Spectral X-ray material fractionation method

Abstract

A method for material fractionation of an object based on spectral X-ray scan data of the object and based on the application of a frequency separation approach. The method includes using two AI models in parallel to perform material fractionation analysis based on input spectral X-ray data. The models are configured so that one exhibits higher bias and lower variance (lower noise) than the other. The input spectral X-ray data is fed to 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, either before or after filtering. The resulting combined material fractionation data has both lower bias and noise compared to the output produced using only one AI model.

Description

[Technical Field]

[0001] The present invention relates to a method for material fractionation using spectral X-ray data. [Background technology]

[0002] Multi-energy X-ray detectors can be used to measure the spectrum of X-ray energies or frequencies received at different locations on the X-ray detector during X-ray imaging of an object. This is known as spectral X-ray imaging. Spectral X-ray data can identify and quantify materials contained in the scanned object. This can be done using data from conventional X-ray imaging or fluoroscopy, or 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 rotating gantry facing one or more integrating detectors. The x-ray tube rotates around an examination region between the x-ray tube and the one or more detectors, emitting polychromatic radiation that traverses the examination region and the subject and / or object disposed therein. The one or more detectors detect the radiation traversing the examination region and generate signals, i.e., projection data, indicative of the examination region and the subject and / or object disposed therein. The projection data may be, for example, raw detector data or projection sinograms, the latter being a visual representation of the projection data captured by the detectors. A reconstructor is typically further 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 that can distinguish between different X-ray energies (e.g., energy-discriminating photon-counting detectors or energy-integrating detectors). X-ray spectral CT allows for material fractionation of scanned objects.

[0005] For material fractionation, a forward model can be used, which models the expected photon counts or expected integrated photon energy in different photon energy windows (bins) at each pixel of the detector as a function of a particular set of materials (basis) and their respective equivalent path lengths through the scanned materials. To fractionate the material composition of an object, this forward model can be inverted to determine the material path lengths for each of the basis materials based on the spectral projection data. Summary of the Invention [Problem to be solved by the invention]

[0006] One way to do this is to numerically invert the equation using a statistical estimation algorithm, such as the maximum likelihood estimation (MLE) algorithm. However, this leads to an oversimplification and fails to consider real-world detector effects such as pulse pile-up and charge sharing. Furthermore, MLE algorithms require significant computational resources due to the iterative optimization techniques involved in finding a solution.

[0007] Another approach is to use artificial intelligence-based methods. Here, an AI algorithm, such as a machine learning algorithm, is trained with known material path lengths and the corresponding photon counts or detection signals per energy bin to learn the system forward model inversion. This is done before a single scan (offline) per material basis. The advantage of this method is that AI model inference is less computationally intensive than applying an MLE approach to the fractionation problem. However, AI-based approaches can also lead to inaccurate results, for example, due to bias in the model algorithm.

[0008] It is desirable to improve the accuracy and reliability of AI-based approaches to spectral X-ray material fractionation.

[0009] The inventors recognize that AI-based approaches to material fractionation must make a known compromise (bias-variance tradeoff) during network training between the amount of noise and the amount of bias in the material path lengths estimated by the AI ​​model. This results in a model with a certain bias and a certain noise. However, for photon counting systems intended for material quantification, it is desirable to have a model with limited bias and noise.

[0010] More generally, in statistics and machine learning, the bias-variance tradeoff is a property of a set of predictive models: models with low bias in their parameter estimates have high variance of parameter estimates across samples, and vice versa. High variance means more noise but less bias. High bias means less noise but potentially more systematic inaccuracies.

[0011] Bias error can be understood as the error from incorrect assumptions in a learning algorithm. High bias can cause the algorithm to miss relevant relationships between features and target outputs (underfitting).

[0012] Variance is the error from sensitivity to small changes in the training set. High variance can cause the algorithm to model random noise in the training data rather than the desired output (overfitting).

[0013] Ideally, the output of an AI algorithm should have low noise and low bias. [Means for solving the problem]

[0014] The invention is defined by the claims.

[0015] According to an example in accordance with one aspect of the present invention, there is provided a method for material fractionation of an object using spectral X-ray data of the object, the method comprising: accessing a data storage device, a first AI model and a second AI model are stored in the data storage device, and each of the first and second AI models receives the X-ray spectrum data as model input data and generates material fractionation data as model output data; accessing, wherein the first AI model is configured to exhibit lower bias than the second AI model and higher variance than the second AI model; obtaining input spectral X-ray data; providing the input spectral X-ray data as model input data to a first AI model, and then applying low-pass filtering to model output data of the first AI model to derive first material fractionation data; providing the input spectral X-ray projection data as model input data to a second AI model, and then applying high-pass filtering to model output data of the second AI model to derive second material fractionation data; and linearly combining the first and second material fractionation data to derive third material fractionation data.

[0016] An embodiment of the present invention is based on the concept of using two trained AI models for material fractionation: one that provides low-bias but high-noise output material fractionation data, and the other that provides high-bias but low-noise output material fractionation data. The outputs of the two models are later combined via a frequency separation approach to achieve a low-bias, low-noise result. The frequency separation approach essentially involves selecting only the low-frequency components from the low-bias (high-noise) AI model and the high-frequency components from the high-bias (low-noise) model, and linearly combining them. This can be done using filtering.

[0017] More specifically, low-pass filtering extracts or selects only the low-frequency or low-energy components of the output of the first AI model (low bias), and high-pass filtering extracts or selects only the high-frequency or high-energy components of the output of the second AI model (high bias). Linearly combining the two results in output material fractionation data with low noise and low bias. This is based on the premise that noise is primarily a high spatial frequency effect, and bias is primarily a low-frequency effect.

[0018] Preferably, the method includes applying low-pass spatial filtering to the model output data of the first AI model, preferably in at least one dimension of the output data, where low-pass spatial filtering means applying low-pass filtering of the spatial frequencies of the output material fractionation data.

[0019] Preferably, the method includes applying high-pass spatial filtering to the model output data of the second AI model, preferably in at least one of the same dimensions as above, where high-pass spatial filtering refers to applying high-pass filtering of the spatial frequencies of the output material fractionation data.

[0020] For example, the model output data may include data values ​​spanning two or more dimensions. For example, the model output data may include material fractionation data values ​​for each of a plurality of pixels in an array of pixels, the pixel array spanning at least two dimensions. For example, low-pass spatial filtering may be for filtering spatial frequency components of the set of model output data in at least one spatial dimension of the data. For example, high-pass spatial filtering may be for filtering spatial frequency components of the set of model output data in the same at least one spatial dimension of the data.

[0021] Spatial frequency filtering generally means that, for example, an image is segmented by periodic patterns (this can be any image), each periodic pattern is weighted with a weighting factor, and then the weighted periodic patterns are summed to obtain a modified (filtered) image. This modification is due to the weighting. In low-pass filtering, higher weights are applied to higher frequencies of the periodic pattern, and similarly in 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 2D or 3D images.

[0022] The model output data can be understood as an "image" of the material fractionation data because it contains data values ​​for each of a set of pixel locations spanning, for example, a 2D or 3D grid. Thus, spatial filtering can be applied to the output data of the two models in the same way as it is applied to standard images.

[0023] In practice, there are various options for how spatial filtering is performed on the model outputs of the first and second AI models. For example, in one non-limiting example, the model output data includes data values ​​(e.g., material path length values ​​or other values) for each of a set of materials, and for each of a set of images, and for each (2D) array of pixels as corresponding 2D "image" datasets for each image and each material, and spatial filtering is applied to each of these 2D "image" datasets. As a variation of this example, a set of 3D image datasets may be formed for each material. Each 3D dataset may include a stack of the above-mentioned 2D "images," with the third dimension corresponding to a different image, and spatial filtering is applied in the 3D domain. That is, spatial filtering is applied to each of the 3D datasets for each material, e.g., in one or more dimensions of this 3D array of values. It will be understood that in this latter option, the third dimension need not necessarily be a physical spatial dimension. Thus, spatial filtering need not be applied along a physical spatial dimension, but may be applied to a periodic pattern in the output data across another dimension of the output data array. Other options will be apparent to those skilled in the art.

[0024] Two different AI models can be generated either by differing the training data fed to the two models during training (e.g., one is fed with noisier data than the other), or by differing the training procedure itself between the two models (e.g., training one model with a different cost function (configured to favor noise) than the other). Each of the AI ​​models may include one or more machine learning algorithms (e.g., one or more artificial neural networks).

[0025] For the avoidance of doubt, it should be noted that the first AI model is configured to exhibit lower bias in the output material fractionation data values ​​than the second AI model and higher variance in the output material fractionation data values ​​than the second AI model.

[0026] The above method is broadly applicable to a wide variety of specific applications, in particular, various types of input X-ray data may be used (e.g., conventional X-ray data or CT data), and various types of output material composition data may be generated (e.g., material projection data or material image data).

[0027] Thus, in different embodiments, the input spectral X-ray data may be spectral projection data (i.e., unreconstructed spectral projection data, e.g., in the form of photon count data for a set of different energy bins) or spectral image data (i.e., reconstructed spectral projection data). The input spectral X-ray data may include raw projection data generated by an X-ray scanning device, or may be pre-processed, e.g., reconstructed data to form spectral image data.

[0028] In different embodiments, the material fractionation data may be material projection data or material image data.

[0029] In some examples, the spectral X-ray data may include spectral computed tomography (CT) data of the object (including projection data for multiple projection angles), or may include conventional X-ray or fluoroscopic data captured from a single projection angle.

[0030] In different embodiments, the entire process may be performed in the projection data domain, the image data domain (by reconstructing the spectral projection data into image data before feeding it to the two AI models), or part of the process may be performed in the projection domain and part in the image domain. For example, the AI ​​model may accept input spectral projection data and output material projection data, while part of the frequency filtering is performed in the image domain after the material projection data is reconstructed. The various options are described in more detail below.

[0031] The first and second AI models may be machine learning models trained using a supervised learning procedure.

[0032] According to one or more embodiments, a first AI model is trained based on training data T1, and a second AI model is trained based on training data T2, where T1 is generated to have less noise than T2. ​​The reduced noise in T1 can be achieved, for example, by preprocessing the training data so that it has less 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, where T1 is generated by applying noise suppression to T2. However, in other examples, noise can be artificially added to the training data set T2. That is, the first AI model is trained on training data T1, and the second AI model is trained on training data T2, where T2 is generated by adding noise to T1. The added noise can be simulated, for example, using a random noise generator. In further examples, one or both of T1 and T2 can be generated by simulation (i.e., using an X-ray simulation program). Simulation allows for the noise level of the obtained data to be adjusted as desired, so one or both of T1 and T2 can be obtained in this manner.

[0033] In this approach, two AI models start out as two copies of the same machine learning network or algorithm, but are trained on separate and distinct training data to introduce differences in their biases and variances.

[0034] As an alternative to this approach, the learning algorithms included in the first and second machine learning models can follow different learning or training strategies: one that favors low bias (high noise) and one that favors low noise (high bias).

[0035] For example, by configuring the first model with a different cost function, the bias of the first model is reduced, so that the cost function of the second model provides more noise and the cost function of the first model provides less noise.

[0036] According to one or more embodiments, each of the first and second AI models receives spectral X-ray projection data as model input data and generates material composition projection data as model output data, whereby material fractionation occurs in the projection data domain.

[0037] The method may be, for example, a basis material fractionation method, and the output material fractionation data may include, for each of a set of basis materials with known attenuation coefficients, the derived radiation path length l through the scanned material. s where the output material fractionation data includes material projection data, i.e., unreconstructed material fractionation data rather than, for example, image data.

[0038] If the input spectral X-ray data is spectral CT data, the output material fractionation data is the path length sinogram vector

number

[0039] A sinogram is a visual representation of the raw data acquired by a computed tomography operation. A pathlength sinogram is a visual representation of the material basis pathlengths for each pixel of the detector.

[0040] According to one or more embodiments, the method further includes applying a reconstruction operation to the first and second material fractionation data images, either before or after the linear combination, so that the output material fractionation data comprises output material image data.

[0041] Material image data refers to a visual or graphical representation of material classification information. For example, material image data may include a set of images, each image representing one material. A further option is to use color coding to represent different materials in the imaged area within one image (e.g., each material is represented by a dedicated color).

[0042] According to one or more embodiments, the input spectral X-ray data may be projection data, e.g., spectral photon count data c indicative of X-ray photon counts detected in each of a plurality of energy bins at an X-ray detector. s Includes:

[0043] According to a further set of embodiments, each of the first and second models receives spectral X-ray image data as model input data and produces material fractionation image data as model output data, where material fractionation is thus performed in the image domain rather than the projection domain.

[0044] As mentioned above, according to one or more embodiments, the input spectral X-ray data is spectral CT data, which in different implementations may be spectral CT projection data or spectral CT image data. Alternatively, the input spectral X-ray data may be conventional spectral X-ray data.

[0045] According to this set of embodiments, the input spectral X-ray data is generated as spectral photon count data c indicating the X-ray photon counts detected in each of a plurality of energy bins at the X-ray detector, for each of a plurality of pixels p, and for each of a plurality of projections v of the input spectral X-ray data. s Contains (p,v).

[0046] An example according to a further aspect of the present invention provides a computer program product comprising computer program code executable on a processor or computer, the code causing the processor to perform a method according to any example or embodiment above or below, or according to any claim of the present application.

[0047] An example according to a further aspect of the present invention provides a processing device for processing X-ray spectral data to derive material fractionation information, the processing device comprising: accessing a data storage device, a first AI model and a second AI model are stored in the data storage device, and each of the first and second AI models receives the X-ray spectrum data as model input data and generates material fractionation data as model output data; accessing, the first AI model being configured to exhibit lower bias than the second AI model and higher variance than the second model; obtaining input spectral X-ray data; providing the input spectral X-ray data as model input data to a first AI model, and then applying low-pass filtering to model output data of the first AI model to derive first material fractionation data; providing the input spectral X-ray data as model input data to a second AI model, and then applying high-pass filtering to model output data of the second AI model to derive second material fractionation data; and linearly combining the first and second material fractionation data to derive third material fractionation data.

[0048] According to some embodiments, the processing device may include a data storage device having the first and second AI models stored thereon.

[0049] An example according to a further aspect of the present invention provides an X-ray imaging system, the X-ray imaging system comprising: an X-ray scanning assembly including an X-ray radiation source and an X-ray radiation detector, wherein a scan region for receiving an object to be scanned is between the X-ray radiation source and the X-ray radiation detector, and an X-ray radiation path between the X-ray radiation source and the X-ray radiation detector passes through the scan region; and a processing device according to any example or embodiment above or below, or according to any claim of the present application, wherein the processing device is communicatively coupled to the X-ray radiation detector to receive input spectral X-ray data from the X-ray detector.

[0050] The scanner assembly may be a CT scanning assembly, which therefore allows for an adjustable angular position of the x-ray source relative to the x-ray detector.

[0051] The X-ray detector may be an X-ray photon counting detector that detects photon counts in each of a set of radiation frequency or energy bins.

[0052] An embodiment according to a further aspect of the present invention provides a method for training first and second AI models for use in deriving material fractionation data from input spectral X-ray data, the method comprising: obtaining a preliminary AI model that obtains spectral X-ray data as model input data and generates material fractionation data as model output data; training a first copy of the preliminary AI model with first training data T1 to obtain a first AI model; training a second copy of the preliminary AI model with second training data T2 to obtain a second AI model; each of the first and second training data includes an input data entry in the form of sample spectral X-ray data and a corresponding output data entry in the form of known material fractionation data for each sample spectral X-ray data; The first training data T1 is generated or pre-processed to reduce noise in the spectral X-ray data forming the input data entries so that the first training data is less noisy compared to the second training data.

[0053] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter. [Brief explanation of the drawings]

[0054] For a better understanding of the present invention and to show more clearly how it may be carried into effect, reference will now be made, by way of example only, to the accompanying drawings in which:

[0055] [Figure 1] FIG. 1 illustrates, in block diagram form, an overview of exemplary method steps in accordance with one or more embodiments. [Figure 2] FIG. 2 shows comparative material fractionation results obtained using an embodiment of the present invention. [Figure 3] FIG. 3 illustrates an exemplary processing device for implementing methods according to one or more embodiments of the present invention. [Figure 4] FIG. 4 illustrates a schematic diagram of an exemplary X-ray system for material separation of a scanned object. [Figure 5] FIG. 5 outlines the steps of an exemplary method for training two AI models for use in material fractionation based on spectral X-ray data. DETAILED DESCRIPTION OF THE INVENTION

[0056] The present invention will now be described with reference to the figures.

[0057] It should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the devices, systems, and methods, are for purposes of illustration only and are not intended to limit the scope of the invention. These and other features, aspects, and advantages of the devices, systems, and methods of the present invention will become better understood from the following description, the appended claims, and the accompanying drawings. It should be understood that the figures are schematic representations only and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the figures to indicate the same or similar parts.

[0058] An embodiment of the present invention provides a method for material fractionation of an object based on spectral X-ray scan data of the object and the application of an AI model and a frequency separation approach. The method includes using two AI models in parallel to perform material fractionation based on input spectral X-ray data. The models are configured so that one exhibits higher bias and lower variance (low noise) than the other. The input spectral X-ray data is fed to 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. The resulting combined material fractionation data has both lower bias and noise compared to the output generated using only one AI model.

[0059] By using a multi-energy X-ray detector, it is possible to measure the spectrum of X-ray energies or frequencies received at different locations on the X-ray detector during X-ray imaging of an object. This is known as spectral X-ray imaging. Spectral X-ray data allows the identification and quantification of materials contained in the scanned object. This can be done using data from conventional X-ray imaging modalities or CT imaging.

[0060] Material fractionation can be achieved based on pixel-by-pixel inversion of a system model, which describes the theoretical signal expected at each X-ray detector energy bin as a function of a specific set of materials (basis) contained in the object and their respective equivalent path lengths. This procedure is called basis material fractionation.

[0061] Material fractionation using spectral photon-counting X-ray data is a nonlinear, ill-posed, and computationally expensive problem. Using AI models for fractionation can significantly reduce computation time, but the ill-posed nature of the problem can still result in strong noise amplification and bias in the output of the AI ​​model.

[0062] In an embodiment of the present invention, a classification method using two or more AI models is proposed to reduce bias and noise. This classification method utilizes the concept of frequency separation based on the premise that noise has high spatial frequencies in the projection and image domains, while bias is a low-frequency effect.

[0063] The basic principles of spectral X-ray based material fractionation as applied to embodiments of the present invention are briefly described below. A more detailed description of the principles underlying material fractionation as applied herein can also be found in the following journal article: KC Zimmerman and TG Schmidt, "Experimental comparison of empirical material decomposition methods for spectral CT," Physics in Medicine & Biology, 60 (2015) 8.

[0064] 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. The charge is converted into a voltage using a charge-integrating amplifier. Using an analog comparator, a digital counter is incremented when the voltage of the accumulated charge exceeds a set threshold level. At the end of the acquisition, the number of detected photons with energies above the threshold is counted. Successive counter measurements are subtracted to generate energy bin data corresponding to the number of photons detected between two threshold levels. In this way, photon counts for multiple X-ray energy bins are generated.

[0065] It should be noted that the following description assumes the use of a spectral photon counting detector, however other types of detectors are possible, such as a spectral energy integrating detector (EID) that can detect radiation within a specific window in the energy spectrum of the X-ray source.

[0066] Consider an X-ray measurement through a material of thickness x and attenuation coefficient μ(E), where E is the energy of the photon traversing the material. The X-ray attenuation of a composite material can be expressed as a linear combination of the attenuations of the set of basis materials contained in the material. More specifically, and as explained in KC Zimmerman (2015), the attenuation of an X-ray through this material corresponds to the attenuation of a unique combination of any two other materials (without a K-edge), as expressed in Equation (1): where μ(E) and μ(E) are the energy-dependent attenuation coefficients of each of the two basis materials, and l(E) and l(E) are the path lengths of each of the two basis materials. This distinction is possible because of the existence of two main attenuation phenomena in the diagnostic X-ray energy range: Compton scattering and photoelectric absorption. xμ(E)=l1μ1(E)+l2μ2(E) (1)

[0067] By extending this basis expansion, we obtain the ith energy bin c of an ideal photon-counting detector. i We can model the expected number of photons detected at

number

[0068] In general, material fractionation involves estimating the thickness l of the underlying material from the acquired spectral data c. One way to estimate the thickness l of the underlying material from the number of detected photons c is to numerically invert equation (2) using a statistical estimation algorithm, for example, maximum likelihood estimation (MLE). However, this leads to an oversimplification and fails to consider real-world detector effects such as pulse pile-up and charge sharing.

[0069] Another approach that can overcome some of these problems is an artificial intelligence (AI) model-based approach, which is used in embodiments of the present invention.

[0070] According to an embodiment of the present invention, a machine learning algorithm, such as a neural network trained with energy bin data from known thicknesses of basis materials, is employed to perform basis material fractionation in the projection domain. For example, a database of various reference path length vectors l(p) for a given material basis of interest is generated along with the resulting photon count data vector c(p), both of which are functions of detector pixel p.

[0071] The trained AI model f estimates the thickness of the underlying material of the scanned object from the spectral X-ray measurements, which can be expressed as:

number

number

[0072] Training the AI ​​model f can be based on using the aforementioned database of reference path length vectors l(p) and photon counts, and training an AI model for each pixel in the database, such that the resulting AI model f minimizes the error Er between the predicted path length (the output of the model) and the actual path length:

number

[0073] The above principles can be applied in accordance with various embodiments of the present invention. Some embodiments of the present invention will now be described in more detail.

[0074] FIG. 1 illustrates, in block diagram form, an overview of steps in an exemplary method, according to one or more embodiments, for material fractionation of an object using spectral X-ray data of the object.

[0075] The method includes accessing (12) a data storage device having a first AI model and a second AI model stored therein, each of the first and second AI models receiving the X-ray spectral data as model input data and generating material fractionation data as model output data.

[0076] The first AI model exhibits lower bias than the second AI model. The first AI model exhibits higher variance than the second AI model.

[0077] The method further includes acquiring (14) input spectral X-ray data, which may be input projection data or reconstructed spectral image data. The input X-ray data may be acquired directly from an X-ray scanning device, e.g., for real-time processing; from a data store, e.g., for offline processing; or from a processing component or unit that pre-processes data from a reconstructor that receives raw projection data and outputs reconstructed spectral image data. Thus, the form of the spectral X-ray data received and provided as input to the two AI models may vary in various embodiments.

[0078] The method further includes providing (16) the input spectral X-ray data to the first AI model as model input data.

[0079] The method further includes providing (18) the input spectral X-ray projection data to a second AI model as model input data.

[0080] The method further includes applying low-pass filtering to the model output data of the first AI model and applying high-pass filtering to the model output data of the second AI model (20).

[0081] The low pass filtering can be any linear low pass filter, including, but not limited to, a moving average filter, a Gaussian window filter, a Hann window filter, a Hamming window filter, a Kaiser window filter, or any other suitable low pass filter as would be apparent to one skilled in the art.

[0082] Again, any suitable high-pass filter can be used for high-pass filtering, as will be apparent to those skilled in the art. By way of non-limiting example, high-pass filtering can be performed using a Dirac pulse minus a low-pass filter, or, for example, by low-pass filtering the data and then subtracting the low-pass filtered data from the original data.

[0083] The method further includes linearly combining the outputs of the first and second AI models to derive output material fractionation data (22). The linear combination can be performed after the model outputs have been filtered.

[0084] The combination of filtering and linear combination of model outputs (as described above) is called the frequency separation approach or procedure, and this term is used herein for simplicity.

[0085] Optionally, the method further comprises a reconstruction step, which comprises applying an image reconstruction operation to the outputs of the first and second AI models, either before or after filtering and / or linear combination, such that the output material fractionation data comprises output material image data.

[0086] Material image data refers to a visual or graphical representation of material classification information. For example, material image data may include a set of images, each image representing one material. A further option is to use color coding to represent different materials in the imaged area within one image (e.g., each material is represented by a dedicated color).

[0087] In different embodiments of the invention, the AI ​​model receives input spectral x-ray data in various formats and generates output material composition data in various formats, for example, the input data can be projection data or reconstructed image data, and the output material composition data from the model can be in the projection domain (material projection data) or the image domain (material image data).

[0088] Four main different advantageous options are briefly outlined below. (1) The input spectral X-ray data is data from a conventional X-ray modality system. Here, the acquired spectral X-ray projection data is fractionated using two AI models to obtain two sets of output material projection data. These two sets of material projection data are then combined in the frequency separation step described above to become the final material projection data. This provides the output material fractionation data, which in some instances can be directly used for visualization. (2) Spectral X-ray data is data from a computed tomography (CT) system. The acquired spectral X-ray projection data is provided as input data to two AI models. It is then separated using the two AI models to derive two sets of output material projection data. For example, the material projection data includes vectors of material path lengths within the scanned object for a set of basis materials. These two sets of material projection data are combined using the frequency separation steps (filtering and linear combination) described above to derive the final material projection data. Optionally, they can be subsequently processed in a reconstruction procedure to derive the final material image data. (3) The input spectral X-ray data is data from a computed tomography (CT) system. The acquired spectral X-ray projection data is provided as input data to two AI models. The data is separated using the two AI models to derive two sets of material projection data. These two sets of material projection data sets are reconstructed together to derive two sets of material image data. These image data sets are combined in a frequency separation step (filtering and linear combination) as described above to derive the final material image data. (4) Spectral X-ray data is data from a computed tomography (CT) system. The acquired spectral X-ray projection data is reconstructed to derive a set of spectral image data as input spectral X-ray data, which is provided as input data to two AI models. The spectral image data is fractionated by the two AI models to derive two sets of material image data, which are then combined using the frequency separation procedure (filtering and linear combination) described above to derive the final material image data.

[0089] As such, options (2) to (4) differ in the order of the various processing steps (classification / frequency separation / reconstruction).Option (1) does not have a reconstruction step.

[0090] To obtain the results of the above method, two AI models are required, both capable of material fractionation but with different biases and variances.

[0091] Each AI model includes one or more machine learning algorithms.

[0092] A machine learning algorithm is any self-training algorithm that processes input data to generate or predict output data, where the input data includes spectral x-ray data and the output data includes material fractionation data.

[0093] Machine learning algorithms suitable for use in the present invention will be apparent to those skilled in the art. Examples of suitable machine learning algorithms include decision tree algorithms and artificial neural networks. Other machine learning algorithms, such as polynomial regression, support vector regression, or kernel regression, are suitable alternatives.

[0094] The structure of an artificial neural network (or simply a neural network) is inspired by the human brain. A neural network is made up of layers, each containing multiple neurons. Each neuron contains a mathematical operation. In particular, each neuron may contain different weighted combinations of a single type of transformation (e.g., the same type of transformation, such as sigmoid, but with different weights). In the process of processing input data, each neuron's mathematical operation is performed on the input data to generate a numerical output, and the output of each layer of the neural network is fed sequentially to the next layer. The final layer provides the output.

[0095] Methods for training machine learning algorithms are well known. Typically, such methods involve obtaining a training dataset containing 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 the corresponding training output data entry is used to correct the machine learning algorithm. This process is repeated until the error converges, i.e., until the predicted output data entries are sufficiently similar to the training output data entries (e.g., ±1%). This is commonly known as a supervised learning technique.

[0096] For example, if a machine learning algorithm is formed from a neural network, the mathematical operations (weights) of each neuron can be modified until the error converges. Known methods for modifying neural networks include gradient descent algorithms and backpropagation algorithms.

[0097] The training input data entries correspond to exemplary input spectral x-ray data for a given material basis (e.g., sample photon energy bin counts for a material basis). The training output data entries correspond to material fractionation data.

[0098] As previously mentioned, in different embodiments of the present invention, the AI ​​model receives input x-ray data in various formats and generates output material composition data in various formats. For example, the input data can be projection data or reconstructed image data, and the output material composition data from the model can be in the projection domain (material projection data) or the image domain (material image data).

[0099] To better understand the principles of the present invention, we will describe in more detail an example in which the input data to the AI ​​model is CT projection data and the output material fractionation data from the AI ​​model is fractionated projection data, but the same principles can be applied to models that input and output data in different formats.

[0100] According to this example, the fractionation method is a basis material fractionation method. The final output material fractionation data in this case is the derived radiation path length l through the scanned material for each of a set of basis materials with known attenuation coefficients. s The theory behind this approach is explained in more detail above, and the reader is referred to this description for further details.

[0101] The final output material fraction data can be, for example, a path length sinogram vector

number

[0102] As mentioned above, basis material fractionation in the projection domain requires an effective inversion of a system forward model (e.g., in a form similar to equation (2) above), which models the expected photon counts in individual detector energy bins for each pixel element of the X-ray projection data as a function of their basis and equivalent path length. Machine learning algorithms, such as neural networks, can be trained to perform this inversion.

[0103] To train the AI ​​model, a reference database of various path length vectors l(p) for a given material basis of interest can be used along with the resulting photon count data vector c(p), both of which are functions of pixel p. The AI ​​model is then learned for each pixel in the database such that the resulting AI model f minimizes the error Er between the predicted path length (the output of the model) and the actual path length, where w(p) is the parameter vector for the AI ​​model f for pixel p:

number

number

number

[0104] As mentioned above, the resulting trained model aims to compromise both phenomena to some extent in terms of output path length bias and noise. The goal according to embodiments of the present invention is to provide an overall method that results in both low bias and low noise.

[0105] For this reason, and as explained, the first AI model exhibits lower bias than the second AI model and higher variance than the second AI model. In this context, bias error is understood as the deviation of the average of the AI ​​model's estimated output path lengths from the actual path lengths. Variance error is understood as the variance of said model output path length estimates relative to the actual path lengths.

[0106] There are various ways to achieve a first AI model that exhibits lower bias than a second AI model and higher variance than a second AI model.

[0107] One approach is to first generate two copies of the preliminary network f and train the two copies on two different types of photon count data (one noise-free and one with the complete full noise expected during a typical X-ray or CT scan). For example, generate two copies of the initial training dataset, then preprocess one of the datasets to reduce the noise. Train the first AI model using the low-noise data and train the second AI model using the full-noise data. After training with the noiseless data, a first AI model with low bias and high noise (high variance) is achieved:

number

number

[0108] Another way to introduce bias and variance differences between two AI models is to use the same training data for both but change the training strategies employed by the two models. That is, change the learning algorithms encoded in the models so that one model favors low bias and the other favors low noise (low variance). For example, by configuring the first model with a different cost function, the bias of the first model is reduced. Thus, the cost function of the second model favors more noise and the cost function of the first model favors lower noise. Both models are then trained using full-noise training data.

[0109] For example, according to one or more embodiments, the bias that occurs is calculated during the training of a given model, and the absolute value of the bias is used as a term in the cost function. The more weight is given to this term, the lower the bias of the trained model. Therefore, when training the first AI model, a higher weight is given to this term in the cost function.

[0110] For example, if several (e.g.) noisy photon count realizations are available in the training data set for a given material combination, the bias can be calculated during training.

[0111] The bias and variance of an AI model (e.g., a neural network) are measurable and testable characteristics. Measurement of the bias and variance of a given model (the model is trained to generate output material fractionation information based on input spectral X-ray data) can be performed empirically, for example, by scanning a phantom with inserts featuring different known material concentrations (e.g., iodine and water) and generating corresponding spectral X-ray data for the phantom.

[0112] By feeding this input spectral X-ray data into a model of the test object, bias can be measured as the deviation of the mean density value measured in a region of interest (ROI) of the corresponding image relative to the known density. Noise or variance can be quantified as the standard deviation of the values ​​measured in each homogeneous portion of the material image.

[0113] Following the separation performed by the two AI models, as described above, a "frequency separation" step is applied, which involves selective filtering of the outputs of the first and second AI models and combining the results with a linear combination.

[0114] In particular, in the context of the above example where the input spectral X-ray data is spectral CT projection data, this involves filtering the model output using a linear low-pass filter (G) and a matching linear high-pass filter (1-G), respectively, and then combining the individual pathlength sinograms resulting from applying the f1 and f2 models, respectively:

number

[0115] The resulting combined sinogram vector

number

[0116] The resulting combined path length sinogram

number

[0117] By way of illustration, Figure 2 shows exemplary simulation results of a material fractionation method according to an embodiment of the present invention. The first column shows actual material fractionation images of an iodine sample (top row) and an actual material fractionation image of a water sample (bottom row). The second, third, and fourth columns show image reconstructions of material fractionation sinogram data resulting from fractionation applied by the first AI model f1 only, the second AI model f2 only, and the fully coupled fractionation method (including frequency separation method) according to the present invention, respectively. As mentioned above, the first model f1 is a bias-optimal model (low bias), and the second model f2 is a noise-optimal model (low noise).

[0118] It can be seen that the results from the combination of the two models (fourth column) yields the best results with bias and noise levels comparable to those provided by the bias-optimal model (f1) and the noise-optimal model (f2), respectively.

[0119] The above description is about training an AI model that receives spectral CT data (projection data) in the projection domain and outputs material fractionation data in the projection domain, i.e., the path length vectors in a given material basis.

number

[0120] However, the same principles can equally be applied to training and using AI models that receive different types of spectral X-ray data and output different types of fractionation information. To derive these models, one simply needs to select the training data appropriately, so that the training input data entries are of the correct modality or format, and the training output data entries are material composition data in the correct format.

[0121] For example, in various embodiments, the input spectral X-ray data provided to each AI model may be spectral computed tomography (CT) data, as in the example above, or in further embodiments, spectral X-ray data from a conventional (non-CT) X-ray scanner or a fluoroscopy scanner. In the latter case, the input spectral X-ray data to the AI ​​model may be photon count projection data c that provides the photon count for each energy bin as a function of detector pixels. s (p). A conventional X-ray scanner may only have one image, so the photon count data may only be a function of pixels. In another case, the conventional spectral X-ray data is reconstructed into image data before applying an AI model, and the AI ​​model can be trained to receive input spectral image data and output material image information.

[0122] When the input spectral X-ray data is spectral CT data, the two AI models can receive either the input spectral CT projection data or reconstructed spectral CT image data reconstructed from the raw spectral projection data. In the former case, the AI ​​model is trained to output material fractionation data in the projection domain, and in the latter case, the model is trained to output material composition data in the image domain.

[0123] An example according to a further aspect of the present invention provides a computer program product comprising computer program code executable on a processor or computer, the code causing the processor to perform a method according to any example or embodiment above or below, or according to any claim of the present application.

[0124] FIG. 3 illustrates components of an exemplary apparatus for performing material fractionation methods according to embodiments disclosed herein.

[0125] In particular, one aspect of the present invention provides a processing unit 32 that processes the spectral X-ray data to derive material fractionation information. In this example, the processing unit includes a single processor component (e.g., an integrated circuit (IC)).

[0126] The processing device 32 has access to a 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 receives spectral X-ray data as model input data and generates material fractionation data as model output data. The first AI model 42 exhibits lower bias than the second AI model 44 and higher variance than the second AI model.

[0127] The processing device obtains input spectral X-ray data, for example, from an X-ray scanning device communicatively coupled to the processing device, or, in other examples, from an additional data store.

[0128] The processing unit 32 provides the input spectral X-ray data as model input data to the first AI model 42, and then applies low-pass filtering to the model output data of the first AI model to derive first material fractionation data.

[0129] The processing unit 32 provides the input spectral X-ray data as model input data to the second AI model 44, and then applies high-pass filtering to the model output data of the second AI model to derive second material fractionation data.

[0130] The processing device 32 linearly combines the first and second material fractionation data to derive third material fractionation data.

[0131] According to one or more embodiments, the processing device 32 may include a data storage device 36 that stores the first AI model 42 and the second AI model 44. Alternatively, the processing device 32 may include inputs / outputs for connecting to the data storage device.

[0132] An example according to a further embodiment provides a spectral X-ray system, an exemplary system 60 is shown schematically in FIG.

[0133] The system includes an X-ray scanning assembly 62 that includes an X-ray radiation source 64 (X-ray generator) and an X-ray radiation detector 66, with a scan region between the X-ray radiation source 64 and the X-ray radiation detector 66 for receiving an object to be scanned. An X-ray radiation path 68 between the source and detector passes through the scan region.

[0134] The system further includes a processing device 32 according to any of the examples or embodiments described above or below, or according to any claim of the present application, the processing device being communicatively coupled to the X-ray radiation detector 66 to receive input spectral X-ray data from the X-ray detector.

[0135] 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 positions, thereby enabling the capture of multiple projection images corresponding to different X-ray path angles through the scan region.

[0136] In a preferred example, the X-ray detector 66 is an X-ray photon counting detector that detects photon counts in each of a set of radiation frequency or energy bins.

[0137] Alternatively, the X-ray detector 66 may include one or more energy integrating detectors (EIDs). The EIDs are capable of selectively detecting incident X-ray radiation in a selected radiation window. In one possible arrangement, two or more EIDs are used in a stacked or bi-layer configuration, each sensitive to a different energy window in the X-ray spectrum of the X-ray source (e.g., X-ray tube).

[0138] Another possible device uses a single EID. In this case, 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 levels of the radiation are controlled by the source rather than being differentially detected by the detector.

[0139] The basic structure of an X-ray scanning assembly suitable for use in embodiments of the present invention is described, for example, in International Patent Publication WO 2014 / 167450, from page 3, line 26 to page 5, line 2, and shown therein in Figure 1. The radiation detector array may take the form, for example, of a photon-counting detector or one or more energy-integrating detectors.

[0140] Details of the structure, operation, and implementation of a suitable X-ray scanning assembly can be found in the following book: Spectral, Photon Counting Computed Tomography (Technology and Applications) by Katsuyuki Taguchi (2020).

[0141] The main principles of proper CT scanning assembly are described in detail in the following paper: Willemink, M. et al., "Photon-counting CT: Technical Principles and Clinical Prospects" (2018).

[0142] As mentioned above, one way to obtain two AI models with different biases and variances is to train the two models on different training data.

[0143] Thus, an example according to a further aspect of the present invention provides a method for training first and second AI models for use in deriving material fractionation data from input spectral X-ray data. Figure 5 shows, in block diagram form, an overview of the steps of one example of such a method.

[0144] The method includes obtaining (82) a preliminary AI model that receives spectral X-ray data as model input data and generates material fractionation data as model output data.

[0145] The method further includes obtaining (84) first training data T1 and second training data T2, where each of the first and second training data includes input data entries in the form of sample X-ray spectral data and corresponding output data entries in the form of known material fractionation data for each sample X-ray spectral data. The first training data T1 is provided having X-ray spectral data (forming the input data entries) that are less noisy than T2. ​​Thus, the first training data is less noisy than the second training data. This can be achieved, for example, by preprocessing the first training data to reduce noise. For example, T1 is generated by applying noise suppression to a second copy of T2. Alternatively, the second training data can be preprocessed to have additional noise (compared to T1). That is, noise can be artificially added to the training data set T2. That is, T2 is generated by adding noise to a second copy of T1. The added noise can be simulated, for example, using a random noise generator. In a further example, one or both of T1 and T2 can be generated by simulation (i.e., using an X-ray simulation program). In simulations, either or both T1 and T2 can be acquired in this manner, as the noise level of the acquired data can be arbitrarily adjusted.

[0146] The method further includes training 86 a first copy of the preliminary AI model with the first training data T1, thereby obtaining a first AI model.

[0147] The method further includes training 88 a second copy of the preliminary AI model with second training data T2, thereby obtaining a second AI model.

[0148] This results in a first AI model that has less bias and more variance than the second model.

[0149] In the above example, the two models start out as copies of the same preliminary model, but this is not required. In further embodiments, a first preliminary model can be trained with first training data T1 to obtain a first AI model, and a second preliminary model can be trained with second training data T2 to obtain a second AI model, and the first and second preliminary models can be different. For example, one preliminary model is a neural network and the other is a linear regression model. Even if different training data is used to train the preliminary models (as in the manner described above), the bias and variance of the two models will be different.

[0150] As mentioned above, another way to introduce 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: that is, change the learning algorithms encoded in the models so that one model favors low bias and the other favors low noise (low variance).

[0151] Embodiments of the present invention have a wide range of applications in a variety of different fields. For example, the methods and systems described above can be used to process medical imaging data, such as medical CT data. Material fractionation in the medical field is often used, for example, to analyze the composition of tumors to determine whether tissue is cancerous. However, embodiments can also be used outside of the medical field, for example, to analyze the material composition of inanimate objects.

[0152] As previously mentioned, systems use processors to process data. Processors can be implemented in a variety of ways using software or hardware to perform the various functions required. Typically, a processor uses one or more microprocessors that are programmed using software (e.g., microcode) to perform the required functions. A processor can be implemented as a combination of dedicated hardware to perform some functions and one or more programmed microprocessors and associated circuitry to perform other functions.

[0153] Examples of circuitry that may be used in various embodiments of the present disclosure include, but are not limited to, conventional microprocessors, application specific integrated circuits (ASICs), and field programmable gate arrays (FPGAs).

[0154] In various implementations, a processor may be associated with one or more storage media, such as volatile and non-volatile computer memory, such as RAM, PROM, EPROM, and EEPROM. The storage media may be encoded with one or more programs that, when executed on the one or more processors and / or controllers, perform the necessary functions. The various storage media may be fixed within the processor or controller, or may be transportable so that one or more programs stored thereon can be loaded into the processor.

[0155] Variations of the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps and the singular does not exclude a plurality.

[0156] A single processor or other unit may fulfill the functions of several items recited in the claims.

[0157] The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

[0158] The computer program may be stored / distributed on any suitable medium, such as an optical storage medium or a solid-state medium, supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless communication systems.

[0159] It should be noted that when the term "adapted to" is used in the claims or description, the term "adapted to" is intended to be equivalent to the term "configured to."

[0160] Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. 1. A method for material fractionation of an object using spectral X-ray data of the object, the method comprising: accessing a data storage device, a first AI model and a second AI model are stored in the data storage device, and each of the first AI model and the second AI model receives spectral X-ray data as model input data and generates material fractionation data as model output data; the first AI model is configured to exhibit lower bias in output material fractionation data values ​​than the second AI model when applied to the same input data set, and to exhibit higher variance in the output material fractionation data values ​​than the second AI model when applied to the same input data set; acquiring input spectral X-ray data; providing the input spectral X-ray data as the model input data to the first AI model; providing the input spectral X-ray data as the model input data to the second AI model; applying low-pass spatial filtering to the model output data of the first AI model; applying high-pass spatial filtering to the model output data of the second AI model; linearly combining the filtered outputs of the first AI model and the second AI model to derive output material fractionation data; Including, wherein the high-pass spatial filtering is not applied to the model output data of the first AI model, and the low-pass spatial filtering is not applied to the model output data of the second AI model, The first AI model and the second AI model are artificial neural networks derived using a supervised training method, the supervised training method comprising: obtaining a preliminary AI model that receives X-ray spectral data as model input data and generates material fractionation data as model output data; A first copy of the preliminary AI model is then transferred to a first training data T １ to obtain the first AI model; A second copy of the preliminary AI model is used as a second training data T ２ to obtain the second AI model; Including, each of the first training data and the second training data includes an input data entry in the form of sample spectral X-ray data and a corresponding output data entry in the form of known material fractionation data for each of the sample spectral X-ray data; The first training data T １ is set to the second training data T such that the first training data is less noisy compared to the second training data. ２ by preprocessing the second training data T ２ wherein the sample spectral X-ray data, generated from and forming the input data entries, is noise reduced.

2. The method of claim 1 , wherein the first AI model and the second AI model each receive spectral X-ray projection data as model input data and generate material fractionation projection data as model output data.

3. The method is a basis material fractionation method, and the output material fractionation data includes, for each of a set of basis materials with known attenuation coefficients, a radiation path length l derived through the scanned material. ｓ The method of claim 2 , comprising the set:

4. The output material fractionation data is a path length sinogram vector [0016] and the path length sinogram vector comprises derived basis material path lengths for each of the set of basis materials for radiation reaching each of a set of pixels p and for each of a set of projections v of the input spectral X-ray data.

5. 5. The method of claim 2, further comprising applying an image reconstruction operation to the model output data of the first AI model and the model output data of the second AI model before the linear combination, or applying an image reconstruction operation to the output material fractionation data after the linear combination, so that the output material fractionation data comprises output material image data.

6. The input spectral X-ray data is projection data, and spectral photon count data c indicates X-ray photon counts detected in each of a plurality of energy bins at an X-ray detector. ｓ 6. The method of claim 2, comprising:

7. The method of claim 1 , wherein the first AI model and the second AI model each receive spectral X-ray image data as model input data and generate material fractionation image data as model output data.

8. The method of claim 1 , wherein the input spectral X-ray data is spectral CT data.

9. The input spectral X-ray data includes spectral photon count data c indicating X-ray photon counts detected in each of a plurality of energy bins at the X-ray detector, for each of a plurality of pixels p, and for each of a plurality of projections v of the input spectral X-ray data. ｓ The method of claim 8 , comprising (p, v).

10. a computer program comprising computer program code, said computer program code being executable on a processor; 10. A computer program, the computer program code of which causes the processor to carry out the method of any one of claims 1 to 9.

11. 1. A processing device for processing X-ray spectral data to derive material fractionation information, the processing device comprising: accessing a data storage device, a first AI model and a second AI model are stored in the data storage device, and each of the first AI model and the second AI model receives spectral X-ray data as model input data and generates material fractionation data as model output data; the first AI model is configured to exhibit lower bias in output material fractionation data values ​​than the second AI model when applied to the same input data set, and to exhibit higher variance in the output material fractionation data values ​​than the second AI model when applied to the same input data set; acquiring input spectral X-ray data; providing the input spectral X-ray data as model input data to the first AI model; providing the input spectral X-ray data as model input data to the second AI model; applying low-pass spatial filtering to the model output data of the first AI model; applying high-pass spatial filtering to the model output data of the second AI model; linearly combining the filtered outputs of the first AI model and the second AI model to derive output material fractionation data; wherein the high-pass spatial filtering is not applied to the model output data of the first AI model and the low-pass spatial filtering is not applied to the model output data of the second AI model; The first AI model and the second AI model are artificial neural networks derived using a supervised training method, the supervised training method comprising: obtaining a preliminary AI model that receives X-ray spectral data as model input data and generates material fractionation data as model output data; A first copy of the preliminary AI model is then transferred to a first training data T １ to obtain the first AI model; A second copy of the preliminary AI model is used as a second training data T ２ to obtain the second AI model; Including, each of the first training data and the second training data includes an input data entry in the form of sample spectral X-ray data and a corresponding output data entry in the form of known material fractionation data for each of the sample spectral X-ray data; The first training data T １ is set to the second training data T such that the first training data is less noisy compared to the second training data. ２ by preprocessing the second training data T ２ wherein the sample spectral X-ray data forming the input data entries is generated from and noise reduced.

12. The processing device of claim 11 , further comprising the data storage device storing the first AI model and the second AI model.

13. an X-ray scanning assembly including an X-ray radiation source and an X-ray radiation detector, wherein a scan region is between the X-ray radiation source and the X-ray radiation detector for receiving an object to be scanned, and an X-ray radiation path between the X-ray radiation source and the X-ray radiation detector passes through the scan region; 13. A processing device according to claim 11 or 12, communicatively coupled to the X-ray radiation detector to receive input spectral X-ray data from the X-ray radiation detector; A spectral X-ray imaging system comprising:

14. 10. A method of training a first AI model and a second AI model for use in the method of any one of claims 1 to 9 for deriving material fractionation data from input X-ray spectral data, wherein the first AI model and the second AI model are artificial neural networks, the method comprising: obtaining a preliminary AI model that receives X-ray spectral data as model input data and generates material fractionation data as model output data; A first copy of the preliminary AI model is then transferred to a first training data T １ to obtain a first AI model; A second copy of the preliminary AI model is used as a second training data T ２ to obtain a second AI model; Including, each of the first training data and the second training data includes an input data entry in the form of sample spectral X-ray data and a corresponding output data entry in the form of known material fractionation data for each of the sample spectral X-ray data; The first training data T １ is set to the second training data T such that the first training data is less noisy compared to the second training data. ２ by preprocessing the second training data T ２ wherein the sample spectral X-ray data, generated from and forming the input data entries, is noise reduced.

Images