was performed using 32 out of 40 tests, and validation was performed using 8 tests. Training parameters included: 1000 epochs, 2x10⁻¹⁰... -4 The learning rate and 5x10 -4 *0.1 时期数 / 100 The LR decay function and the loss function of MAE+SSIM.

[0057] Figure 7Different types of sinograms and corresponding image reconstructions are depicted. Sinograms 204, 206, 208 represent a correct sinogram (i.e., without any missing values due to bad pixels), a bad sinogram with missing pixel values as indicated by line 209 due to bad pixels, and a corrected sinogram output with the deep learning based sinogram correction model as described above. As depicted in sinogram 208, line 209 in sinogram 206 is no longer present as the pixel value for that missing pixel value has been estimated. Reconstructed images 210, 212, and 214 represent the corresponding images derived from sinograms 204, 206, and 208. As depicted in image 212, due to the missing data from the bad pixels, there are ring and streak artifacts present. As depicted in image 214, there are no ring and streak artifacts present and image 214 looks similar to image 210.

[0058] Technical effects of the disclosed subject matter include providing a deep learning based technique for correcting missing pixel values due to one or more bad pixels in a CT detector. The missing pixel values are predicted in a trained network with multi-channel input data in a sinogram that has complementary information. Specifically, the prediction of missing pixel values occurs entirely in the sinogram domain to provide faster predictions. The correction occurs only in the sinogram domain such that the process is independent of the anatomy, display field of view, and reconstruction parameters including the reconstruction kernel. By correcting bad pixels, the deep learning based technique provides high quality reconstructed images suitable for diagnostic purposes that are free of ring and streak artifacts due to bad detector pixels. The deep learning based technique enables relaxation of restrictions during production of CT detectors and utilization of CT detectors in field operations. Specifically, CT detectors with bad pixels in the central region of the detector or with multiple bad pixels typically on the detector can still be utilized. This can reduce service costs (e.g., associated with detector panel replacement) due to bad pixels.

[0059] This written description uses examples to disclose the application, including the best mode, and also to enable any person skilled in the art to practice the application, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the application is defined by the claims, and can include other examples that occur to those skilled in the art. Such other examples are intended to fall within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they include equivalent

Claims

1. A computer-implemented method for correcting artifacts in computed tomography (CT) data from a CT system, the computer-implemented method comprising: obtaining the CT data from a CT scan based on a number of different angles around a patient or subject measuring a series of views or projections; combining a plurality of projections acquired at different viewing angles into a sinogram; inputting the sinogram into a trained sinogram correction network, wherein the sinogram is missing a pixel value of at least one pixel, wherein the sinogram is a combination of a plurality of projections acquired at different viewing angles; processing the sinogram via one or more layers of the trained sinogram correction network, wherein processing the sinogram includes deriving complementary information from the sinogram and estimating the pixel value of the at least one pixel based on the complementary information; and outputting a corrected sinogram with the estimated pixel value from the trained sinogram correction network.

2. The computer-implemented method of claim 1, wherein the complementary information comprises a multi-channel patch, and wherein the multi-channel patch comprises: a local neighboring patch from the sinogram corresponding to a portion along a channel view direction of a row having the at least one pixel, a neighboring row patch from the sinogram corresponding to a neighboring row of the row having the at least one pixel, a conjugate patch from the sinogram corresponding to a conjugate region relative to the at least one pixel, or dual or multi-energy computed tomography complementary information extracted from other energy channels.

3. The computer-implemented method of claim 1, comprising reconstructing an image from the corrected sinogram with reduced artifacts that occurred due to the missing pixel value of the at least one pixel.

4. The computer-implemented method of claim 1, wherein the sinogram is raw projection data as a measurement captured by a detector of an x-ray computed tomography system.

5. The computer-implemented method of claim 1, comprising training a neural network using supervised learning to generate the trained sinogram correction network, wherein training data for the supervised learning includes sinograms without any missing pixel values and corresponding sinograms with missing pixel values simulated from the sinograms without any missing pixel values.

6. A computer-implemented method for generating a trained neural network to estimate missing values in computed tomography (CT) data from a CT system, the computer-implemented method comprising: providing training data including sinograms and complementary information derived from the sinograms, wherein sinograms are combinations of a plurality of projections acquired at different viewing angles in a CT scan, wherein the sinograms include sinograms without any missing pixel values and corresponding sinograms with missing pixel values simulated from the sinograms without any missing pixel values; and training a neural network using the training data to correct sinograms with missing pixel values of at least one pixel based on a combined training loss derived from both a sinogram domain of the training data and an image reconstruction domain of images reconstructed from the training data.

7. The computer-implemented method of claim 6, wherein the complementary information includes a multi-channel patch derived from each of the sinograms having a missing pixel value.

8. The computer-implemented method of claim 7, wherein the multi-channel patch for a respective sinogram having missing pixel values comprises: a locally neighboring patch corresponding to a portion along a channel view direction of a row having the missing pixel value, a neighboring row patch corresponding to a neighboring row of the row having the missing pixel value, a conjugate patch corresponding to a conjugate region relative to the missing pixel value, or dual or multi-energy computed tomography complementary information extracted from other energy channels.

9. The computer-implemented method of claim 6, wherein training the neural network to correct the sinogram having the missing pixel value of the at least one pixel includes training the neural network to estimate the pixel value independent of an anatomy of a scanned subject, a display field of view, and reconstruction parameters used to reconstruct the sinogram into a reconstructed image.

10. A deep learning based sinogram correction system for correcting artifacts in computed tomography (CT) data from a CT system, the system comprising: a memory encoding processor executable routines; a processing component configured to access the memory and execute the processor executable routines, wherein the routines, when executed by the processing component, cause the processing component to: input a sinogram into a trained sinogram correction network, wherein the sinogram is a combination of a plurality of projections acquired at different view angles in a CT scan, wherein the sinogram is missing a pixel value of at least one pixel; process the sinogram via one or more layers of the trained sinogram correction network, wherein processing the sinogram includes deriving complementary information from the sinogram and estimating the pixel value of the at least one pixel based on the complementary information, wherein the complementary information includes a multi-channel patch, and the multi-channel patch includes: a locally neighboring patch from the sinogram corresponding to a portion along a channel view direction of a row having the at least one pixel, a neighboring row patch from the sinogram corresponding to a neighboring row of the row having the at least one pixel, and a conjugate patch from the sinogram corresponding to a conjugate region relative to the at least one pixel or dual or multi-energy computed tomography complementary information extracted from other energy channels; and output a corrected sinogram having the estimated pixel value from the trained sinogram correction network.

11. The system of claim 10, wherein the routines, when executed by the processing component, cause the processing component to train a neural network using supervised learning to generate the trained sinogram correction network, wherein training data for the supervised learning includes sinograms having no missing pixel values and corresponding sinograms having missing pixel values simulated from the sinograms having no missing pixel values, and wherein a training loss is derived only from a sinogram domain of the training data or from both a sinogram domain and an image reconstruction domain of images reconstructed from the training data.

Images