Spectroscopic x-ray material decomposition method

Abstract

A method for material decomposition of an object based on energy-spectral X-ray scan data and an application of a frequency splitting method. The method includes: using two parallel AI models to perform material decomposition analysis based on input energy-spectral X-ray data, wherein the models are configured such that one exhibits higher bias and lower variance (lower noise) than the other. The input energy-spectral X-ray data is fed into both models. The output material composition data from the low-bias model is low-pass filtered, and the output material composition data from the low-variance model is high-pass filtered. The outputs from the two models are linearly combined before or after filtering. Compared to the output generated using only one AI model, the resulting combined material decomposition data has both lower bias and lower noise.

Description

Technical Field

[0001] This invention relates to a method for material decomposition using spectral X-ray data. Background Technology

[0002] By using a multi-energy X-ray detector, it is possible to measure the energy or frequency spectrum of X-rays received at different locations across the X-ray detector during X-ray imaging of an object. This is known as spectral X-ray imaging. Spectral X-ray data allows for the identification and quantification of the material contained within the scanned object. It can be performed using data from conventional X-ray imaging or fluorescence fluoroscopy, or using data from computed tomography (CT) imaging.

[0003] Referring first to CT imaging, a conventional computed tomography (CT) scanner includes an X-ray tube mounted on a rotatable gantry opposite one or more integrating detectors. The X-ray tube rotates around an examination area located between the X-ray tube and the one or more detectors, emitting polychromatic radiation that passes through the examination area and the subject and / or object positioned within it. The one or more detectors detect the radiation passing through the examination area and generate signals or projection data indicative of the examination area and the subject and / or object positioned therein. Projection data may be, for example, raw detector data or a projection sine curve, the latter being a visual representation of the projection data captured by the detector(s). A reconstructor is also typically used to process the projection data and reconstruct a volumetric image of the subject or object.

[0004] X-ray spectral CT is an imaging modality that extends the capabilities of conventional CT systems by incorporating detectors capable of distinguishing different X-ray energies (such as energy-discriminating photon counting detectors or energy-integrating detectors). X-ray spectral CT allows for the decomposition of the material being scanned.

[0005] In material decomposition, a forward model can be used, which models the expected photon count or expected integrated photon energy in different photon energy windows (bins) at each pixel of the detector as a function of a specific set of materials (base) and its corresponding equivalent path length in the scanned material. To decompose the material composition of an object, this forward model can be reversed to determine the material path length of each in the base material based on the energy spectrum projection data.

[0006] One way to do this is to numerically invert the equation, for example, using statistical estimation algorithms such as maximum likelihood estimation (MLE). However, this can lead to oversimplification and fail to account for real word detector effects such as pulse stacking and charge sharing. Furthermore, the MLE algorithm requires significant computational resources due to the iterative optimization techniques involved in finding the solution.

[0007] An alternative approach is to use an AI-based method. Here, AI algorithms, such as machine learning algorithms, are trained using known material path lengths and corresponding photon counts or probe signals for each energy bin to learn the system's forward model inversion. This is done once for each material basis before scanning (offline). The advantage of such an approach is that AI model inference is computationally less expensive than applying the MLE method to the decomposition problem. However, AI-based methods can also lead to inaccurate results, for example, due to biases in the model algorithm.

[0008] Improved accuracy and reliability of AI-based methods for energy-dispersive X-ray material decomposition are expected. Summary of the Invention

[0009] The inventors have recognized that, for AI-based material decomposition methods, a known trade-off (bias-variance trade-off) must be made between the amount of noise and bias in the material path length estimated by the AI ​​model during network training. This results in a model with a certain amount of bias and noise. However, the expectation for a photon counting system for material quantization is to have a model with finite bias and finite noise.

[0010] More generally, in statistics and machine learning, the bias-variance tradeoff is a property of a set of predictive models where a model with lower bias in parameter estimation has higher variance across samples, and vice versa. Higher variance implies more noise but less bias. Higher bias implies less noise but potentially more systematic inaccuracy.

[0011] Bias error can be understood as the error arising from incorrect assumptions in the learning algorithm. High bias may cause the algorithm to miss the correlation between features and the target output (underfitting).

[0012] Variance is the error in the training set ranging from sensitivity to small fluctuations. High variance may cause the algorithm to model random noise in the training data instead of the expected output (overfitting).

[0013] Ideally, AI algorithm outputs should have both low noise and low bias.

[0014] This invention is defined by the claims.

[0015] According to an example of an aspect of the present invention, a method is provided for material decomposition of said object using energy-spectral X-ray data of said object, the method comprising:

[0016] Access data storage device,

[0017] The data storage device stores a first AI model and a second AI model. Each of the first and second AI models is configured to: receive energy-spectral X-ray data as model input data and generate material decomposition data as model output data.

[0018] The first AI model is configured to exhibit a lower bias and a higher variance than the second AI model.

[0019] Obtain input energy spectrum X-ray data;

[0020] The input energy spectrum X-ray data is provided as model input data to the first AI model, and then a low-pass filter is applied to the model output data of the first AI model to derive the first material decomposition data.

[0021] The input energy spectrum X-ray projection data is provided as model input data to the second AI model, and then a high-pass filter is applied to the model output data of the second AI model to derive the second material decomposition data;

[0022] The first and second material decomposition data are linearly combined to derive the third material decomposition data.

[0023] Embodiments of the present invention are based on the concept of using two trained AI models for material decomposition: one providing low bias but high noise in the output material decomposition data, and the other providing high bias but low noise. The outputs of the two models are then combined via a frequency splitting method to obtain a result with low bias and low noise. The frequency splitting method effectively involves selecting only lower frequency components from the low-bias (high-noise) AI model, selecting higher frequency components from the higher-bias (low-noise) model, and linearly combining them. This can be accomplished using filtering.

[0024] More specifically, the low-pass filter extracts or selects only the lower frequency or energy components of the output of the first AI model (with low bias), and the high-pass filter extracts or selects only the higher frequency or energy components of the output of the second AI model (with higher bias). The linear combination of both provides low-noise and low-bias output material decomposition data. This is based on the assumption that noise is primarily high spatial frequencies, while bias is primarily a low-frequency effect.

[0025] Preferably, the method includes applying a low-pass spatial filter to the model output data of the first AI model. This is preferably applied to at least one dimension of the output data. Low-pass spatial filtering means applying a low-pass filter to the spatial frequencies of the output material decomposition data.

[0026] Preferably, the method includes applying a high-pass spatial filter to the model output data of the second AI model. This is preferably applied in at least one dimension, as described above. High-pass spatial filtering means applying a high-pass filter to the spatial frequencies of the output material decomposition data.

[0027] For example, the model output data may include data values ​​extending in two or more dimensions. For example, the model output data may include material decomposition data values ​​for each pixel in a pixel array extending in at least two dimensions. For example, a low-pass spatial filter may be used to filter the spatial frequency components of the set of model output data in at least one spatial dimension of the data. For example, a high-pass spatial filter may be used to filter the spatial frequency components of the set of model output data in the same at least one spatial dimension of the data.

[0028] Filtered spatial frequency generally refers to, for example, an image being decomposed into periodic patterns (which is possible for any image), each periodic pattern being weighted using a weighting factor, and then the weighted periodic patterns being summed to obtain a modified image (the filtered image), where the modification depends on the weighting. In low-pass filtering, higher weights will be applied to the higher frequencies of the periodic patterns, and vice versa for high-pass filtering. In practice, this process can be replaced by a convolution operation applied directly to the image. This concept can be applied to both 2D and 3D images.

[0029] The model output data may include data values ​​for each of a set of pixel locations spanning, for example, a 2D or 3D grid, and can therefore be understood as an "image" of the material decomposition data. Thus, spatial filtering can be applied to the output data of the two models in a manner similar to applying spatial filtering to a standard image.

[0030] In practice, there are different options for how to implement spatial filtering of the model outputs of the first and second AI models. For example, in a non-limiting example, the model output data may include data values ​​for each of a set of materials (e.g., material path length values ​​or other values), data values ​​for each of a set of views, and data values ​​for each of a pixel (2D) array, as a corresponding 2D “image” dataset for each view and each material, wherein spatial filtering is applied to each of these 2D “image” datasets. As a variation on this example, a set of 3D image datasets may be formed for each material, each 3D dataset comprising a stack of the previously referenced 2D “images,” with a third dimension corresponding to a different view, wherein the spatial filtering is applied in the 3D domain, i.e., for example, in one or more dimensions of the 3D array of values, for each of the 3D datasets for each material. In this latter option, it will be understood that the third dimension is not necessarily a physical space dimension, and therefore the spatial filtering does not have to be applied along a physical space dimension, but can be applied across another dimension of the output data array to periodic patterns in the output data. Other options will be apparent to those skilled in the art.

[0031] The two distinct AI models can be generated by altering the training data supplied to the two models during training (e.g., one model is fed more noisy data than the other), or by changing the training process itself between the two models, for example, training one model using a different cost function configured to favor noise. Each of the AI ​​models may include one or more machine learning algorithms, such as one or more artificial neural networks.

[0032] To avoid ambiguity, it should be noted that the first AI model is configured to exhibit a lower bias and a higher variance in the output material decomposition data values ​​compared to the second AI model.

[0033] The method is widely applicable to a variety of specific applications. In particular, the type of input X-ray data used can vary (e.g., conventional X-ray or CT data), and the type of output material composition data generated can vary, such as material projection data or material image data.

[0034] Therefore, in different examples, the input energy spectrum X-ray data can be energy spectrum projection data (i.e., unreconstructed energy spectrum projection data, for example, in the form of photon count data from a set of different energy bins), or it can be energy spectrum image data (i.e., reconstructed energy spectrum projection data). The input energy spectrum X-ray data can include raw projection data such as that generated by an X-ray scanning device, or it can be pre-processed data, such as data reconstructed to form energy spectrum image data.

[0035] In different examples, the material decomposition data may be material projection data or material image data.

[0036] In some examples, the spectral X-ray data may include spectral computed tomography (CT) data for the object (including projection data for multiple projection angles). Alternatively, it may include conventional X-ray or fluorescence fluoroscopy data captured from a single projection angle.

[0037] In different examples, the entire process can be completed in the projection data domain, or the entire process can be completed in the image data domain (by reconstructing the energy spectrum projection data into image data before feeding it into the two AI models), or part of the process can be completed in the projection domain and part in the image domain. For example, the AI ​​model can accept input energy spectrum projection data and output material projection data, while the frequency filtering part is completed in the image domain after the reconstruction of the material projection data. These various options will be discussed in more detail below.

[0038] The first and second AI models can be machine learning models trained using a supervised learning process.

[0039] According to one or more embodiments, the first AI model can be trained based on training data T1, and the second AI model can be trained based on training data T2, wherein T1 is generated with reduced noise compared to T2. For example, the reduced noise in T1 can be achieved by preprocessing the training data to have reduced noise compared to T2. For example, the first AI model is trained on training data T1, and the second AI model is trained on training data T2, wherein T1 is generated by adding noise suppression to T2. However, in other examples, noise may be artificially added to the training dataset T2. In other words, the first AI model is trained on training data T1, and the second AI model is trained on training data T2, wherein T2 is generated by adding noise to T1. For example, a random noise generator can be used to simulate the added noise. In another example, one or both of T1 and T2 can be generated by simulation, i.e., using an X-ray simulation procedure. Within the simulation, the noise level of the obtained data can be arbitrarily adjusted so that one or both of T1 and T2 can be obtained in this way.

[0040] In this method, the two AI models can start as two copies of the same machine learning network or algorithm, but are trained with separate and different training data to achieve differences in their respective biases and variances.

[0041] As an alternative to this method, the learning algorithms included by the first and second machine learning models can be configured to follow different learning or training strategies, one favoring lower bias (higher noise) and the other favoring lower noise (higher bias).

[0042] For example, by configuring the first model using different cost functions, the first model can be configured to have reduced bias, such that the cost function of the second model gives more noise, while the cost function of the first model gives less noise.

[0043] According to one or more embodiments, each of the first AI model and the second AI model is configured to receive energy spectrum X-ray projection data as model input data and generate material composition projection data as model output data. Therefore, the material decomposition is performed in the projection data domain.

[0044] The method may be, for example, a base material decomposition method, wherein the output material decomposition data includes a set of derived radiation path lengths l through the scanned material for each of a set of base materials having a known attenuation coefficient. s Here, the output material decomposition data includes material projection data. In other words, it is, for example, the unreconstructed material decomposition data, as opposed to image data.

[0045] When the input energy-spectral X-ray data is energy-spectral CT data, the output material decomposition data may include a path length sine curve vector. The path length sine vector includes the derived base material path length for each of the set of base materials, the path length sine vector for the radiation arriving at the detector for each of the set of pixels p, and for each of the set of projection views v of the input energy spectrum X-ray data.

[0046] A sine wave is a visual representation of the raw data obtained during computed tomography (CT) operations. A path length sine wave is a visual representation of the material-based path length for each pixel of the detector.

[0047] According to one or more embodiments, the method may further include applying an image reconstruction operation to the first and second material decomposition data before or after a linear combination of the first and second material decomposition data, such that the output material decomposition data includes output material image data.

[0048] Material image data refers to a visual or graphical representation of the material breakdown information. For example, material image data may include a set of images, each representing a material. Another option is to use color coding in a single image to represent different materials in the imaging area (e.g., each material is represented using a dedicated color).

[0049] According to one or more embodiments, the input energy spectrum X-ray data may include projection data, such as energy spectrum photon counting data. s It indicates the number of X-ray photons detected in each of the multiple energy bins at the X-ray detector.

[0050] According to another set of embodiments, each of the first AI model and the second AI model can be configured to receive energy spectrum X-ray image data as model input data and generate material decomposition image data as model output data. Therefore, here, the material decomposition is performed in the image domain rather than the projection domain.

[0051] As mentioned above, according to one or more embodiments, the input spectral X-ray data can be spectral CT data. In different examples, this can be spectral CT projection data or spectral CT image data. As an alternative to CT data, the input spectral X-ray data can be conventional spectral X-ray data.

[0052] According to this set of examples, the input energy spectrum X-ray data may include energy spectrum photon count data c s(p,v) indicates the number of X-ray photons detected in each of the multiple energy bins at the X-ray detector, and for each of the multiple pixels p and the multiple projection views v for the input energy spectrum X-ray data.

[0053] An example of another aspect of the invention provides a computer program product comprising computer program code executable on a processor or computer, wherein the code is configured to cause the processor to perform a method according to any example or embodiment outlined above or described below, or according to any claim of this application.

[0054] An example of another aspect of the present invention provides a processing apparatus for processing energy-spectral X-ray data to derive material decomposition information, the processing apparatus being configured to:

[0055] Access data storage device,

[0056] The data storage device stores a first AI model and a second AI model. Each of the first and second AI models is configured to receive energy-spectral X-ray data as model input data and generate material decomposition data as model output data.

[0057] The first AI model is configured to exhibit lower bias and higher variance than the second AI model.

[0058] Obtain input energy spectrum X-ray data;

[0059] The input energy spectrum X-ray data is provided as model input data to the first AI model, and then a low-pass filter is applied to the model output data of the first AI model to derive the first material decomposition data.

[0060] The input energy spectrum X-ray data is provided as model input data to the second AI model, and then a high-pass filter is applied to the model output data of the second AI model to derive the second material decomposition data;

[0061] The first and second material decomposition data are linearly combined to derive the third material decomposition data.

[0062] According to some embodiments, the processing apparatus may include a data storage device for storing first and second AI models.

[0063] An example of another aspect of the present invention provides an X-ray imaging system comprising:

[0064] An X-ray scanning assembly includes an X-ray radiation source and an X-ray radiation detector, with a scanning area between them for receiving an object to be scanned, wherein an X-ray radiation path between the source and the detector passes through the scanning area; and a processing device according to any example or embodiment outlined above or described below, or according to any claim of this application. The processing device is communicatively coupled to the X-ray radiation detector to receive input energy spectrum X-ray data from the X-ray detector.

[0065] The scanner assembly can be a CT scanning assembly. Therefore, it can be arranged such that the X-ray source can be angularly positioned relative to the X-ray detector.

[0066] The X-ray detector may be an X-ray photon counting detector suitable for detecting photon counts for each of a set of radiation frequencies or energy bins.

[0067] An example of another aspect of the invention provides a method for training first and second AI models for use in deriving material decomposition data from input energy-spectral X-ray data. The method includes:

[0068] A preliminary AI model is obtained, which is configured to obtain energy spectrum X-ray data as model input data and generate material decomposition data as model output data.

[0069] The first AI model is obtained by training a first copy of the initial AI model using the first training data T1.

[0070] The second AI model is obtained by training a second copy of the initial AI model using the second training data T2.

[0071] Each of the first and second training data includes an input data entry in the form of sample energy spectrum X-ray data, and a corresponding output data entry in the form of known material decomposition data for each sample energy spectrum X-ray data in the sample energy spectrum X-ray data.

[0072] The first training data T1 is generated or preprocessed to have reduced noise in the energy spectrum X-ray data that forms its input data entries, such that the first training data has reduced noise compared to the second training data.

[0073] These and other aspects of the invention will be apparent and set forth with reference to one or more embodiments described below. Attached Figure Description

[0074] To better understand the invention and to more clearly illustrate how it can be implemented, reference will now be made to the accompanying drawings by way of example only, wherein:

[0075] Figure 1 The steps of an example method according to one or more embodiments are summarized in the form of a block diagram;

[0076] Figure 2 The illustration shows comparative results of material decomposition obtained using embodiments of the present invention;

[0077] Figure 3 An example processing apparatus for implementing a method according to one or more embodiments of the present invention is shown;

[0078] Figure 4 An example X-ray system for material decomposition of the scanned object is schematically illustrated; and

[0079] Figure 5 The steps for example methods of training two AI models for use in material decomposition based on energy-spectral X-ray data are outlined. Detailed Implementation

[0080] The invention will be described with reference to the accompanying drawings.

[0081] It should be understood that the detailed descriptions and specific examples are intended for illustrative purposes only and are not intended to limit the scope of the invention when indicating exemplary embodiments of the apparatus, system, and method. These and other features, aspects, and advantages of the apparatus, system, and method of the invention will become better understood from the following description, claims, and drawings. It should be understood that the drawings are merely schematic and not drawn to scale. It should also be understood that the same reference numerals are used throughout the drawings to indicate the same or similar parts.

[0082] Embodiments of the present invention provide a method for material decomposition of an object based on energy-spectral X-ray scan data and by applying an AI model and a frequency splitting method. The method includes: performing material decomposition using two parallel AI models based on input energy-spectral X-ray data, wherein the models are configured such that one exhibits higher bias and lower variance (lower noise) than the other. The input energy-spectral X-ray data is fed into both models. The output material composition data from the low-bias model is low-pass filtered, while the output material composition data from the low-variance model is high-pass filtered. The outputs from the two models are linearly combined. Compared to the output generated using only one AI model, the resulting combined material decomposition data has both lower bias and lower noise.

[0083] By using a multi-energy X-ray detector, it is possible to measure the energy spectrum of X-rays received at different locations across the X-ray detector during X-ray imaging of an object. This is known as spectral X-ray imaging. Spectral X-ray data allows for the identification and quantification of the material contained within the scanned object. It can be performed using data from conventional X-ray imaging modalities or from CT imaging.

[0084] Material decomposition can be achieved through pixel-by-pixel reversal of a system's forward model, which describes the expected theoretical signal in an individual X-ray detector's energy bin as a function of the specific set of materials (base) included in the object and their corresponding equivalent path lengths. This process is called base material decomposition.

[0085] Material decomposition using spectral photon counting X-ray data is a nonlinear, ill-posed, and computationally expensive problem. Using AI models for decomposition can significantly reduce computation time, but the ill-posed nature of the problem can still lead to strong noise amplification and bias in the AI ​​model output.

[0086] According to embodiments of the present invention, a decomposition method using two or more AI models is proposed to reduce bias and noise. It utilizes the concept of frequency splitting, based on the assumption that noise has high spatial frequencies in both the projection and image domains, while bias is a low-frequency effect.

[0087] The basic principles of energy-spectral X-ray-based material decomposition, as applied in embodiments of the present invention, will now be briefly discussed. A more detailed discussion of the underlying principles of the material decomposition applied herein can also be found in the article: KC Zimmerman and TGSchmidt, Experimental comparison of empirical material decomposition methods for spectral CT (Physics in Medicine & Biology, 60 (2015) 8).

[0088] When an X-ray photon strikes a photon counting detector, the photon is converted into a charge proportional to the energy of the incident photon. This charge can be converted into a voltage using a charge integrating amplifier. An analog comparator can be used to increment a digital counter when the accumulated charge voltage exceeds a set threshold level. At the end of the acquisition, the number of photons detected with energies above the threshold can be counted. Based on the subtraction of the continuous counter measurements, energy bin data corresponding to the number of photons detected between two threshold levels can be generated. In this way, photon counts for multiple X-ray energy bins can be generated.

[0089] It should be noted that, for the purposes of the following description, it will be assumed that an energy spectrum photon counting detector is used. However, other types of detectors are also possible, such as an energy spectrum energy integration detector (EID) capable of detecting radiation within a specific window of the energy spectrum of an X-ray source.

[0090] Consider the case of X-ray measurements of a material with thickness x and attenuation coefficient μ(E), where E is the energy of a photon passing through the material. The X-ray attenuation of the composite material can be expressed as a linear combination of the attenuations of the set of base materials included in the material. More specifically, and as explained in KCZimmerman (2015), the X-ray attenuation through this material is equivalent to the attenuation of any unique combination of two other materials (in the absence of a K-edge), as shown in the following equation (1), where μ1(E) and μ2(E) are the energy-dependent attenuation coefficients of each of the two base materials, and l1 and l2 are the path lengths of each of the two base materials. This decomposition is possible because there are two main attenuation phenomena in the diagnostic X-ray energy range: Compton scattering and photoelectric absorption.

[0091] xμ(E)=l1μ1(E)+l2μ2(E) (1)

[0092] This basic extension can be extended to make the i-th energy bin c of an ideal photon counting detector... i The expected number of photons detected can be modeled as follows:

[0093]

[0094] Where S(E) is the X-ray source energy spectrum, and l j An element of l has a decay coefficient function μ. j The thicknesses of the M basic materials. The energy spectrum measurement results are represented as a vector of detected photon counts, c = [c1, c2, ..., c K ] T , where K is the total number of energy measurement results.

[0095] Generally, material decomposition involves estimating the thickness l of the underlying material based on the acquired energy spectrum data c. One approach to estimating the thickness l of the underlying material based on the number of detected photons c is, for example, to numerically invert equation (2) using a statistical estimation algorithm such as maximum likelihood estimation (MLE). However, this can lead to oversimplification and fail to account for real-world word detector effects such as pulse stacking and charge sharing.

[0096] Another approach that can overcome some of these problems is based on artificial intelligence (AI) models. It is this approach that is used in embodiments of the present invention.

[0097] According to embodiments of the invention, in order to perform base material decomposition in the projection domain, machine learning algorithms, such as neural networks, can be employed, which are trained using energy binning data from known thicknesses of the base material. For example, a database of various reference path length vectors l(p) for a given base material of interest is generated together with the resulting photon count data vector c(p), both of which are functions of the detector's pixel p.

[0098] The trained AI model f is configured to estimate the thickness of the underlying material of the scanned object based on energy-spectral X-ray measurements. This can be expressed as follows:

[0099]

[0100] in, This is the path length vector calculated for the basic material of pixel p, and w(p) is the weighted parameter vector of the AI ​​model f for pixel p. When the input energy spectrum data c is CT data, the calculated path length data can be a sine wave vector. It is a function of detector pixel p and the projected view v acquired during scanning.

[0101] The training of the AI ​​model f can be based on a database of reference path length vector l(p) and photon counts discussed above, and the AI ​​model can be trained on each pixel in the database to minimize the error Er between the predicted path length (the model's output) and the true path length.

[0102]

[0103] The above principles can be applied according to various embodiments of the present invention. Several embodiments of the present invention will now be discussed in more detail.

[0104] Figure 1 The steps of an example method according to one or more embodiments are outlined in block diagram form for material decomposition of an object using energy-spectral X-ray data of the object.

[0105] The method includes accessing 12 data storage devices. The data storage devices store a first AI model and a second AI model. Each AI model is configured to receive energy-spectral X-ray data as model input data and generate material decomposition data as model output data.

[0106] The first AI model was configured to exhibit lower bias than the second AI model. The second AI model was configured to exhibit higher variance than the second AI model.

[0107] The method also includes 14 obtaining input energy spectrum X-ray data. This can be input projection data or reconstructed energy spectrum image data. The input X-ray data can be obtained directly from the X-ray scanning apparatus, for example, for real-time processing. It can be obtained from a data storage device, for example, for offline processing. It can be obtained from a processing unit or unit configured to preprocess data, for example, from a reconstructor that receives raw projection data and outputs reconstructed energy spectrum image data. In this way, the form of the energy spectrum X-ray data received and supplied as input to two AI models can vary in different embodiments.

[0108] The method also includes providing input energy spectrum X-ray data as model input data to the first AI model.

[0109] The method also includes providing the input energy spectrum X-ray projection data as model input data to the second AI model.

[0110] The method also includes applying a low-pass filter to the model output data of the first AI model and applying a high-pass filter to the model output data of the second AI model.

[0111] For low-pass filtering, any linear low-pass filter can be used. For example, by way of non-limiting example, a moving average filter, a Gaussian window filter, a Hann window filter, a Hamming window filter, a Kaiser window filter, or any other suitable example low-pass filter as will be apparent to those skilled in the art.

[0112] For high-pass filtering, again, as will be apparent to those skilled in the art, any suitable high-pass filter can be used. By way of non-limiting example, high-pass filtering can be accomplished by subtracting a low-pass filter from a Dirac pulse, or, for example, by low-pass filtering the data and then subtracting the low-pass filtered data from the original data.

[0113] The method also includes linearly combining the outputs of the first and second AI models 22 to derive output material decomposition data. This linear combination can be performed after the model outputs have already been filtered.

[0114] The filtering and linear combination of the model outputs (as outlined above) can be referred to as a frequency splitting method or frequency splitting process. For the sake of brevity, this article will use this term.

[0115] Optionally, the method may also include a reconstruction step, comprising applying an image reconstruction operation to the outputs of the first AI model and the second AI model before or after their filtering and / or linear combinations thereof, such that the output material decomposition data includes the output material image data.

[0116] Material image data refers to a visual or graphical representation of material breakdown information. For example, material image data may include a set of images, each representing a material. Another option is to use color coding in a single image to represent different materials in the imaging area (e.g., each material is represented using a dedicated color).

[0117] In different embodiments of the invention, the AI ​​model can be configured to receive different forms of input energy spectrum X-ray data and generate different forms of output material composition data. For example, the input data can be projection data or it can be reconstructed image data, and the output material composition data from the model can be in the projection domain (material projection data) or it can be in the image domain (material image data).

[0118] The following is a brief overview of the four main advantageous options:

[0119] (1) The input energy-spectral X-ray data comes from a conventional X-ray modal system. Here, two AI models are used to decompose the acquired energy-spectral X-ray projection data to obtain two sets of output material projection data. These two sets of material projection data are combined with the frequency splitting step discussed above to obtain the final material projection data. This provides the output material decomposition data. In some examples, this can be used directly for visualization.

[0120] (2) The energy-spectral X-ray data is from a computed tomography (CT) system. The acquired energy-spectral X-ray projection data is provided as input to two AI models. The data is decomposed using the two AI models to derive two sets of output material projection data. The material projection data may, for example, include a vector of material path lengths in the scanned object from a base material. These two sets of material projection data are combined with the frequency splitting steps discussed above (filtering and linear combination) to derive the final material projection data. Optionally, this can then be processed using a reconstruction pipeline to derive the final material image data.

[0121] (3) The input energy-spectral X-ray data comes from a computed tomography (CT) system. The acquired energy-spectral X-ray projection data is provided as input to two AI models. The two AI models are used to decompose the data to derive two sets of material projection data. These two sets of material projection data are then reconstructed to derive two sets of material image data. These image datasets are combined with the frequency splitting steps (filtering and linear combination) discussed above to derive the final material image data.

[0122] (4) The energy-spectral X-ray data is from a computed tomography (CT) system. The acquired energy-spectral X-ray projection data is reconstructed to derive a set of energy-spectral image data as input energy-spectral X-ray data. This is provided as input data to two AI models. The energy-spectral image data is decomposed using the two AI models to derive two sets of material image data, which are then combined with the frequency splitting process (filtering and linear combination) discussed above to derive the final material image data.

[0123] As can be seen, options (2)-(4) differ in the order of the various processing steps (decomposition / frequency splitting / reconstruction). Option (1) does not have a reconstruction step.

[0124] To achieve the desired results, two AI models are required, both capable of material decomposition but differing in their exhibited bias and variance.

[0125] Each of the AI ​​models can include one or more machine learning algorithms.

[0126] A machine learning algorithm is any self-trained algorithm that processes input data to produce or predict output data. Here, the input data includes energy-spectral X-ray data, and the output data includes material decomposition data.

[0127] Suitable machine learning algorithms used in this invention will be apparent to those skilled in the art. Examples of suitable machine learning algorithms include decision tree algorithms and artificial neural networks. Other machine learning algorithms, such as multinomial regression, support vector regression, or kernel regression, are suitable alternatives.

[0128] The structure of artificial neural networks (or simply neural networks) is inspired by the human brain. A neural network consists of layers, each containing multiple neurons. Each neuron performs a mathematical operation. Specifically, each neuron can include different weighted combinations of a single type of transformation (e.g., transformations of the same type, sigmoid transformations, etc., but with different weights). In processing the input data, the mathematical operation of each neuron is performed on the input data to produce a numerical output, and the outputs of each layer in the neural network are sequentially fed into the next layer. The final layer provides the output.

[0129] Methods for training machine learning algorithms are well-known. Typically, such methods involve obtaining a training dataset, which includes training input data entries and corresponding training output data entries. An initialized machine learning algorithm is applied to each input data entry to generate a predicted output data entry. The error between the predicted output data entry and its corresponding training output data entry is used to modify the machine learning algorithm. This process can be repeated until the error converges and the predicted output data entry is sufficiently similar to the training output data entry (e.g., ±1%). This is often referred to as a supervised learning technique.

[0130] For example, in machine learning algorithms formed by neural networks, the mathematical operations (weighting) of each neuron can be modified until the error converges. Known methods for modifying neural networks include gradient descent, backpropagation, and others.

[0131] The training input data entries correspond to example input energy spectrum X-ray data for a given material basis (e.g., sample photon energy bin counts for the material basis). The training output data entries correspond to material decomposition data.

[0132] As discussed, in different embodiments of the invention, the AI ​​model can be configured to receive different forms of input X-ray data and generate different forms of output material composition data. For example, the input data can be projection data or it can be reconstructed image data, and the output material composition data from the model can be in the projection domain (material projection data) or it can be in the image domain (material image data).

[0133] To better understand the principles of the invention, examples will be described in more detail, wherein the input data of the AI ​​model is CT projection data, and the output material decomposition data from the AI ​​model is decomposition projection data. However, the same principle can be applied to models with input and output data in different forms.

[0134] In this example, the decomposition method is the basic material decomposition method. In this case, the final output material decomposition data includes a set of derived radiation path lengths l. s The method involves scanning a set of base materials, each with a known attenuation coefficient. The theory behind this method has been discussed in more detail above, and we invite the reader to refer to that explanation for further details.

[0135] The final output material decomposition data may include, for example, a path length sine vector. This includes the derived base material path length for each of a set of base materials, the path length sine vector for radiation reaching each of a set of pixels p, and for each of a set of projection views v of the input energy spectrum X-ray data.

[0136] As discussed above, the fundamental material decomposition in the projection domain requires effectively reversing the forward model of the system (e.g., in a form similar to equation (2) above), which models the expected photon count in the individual detector energy bins as a function of the corresponding fundamental and equivalent path length of each pixel element of the X-ray projection data. Machine learning algorithms, such as neural networks, can be trained to perform this reversal.

[0137] To train the AI ​​model, a reference database of various path length vectors l(p) for a given material of interest can be used, along with the obtained photon count data vector c(p), both of which are functions of pixel p. Then, the AI ​​model is trained for each pixel in the database such that the resulting AI model f minimizes the error Er between the predicted path length (the model's output) and the true path length, where w(p) is the parameter vector of the AI ​​model f for pixel p.

[0138]

[0139] The training database can be generated experimentally using a calibration phantom, or it can be generated, for example, based on performing a real system simulation. Once trained, the model can be applied to the photon count projection data cs(p,v) of each projected view v acquired during the scan to estimate the equivalent path length sine curve.

[0140]

[0141] As previously mentioned, the resulting trained model will aim for a trade-off between output path length bias and noise, leading to some levels of both phenomena. The aim of embodiments of the invention is to provide an overall method for achieving low bias and low noise.

[0142] Therefore, and as discussed, the first AI model is configured to exhibit lower bias and higher variance than the second AI model. In this context, bias error can be understood as the deviation of the average estimated output path length of the AI ​​model from the true path length. Variance error can be understood as the variability of the aforementioned model's output path length estimate relative to the true path length.

[0143] There are different ways to implement the first AI model, which exhibits lower bias and higher variance than the second AI model.

[0144] One approach is to initially generate two copies of the initial network f and train both copies on two different types of photon counting data: one with no noise and one with the full noise expected during a typical X-ray or CT scan. For example, two copies of the initial training dataset can be generated, and one of the datasets can then be preprocessed to reduce noise. The lower-noise data is used to train the first AI model, and the full-noise data is used to train the second AI model. As a result of training with the noise-free data, a first AI model with low bias and high noise (high variance) is achieved. On the other hand, training with noisy data generates a second AI model with high bias and low noise (low variance):

[0145] An alternative to achieving differences in bias and variance between two AI models is to use the same training data for both but change the training strategies employed by the two models. This involves altering the learning algorithms encoded in the models so that one model favors lower bias and the other favors lower noise (lower variance). For example, by configuring the first model with different cost functions, it can be configured to have reduced bias, causing the second model's cost function to output more noise, while the first model's cost function outputs less noise. Both models are then trained using fully noisy training data.

[0146] For example, according to one or more embodiments, the bias that occurs during the training of a given model can be calculated, and the absolute value of the bias is used as a term in the cost function. The greater the weight given to this term, the lower the bias of the trained model. Therefore, when training the first AI model, a higher weight can be given to this term in the cost function.

[0147] For example, if several (example) noisy implementations of photon counting are available as a reference within a training dataset of a given combination of materials, the bias can be calculated during training.

[0148] The bias and variance of an AI model (e.g., a neural network) are measurable and testable properties. Measuring the bias and variance of a given model, which is trained to generate output material decomposition information based on input energy-spectral X-ray data, can be accomplished empirically, for example, by scanning a model with inserts characterized by different known material concentrations (e.g., iodine or water) and generating corresponding energy-spectral X-ray data for the phantom.

[0149] By feeding this input energy-spectral X-ray data into the model under test, the bias can be measured as the deviation of the average concentration value measured in the corresponding region of interest (ROI) from the known concentration. Noise, or variance, can be quantified as the standard deviation of the values ​​measured in a homogeneous portion of each image of the material.

[0150] As discussed above, after the decomposition performed by the two AI models, a “frequency splitting” step is applied, which involves selectively filtering the outputs of the first and second AI models and linearly combining their results.

[0151] Specifically, in the context of the above-described example where the input energy spectral X-ray data is energy spectral CT projection data, it can include, after filtering the model output using linear low-pass (G) and matched linear high-pass (1-G) filters respectively, combining the individual path length sine plots generated by the application of models f1 and f2 respectively:

[0152]

[0153] The resulting combined sine vector Provide output material decomposition data, in this case in the form of output material projection data (represented as a sine vector of the material path length for each pixel and each projection view for a given material basis).

[0154] The resulting sine curve of combined path lengths It will then exhibit both low bias and low noise.

[0155] As illustrated in the diagram. Figure 2 Example simulation results of the material decomposition method according to an embodiment of the present invention are shown. The first column shows the actual material decomposition images of the iodine sample (upper row) and the water sample (lower row). The second, third, and fourth columns show image reconstructions of the material decomposition sinogram data generated by decomposition using only the first AI model f1, only the second AI model f2, and the fully combined decomposition method according to the present invention (including the frequency splitting method), respectively. As discussed, the first model f1 is the bias-optimal model (low bias), and the second model f2 is the noise-optimal model (low noise).

[0156] It can be seen that the combination of the two models (fourth column) produces the best results, where the bias and noise levels are comparable to those provided by the bias-optimal (f1) and noise-optimal (f2) models, respectively.

[0157] The above explanation has already outlined the training of the AI ​​model, which receives energy spectrum CT data (projection data) in the projection domain and also outputs material decomposition data in the projection domain, i.e., based on the path length vector for a given material basis.

[0158] However, the same principle can be applied in a similar way to training and using AI models that receive different forms of energy-spectral X-ray data and output different forms of decomposition information. To derive these models, it is simply necessary to appropriately select the training data such that the training input data entries have the correct modalities or forms, and the training output data entries are material composition data in the correct form.

[0159] For example, in various embodiments, the input spectral X-ray data provided to each AI model can be spectral computed tomography (CT) data, as in the example summarized above, or in other embodiments it can be spectral X-ray data from a conventional (non-CT) X-ray or fluorescence fluoroscopy scanner. In the latter case, the input spectral X-ray data to the AI ​​model can be photon-counted projection data. s The photon count for each energy sub-bin is given in the form of (p), based on the pixels of the detector. Since there may only be one view for a conventional X-ray scanner, the photon count data may only be a function of the pixels. Alternatively, the conventional energy spectrum X-ray data can be reconstructed into image data before applying the AI ​​model, and the AI ​​model can be trained to receive the input energy spectrum image data and output material image information.

[0160] When the input energy-spectral X-ray data is energy-spectral CT data, the two AI models can be configured to receive either the input energy-spectral CT projection data or reconstructed energy-spectral CT image data reconstructed from the original energy-spectral projection data. In the former case, the AI ​​model is trained to output material decomposition data in the projection domain; in the latter case, the model is trained to output material composition data in the image domain.

[0161] An example of another aspect of the invention provides a computer program product comprising computer program code executable on a processor or computer, wherein the code is configured to cause the processor to perform a method according to any example or embodiment outlined above or described below, or according to any claim of this application.

[0162] Figure 3 Components of an example apparatus for performing a material decomposition method according to embodiments disclosed herein are shown.

[0163] In particular, aspects of the present invention provide a processing apparatus 32 for processing energy-spectral X-ray data to derive material decomposition information. In this example, the processing apparatus includes a single processor component, such as an integrated circuit (IC).

[0164] The processing device 32 is configured to access the data storage device 36. The data storage device stores a first AI model 42 and a second AI model 44. Each of the first and second AI models is configured to receive energy-spectral X-ray data as model input data and generate material decomposition data as model output data. The first AI model 42 is configured to exhibit lower bias and higher variance than the second AI model 44.

[0165] The processing device is configured to obtain input energy spectrum X-ray data, for example, from an X-ray scanning device communicatively coupled to the processing device or, in other examples, from another data storage device.

[0166] The processing device 32 is configured to provide input energy spectrum X-ray data as model input data to the first AI model 42, and then apply low-pass filtering to the model output data of the first AI model to derive first material decomposition data.

[0167] The processing device 32 is configured to provide input energy spectrum X-ray data as model input data to the second AI model 44, and then apply high-pass filtering to the model output data of the second AI model to derive second material decomposition data.

[0168] The processing device 32 is configured to linearly combine the first and second material decomposition data to derive the third material decomposition data.

[0169] According to one or more embodiments, the processing device 32 may include a data storage device 36 for storing the first 42 and the second 44 AI models. In other examples, it may include inputs / outputs for connection to the data storage device.

[0170] An example of another embodiment provides an energy spectrum X-ray system. Figure 4 Example system 60 is illustrated schematically.

[0171] The system includes an X-ray scanning assembly 62, which includes an X-ray radiation source 64 (X-ray generator) and an X-ray radiation detector 66, with a scanning area between them for receiving the object to be scanned, wherein an X-ray radiation path 68 between the source and the detector passes through the scanning area.

[0172] The system also includes a processing device 32 according to any example or embodiment outlined above or described below, or according to any claim of this application. The processing device is communicatively coupled to an X-ray radiation detector 66 to receive input energy spectrum X-ray data from the X-ray detector.

[0173] The X-ray scanning assembly 62 may be a CT scanning assembly, wherein the radiation source 64 and the radiation detector 66 have adjustable relative angular positioning, thereby enabling multiple projection views corresponding to different X-ray path angles through the scanning area to be captured.

[0174] In a preferred example, X-ray detector 66 is an X-ray photon counting detector adapted to detect photon counts for each of a set of radiation frequencies or energy bins.

[0175] Alternatively, the X-ray detector 66 may include one or more energy integrating detectors (EIDs). The EIDs are capable of selectively sensing incident X-ray radiation at a selected radiation window. In one possible arrangement, two or more EIDs are used in a stacked or bilayer configuration, each configured to be sensitive to a different energy window of the X-ray energy spectrum of an X-ray source (e.g., an X-ray tube).

[0176] In another possible setup, a single EID is used, wherein the X-ray source is controlled to cycle through two or more different X-ray tube voltages to sequentially emit radiation of different energy spectra. The single EID sequentially detects the radiation associated with each tube voltage. Thus, in this case, the energy level of the radiation is controlled at the source, rather than being discriminated against at the detector.

[0177] The basic architecture of a suitable X-ray scanning assembly for use in embodiments of the invention is described, for example, between page 3, line 26 and page 5, line 2 of WO2014 / 167450, and in that document... Figure 1 As shown in the diagram, the radiation detector array can take the form of, for example, a photon counting detector or one or more energy integration detectors.

[0178] Further details on the construction, operation, and implementation of suitable X-ray scanning components can be found in Katsuyuki Taguchi's (2020) book: "Spectral, Photon Counting Computed Tomography (Technology and Applications)".

[0179] Details about the key principles of suitable CT scanning components can also be found in the following paper: Willemink, M et al., “Photon-counting CT: Technical Principles and Clinical Prospects” (2018).

[0180] As discussed above, one approach to obtaining two AI models with different biases and variances is to train the two models using different training data.

[0181] Therefore, according to an example of another aspect of the invention, a method is provided for training first and second AI models for use in deriving material decomposition data from input energy spectrum X-ray data. Figure 5 The steps of such a method are outlined in the form of a block diagram.

[0182] The method involves obtaining an 82-preliminary AI model, which is configured to receive energy-spectral X-ray data as model input data and generate material decomposition data as model output data.

[0183] The method further includes obtaining first training data T1 and second training data T2, wherein each of the first and second training data includes input data entries in the form of sample energy spectrum X-ray data and corresponding output data entries in the form of known material decomposition data for each of the sample energy spectrum X-ray data. The first training data T1 is provided with reduced noise in the energy spectrum X-ray data forming its input data entries compared to T2, such that the first training data has reduced noise compared to the second training data. This can be achieved, for example, by preprocessing the first training data to reduce noise. For example, T1 can be generated by applying noise suppression to a second copy of T2. Alternatively, it can be achieved by preprocessing the second training data to have additional noise (compared to T1). In other words, noise may be artificially added to the training dataset T2. In other words, T2 is generated by adding noise to a second copy of T1. For example, a random noise generator can be used to simulate the added noise. In another example, one or both of T1 and T2 can be generated by simulation, i.e., using an X-ray simulation program. Within the simulation, the noise level of the acquired data can be adjusted arbitrarily, allowing one or both of T1 and T2 to be obtained in this way.

[0184] The method also includes training a first copy of the initial AI model using the first training data T1 to obtain the first AI model;

[0185] The method also includes training a second copy of the initial AI model using the second training data T2 to obtain a second AI model.

[0186] The result is that the first AI model obtained has less bias and greater variance than the second model.

[0187] Although in the example above, the two models start as copies of the same initial model, this is not necessary. In another embodiment, a first initial model can be trained using first training data T1 to obtain a first AI model, and a second initial model can be trained using second training data T2 to obtain a second AI model, wherein the first and second initial models can be different. For example, one initial model might be a neural network, and the other a linear regression model. Using different training data to train the initial models (as outlined above) will still result in different biases and variances between the two models.

[0188] As discussed above, an alternative way to achieve the difference in bias and variance between two AI models is to use the same training data for both, but change the training strategies adopted by the two models, that is, change the learning algorithms encoded in the models so that one model favors lower bias and the other favors lower noise (lower variance).

[0189] The embodiments of the present invention have wide applications in a variety of different fields. For example, the methods and systems can be used to process medical imaging data, such as medical CT data. For example, material decomposition in the medical field is often used to analyze the composition of tumors to determine whether tissue is cancerous. However, the embodiments can also be used outside the medical field, for example, for analyzing the material composition of inanimate objects.

[0190] As discussed above, the system utilizes a processor to perform data processing. A processor can be implemented in many ways using software and / or hardware to perform a variety of required functions. A processor typically employs one or more microprocessors that can be programmed using software (e.g., microcode) to perform the required functions. A processor can be implemented as a combination of dedicated hardware performing some functions and one or more programmable microprocessors and associated circuitry performing other functions.

[0191] Examples of circuits that can be employed in the various embodiments of this disclosure include, but are not limited to, conventional microprocessors, application-specific integrated circuits (ASICs), and field-programmable gate arrays (FPGAs).

[0192] In various implementations, the processor may be associated with one or more storage media, such as volatile and non-volatile computer memories, such as RAM, PROM, EPROM, and EEPROM. The storage media may be encoded with one or more programs that, when run on one or more processors and / or controllers, perform the required functions. The various storage media may be fixed within the processor or controller or may be transportable, such that one or more programs stored thereon can be loaded into the processor.

[0193] By studying the accompanying drawings, the disclosure, and the claims, those skilled in the art will be able to understand and implement other variations of the disclosed embodiments in practicing the claimed invention. In the claims, the word "comprising" does not exclude other elements or steps, and the words "a" or "an" do not exclude a plurality.

[0194] A single element or other unit may perform the functions of several items recorded in the claims.

[0195] Although specific measures are described in different dependent claims, this does not imply that combinations of these measures cannot be used advantageously.

[0196] Computer programs can be stored / distributed on suitable media, such as optical storage media or solid-state media provided with or as part of other hardware, but computer programs can also be distributed in other forms, such as via the Internet or other wired or wireless telecommunications systems.

[0197] If the term “suitable” is used in the claims or specification, it should be noted that the term “suitable” is intended to be equivalent to the term “configured as”.

[0198] Any reference numerals in the claims should not be construed as limiting the scope.

Claims

1. A method for material decomposition of said object using energy-spectral X-ray data of the object, the method comprising: Access data storage device, The data storage device stores a first AI model and a second AI model. Each of the first and second AI models is configured to: receive energy-spectral X-ray data as model input data and generate material decomposition data as model output data. The first AI model is configured to exhibit a lower bias in the output material decomposition data values ​​than the second AI model when applied to the same input dataset, and to exhibit a higher variance in the output material decomposition data values ​​when applied to the same input dataset. Obtain input energy spectrum X-ray data; The input energy spectrum X-ray data is provided as model input data to the first AI model; The input energy spectrum X-ray data is provided as model input data to the second AI model; Low-pass spatial filtering is applied to the model output data of the first AI model; High-pass spatial filtering is applied to the model output data of the second AI model; Specifically, the high-pass spatial filter is not applied to the model output data of the first AI model, and the low-pass spatial filter is not applied to the model output data of the second AI model; and The filtered outputs of the first AI model and the second AI model are linearly combined to derive the output material decomposition data. Wherein, the first AI model and the second AI model are artificial neural networks derived using a supervised training method, the supervised training method including: A preliminary AI model is obtained, which is configured to receive energy spectrum X-ray data as model input data and generate material decomposition data as model output data. The first AI model is obtained by training a first copy of the initial AI model using the first training data T1. The second AI model is obtained by training a second copy of the initial AI model using the second training data T2. Each of the first and second training data includes an input data entry in the form of sample energy spectrum X-ray data, and a corresponding output data entry in the form of known material decomposition data for each sample energy spectrum X-ray data in the sample energy spectrum X-ray data. The first training data T1 is generated from the second training data T2 by preprocessing the second training data T2 to reduce noise in the energy spectrum X-ray data that forms the input data entries of the first training data T1, so that the first training data has reduced noise compared to the second training data.

2. The method according to claim 1, wherein, Each of the first AI model and the second AI model is configured to receive energy spectrum X-ray projection data as model input data and generate material decomposition projection data as model output data.

3. The method according to claim 2, wherein, The method is a base material decomposition method, wherein the output material decomposition data includes a set of derived radiation path lengths l through the scanned material for each of a set of base materials having a known attenuation coefficient. s .

4. The method according to claim 3, wherein, The output material decomposition data includes a path length sine vector l̂(p, v), which includes the derived base material path length for each base material in the set of base materials, and the path length sine vector is for the radiation reaching each pixel in the set of pixels p and for each projection view in the set of projection views v of the input energy spectrum X-ray data.

5. The method according to any one of claims 2-4, wherein, The method further includes applying an image reconstruction operation to the first material decomposition data and the second material decomposition data before or after a linear combination of the first material decomposition data and the second material decomposition data, such that the output material decomposition data includes output material image data.

6. The method according to any one of claims 2-5, wherein, The input energy spectrum X-ray data is projection data and includes energy spectrum photon count data c. s The energy spectrum photon count data indicates the number of X-ray photons detected in each of the multiple energy sub-sub ...

7. The method according to claim 1, wherein, Each of the first AI model and the second AI model is configured to receive energy spectrum X-ray image data as model input data and generate material decomposition image data as model output data.

8. The method according to any one of claims 1-7, wherein, The input energy spectrum X-ray data is energy spectrum CT data.

9. The method according to claim 8, wherein, The input energy spectrum X-ray data includes energy spectrum photon count data c s (p, v), the energy spectrum photon count data indicates the detected X-ray photon count in each of the multiple energy sub-bins at the X-ray detector, and the energy spectrum photon count data is for each of the multiple pixels p, and for each of the multiple projection views v of the input CT projection data.

10. A computer program product comprising computer program code, said computer program code being executable on a processor. in, The code is configured to cause the processor to perform the method according to any one of claims 1-9.

11. A processing apparatus for processing energy-spectral X-ray data to derive material decomposition information, the processing apparatus being configured to: Access data storage device, in, The data storage device is arranged to store a first AI model and a second AI model. Each of the first and second AI models is configured to receive energy-spectral X-ray data as model input data and generate material decomposition data as model output data. The first AI model is configured to exhibit a lower bias in the output material decomposition data values ​​than the second AI model when applied to the same input dataset, and to exhibit a higher variance in the output material decomposition data values ​​when applied to the same input dataset. Obtain input energy spectrum X-ray data; The input energy spectrum X-ray data is provided as model input data to the first AI model; The input energy spectrum X-ray data is provided as model input data to the second AI model; Low-pass spatial filtering is applied to the model output data of the first AI model; High-pass spatial filtering is applied to the model output data of the second AI model; Specifically, the high-pass spatial filter is not applied to the model output data of the first AI model, and the low-pass spatial filter is not applied to the model output data of the second AI model; and The filtered outputs of the first AI model and the second AI model are linearly combined to derive the output material decomposition data; Wherein, the first AI model and the second AI model are artificial neural networks derived using a supervised training method, the supervised training method including: A preliminary AI model is obtained, which is configured to receive energy spectrum X-ray data as model input data and generate material decomposition data as model output data. The first AI model is obtained by training a first copy of the initial AI model using the first training data T1. The second AI model is obtained by training a second copy of the initial AI model using the second training data T2. Each of the first and second training data includes an input data entry in the form of sample energy spectrum X-ray data, and a corresponding output data entry in the form of known material decomposition data for each sample energy spectrum X-ray data in the sample energy spectrum X-ray data. The first training data T1 is generated from the second training data T2 by preprocessing the second training data T2 to reduce noise in the energy spectrum X-ray data that forms the input data entries of the first training data T1, so that the first training data has reduced noise compared to the second training data.

12. The processing apparatus according to claim 11, wherein, The processing apparatus includes the data storage apparatus for storing the first AI model and the second AI model.

13. An energy-spectral X-ray imaging system, comprising: An X-ray scanning assembly includes an X-ray radiation source and an X-ray radiation detector, with a scanning area between the X-ray radiation source and the X-ray radiation detector for receiving an object to be scanned, wherein the X-ray radiation path between the source and the detector passes through the scanning area; and The processing apparatus according to claim 11 or 12 is communicatively coupled to the X-ray radiation detector to receive input energy spectrum X-ray data from the X-ray radiation detector.

14. A method for training a first AI model and a second AI model, wherein the first AI model and the second AI model are used in the method according to any one of claims 1-9 to derive material decomposition data based on input energy-spectral X-ray data, wherein, The first AI model and the second AI model are artificial neural networks, and the method includes: A preliminary AI model is obtained, which is configured to receive energy spectrum X-ray data as model input data and generate material decomposition data as model output data. The first AI model is obtained by training a first copy of the initial AI model using the first training data T1. The second AI model is obtained by training a second copy of the initial AI model using the second training data T2. Each of the first and second training data includes an input data entry in the form of sample energy spectrum X-ray data, and a corresponding output data entry in the form of known material decomposition data for each sample energy spectrum X-ray data in the sample energy spectrum X-ray data. The first training data T1 is generated from the second training data T2 by preprocessing the second training data T2 to reduce noise in the energy spectrum X-ray data that forms the input data entries of the first training data T1, so that the first training data has reduced noise compared to the second training data.

Images