Deep learning material decomposition method and device based on physical model, terminal and storage medium

Abstract

This application provides a deep learning-based material decomposition method based on a physical model, comprising: constructing a physical model to map the X-ray response of a CT detector under different physical parameters, and creating a dataset based on the physical model; constructing and training a neural network based on the dataset to fit the CT detector response under different physical parameters; setting up several calibration experiments to determine the physical parameters of the CT detector by reducing the error between the neural network prediction value and the experimental response value; during imaging, calculating the thickness information of the detected material through the neural network based on the physical parameters and actual response of the CT detector. This application constructs a neural network to map the correlation between various physical parameters during the material decomposition process of the CT detector, and obtains the physical parameters of the CT detector through a small number of calibration experiments. Therefore, during material decomposition, the thickness information of the detected material can be inferred from the known physical parameters and response energy spectrum, and the method has good robustness.

Description

Technical Field

[0001] This application relates to the field of CT imaging, and in particular to a deep learning-based material decomposition method, apparatus, terminal, and storage medium based on a physical model. Background Technology

[0002] Computed tomography (CT) is an indispensable diagnostic tool in modern medicine, playing an irreplaceable role in the diagnosis of many diseases. With the development of photon counting detector technology, photon counting spectral CT has received increasing attention. Spectral CT refers to a CT imaging method that utilizes X-ray energy spectrum information. Its purpose is to eliminate the influence of energy in the imaging results by utilizing the differences in the attenuation coefficients of matter under different energy spectra. It can provide material resolution capabilities that traditional CT lacks, which is of great significance for contrast imaging and soft tissue imaging; at the same time, it has lower noise than dual-energy CT, which can reduce the dose received by patients.

[0003] The process of identifying matter using different X-ray energy spectrum information is called matter decomposition. The core of matter decomposition lies in obtaining the thickness information of the object being tested, thereby enabling imaging based on this thickness information to reconstruct the internal structure of the object. However, current photon counting detectors still suffer from some hardware and physical limitations, posing risks to the imaging quality of photon counting CT. For example, signal stacking leads to increased image noise and distortion of the energy spectrum; Compton scattering causes errors in detector energy counting, resulting in errors in the final energy spectrum; charge sharing effects cause adjacent detector units to repeatedly count the same photon, also producing errors; and the escape and reabsorption phenomena of characteristic X-rays also lead to certain errors in the energy spectrum. These physical phenomena all affect the quality of matter decomposition to varying degrees and are problems that urgently need to be solved in current technology.

[0004] There are three main existing methods for decomposing CT materials:

[0005] 1. Image method. The principle is to use traditional image reconstruction algorithms to obtain images under different photon energies and then perform material decomposition to obtain the thickness information of the detected material. This method cannot directly obtain the material decomposition results and has the problems of ray hardening artifacts and amplification noise.

[0006] 2. Model-based method. This method works by establishing a physical model to directly obtain the thickness information of the detected material based on the detector's count values. However, the drawback is that the parameters of the physical model are too complex, requiring extensive calculations and making real-time imaging impossible.

[0007] 3. Calibration method. The principle is to obtain the correlation between the detector's counting response and the thickness information of the detected substance through a large number of calibration experiments and calibration data. Its disadvantage is that the data calibration process is too complicated and inefficient. Summary of the Invention

[0008] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a deep learning-based matter decomposition method, device, terminal and storage medium based on a physical model, to solve the technical problems of complex calculation and complicated calibration in existing matter decomposition methods.

[0009] To achieve the above and other related objectives, the first aspect of this application provides a method comprising: constructing a physical model for mapping the X-ray response of a CT detector under different physical parameters, and creating a dataset based on the physical model; constructing and training a neural network based on the dataset for fitting the detector response under different physical parameters; setting up several calibration experiments to determine the physical parameters of the CT detector by reducing the error between the neural network prediction value and the experimental response value; and during imaging, calculating the thickness information of the detected material through the neural network based on the physical parameters of the CT detector and the actual response of the CT detector.

[0010] In some embodiments of the first aspect of this application, the physical parameters of the CT detector are obtained based on the following calibration experiment: obtaining predicted values ​​of the physical parameters of the CT detector, thickness information of the experimental material, and experimental response energy spectrum of the CT detector; inputting the thickness information of the experimental material and the predicted values ​​of the physical parameters of the CT detector into the pre-trained neural network to obtain the predicted response energy spectrum of the CT detector; calculating the error between the predicted response energy spectrum of the CT detector and the experimental response energy spectrum of the CT detector; if the error is greater than a preset error threshold, updating the predicted values ​​of the physical parameters of the CT detector, and repeating the above steps until the error is less than the preset error threshold; when the error is less than the preset error threshold, setting the physical parameters of the CT detector to be equal to the current predicted values ​​of the physical parameters of the CT detector.

[0011] In some embodiments of the first aspect of this application, the pre-trained neural network is obtained based on the following steps: building a physical model, the physical model being used to map the functional relationship between an initial energy spectrum, physical parameters of a CT detector, thickness information of the material being detected, and the response energy spectrum of the CT detector; generating a dataset based on the physical model, the dataset including physical parameters of different CT detectors under a fixed initial energy spectrum, thickness information of different materials being detected, and the response energy spectra of the CT detectors corresponding to both; and inputting the dataset into an untrained neural network for training the untrained neural network until convergence.

[0012] In some embodiments of the first aspect of this application, the physical parameters of the CT detector include detector thickness, detection efficiency, electric field strength, charge sharing coefficient, chip dead time, chip counting threshold, chip threshold drift coefficient with count rate, and energy resolution.

[0013] In some embodiments of the first aspect of this application, the pre-trained neural network includes a pre-trained first neural network and a pre-trained second neural network; wherein the pre-trained first neural network is used to map the functional relationship between detector thickness, detection efficiency, electric field strength, charge sharing coefficient, thickness information of the detected material, and intermediate energy spectrum of the CT detector, and the pre-trained second neural network is used to map the functional relationship between chip dead time, chip count threshold, chip threshold drift coefficient with count rate, energy resolution, thickness information of the detected material, and final response energy spectrum of the CT detector.

[0014] In some embodiments of the first aspect of this application, the physical model includes a first physical model and a second physical model; wherein the first physical model is used to map the functional relationship between the initial energy spectrum, detector thickness, detection efficiency, electric field strength, charge sharing coefficient, thickness information of the detected material, and the intermediate energy spectrum of the CT detector, and the second physical model is used to map the functional relationship between the initial energy spectrum, chip dead time, chip count threshold, chip threshold drift coefficient with count rate, energy resolution, thickness information of the detected material, and the final response energy spectrum of the CT detector.

[0015] In some embodiments of the first aspect of this application, the untrained neural network includes an untrained first neural network and an untrained second neural network; wherein, when training the untrained neural network, the input of the untrained first neural network is an ideal energy spectrum and detector thickness, detection efficiency, electric field strength, and charge sharing coefficient, and the output is the response energy spectrum of the detector under the detector thickness, detection efficiency, electric field strength, and charge sharing coefficient; the input of the untrained second neural network is the output of the untrained first neural network, chip dead time, chip count threshold, chip threshold drift coefficient with count rate, and energy resolution, and the output is the actual response energy spectrum of the chip.

[0016] To achieve the above and other related objectives, a second aspect of this application provides a deep learning-based material decomposition device based on a physical model, comprising: a parameter acquisition module for acquiring the response energy spectrum of a CT detector and the physical parameters of the CT detector; wherein the physical parameters of the CT detector are obtained by reducing the error between the experimental response energy spectrum and the predicted response energy spectrum of the CT detector through a preset calibration experiment; and a thickness calculation module for calculating the thickness information of the detected material based on the response energy spectrum of the CT detector and the physical parameters of the CT detector, using a pre-trained neural network; wherein the pre-trained neural network is used to map the functional relationship between the physical parameters of the CT detector, the thickness information of the detected material, and the response energy spectrum of the CT detector.

[0017] To achieve the above and other related objectives, a third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method.

[0018] To achieve the above and other related objectives, a fourth aspect of this application provides an electronic terminal, comprising: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the terminal to perform the method.

[0019] As described above, the deep learning-based matter decomposition method, apparatus, terminal, and storage medium based on a physical model of this application have the following beneficial effects:

[0020] This application constructs a neural network to map the complex relationships between various physical parameters during the material decomposition process of a photon-counting CT detector. Through a small number of calibration experiments, the physical parameters of the CT detector are obtained. In practical applications, material decomposition is performed based on the known physical parameters and the trained neural network, enabling rapid acquisition of the most suitable material thickness information. Compared with traditional material decomposition methods, this invention significantly reduces the required computational resources and greatly improves operational efficiency, facilitating flexible deployment in practical applications. It is important to emphasize that this application eliminates the need for the complex calibration procedures of traditional material decomposition methods. Ideal CT detector physical parameters can be obtained through only a few sets of calibration experiments, resulting in a significant improvement in the overall efficiency of material decomposition and a high degree of accuracy, demonstrating promising application prospects. Attached Figure Description

[0021] Figure 1 The diagram shown is a flowchart illustrating a deep learning-based matter decomposition method based on a physical model, according to one embodiment of this application.

[0022] Figure 2The diagram shown is a schematic flowchart of a physical parameter calibration method for a CT detector in one embodiment of this application.

[0023] Figure 3 The diagram shown is a schematic flowchart of a pre-trained neural network training method in one embodiment of this application.

[0024] Figure 4 The diagram shown is an initial energy spectrum schematic diagram in one embodiment of this application.

[0025] Figure 5 The diagram shown is a schematic diagram of the attenuation energy spectrum after X-rays interact with matter in one embodiment of this application.

[0026] Figure 6 The diagram shows the energy spectrum after X-rays interact with the detector in one embodiment of this application.

[0027] Figure 7 The diagram shown is a schematic representation of the entire detector response process in one embodiment of this application.

[0028] Figure 8 The diagram shown is a schematic representation of the detector signal over a period of time in one embodiment of this application.

[0029] Figure 9 The diagram shown is a schematic representation of the internal structure of an ASIC according to an embodiment of this application.

[0030] Figure 10 The diagram shown is a schematic representation of the response energy spectrum in one embodiment of this application.

[0031] Figure 11 The diagram shown is a schematic representation of the energy change from the initial energy spectrum to the response energy spectrum in one embodiment of this application.

[0032] Figure 12 The diagram shown is a schematic diagram of the first neural network structure in one embodiment of this application.

[0033] Figure 13 The diagram shown is a schematic representation of the fitting result of the first neural network in one embodiment of this application.

[0034] Figure 14 The diagram shown is a schematic diagram of the second neural network structure in one embodiment of this application.

[0035] Figure 15 The diagram shown is a schematic representation of the fitting result of the second neural network in one embodiment of this application.

[0036] Figure 16 The diagram shown is a schematic representation of the overall structure of the first neural network and the second neural network in one embodiment of this application.

[0037] Figure 17The diagram shown is a schematic representation of the fitting results of the first neural network and the second neural network in one embodiment of this application.

[0038] Figure 18 The diagram shown is a schematic representation of the entire process of material decomposition in one embodiment of this application.

[0039] Figure 19 The diagram shown is a structural schematic of an electronic terminal according to an embodiment of this application.

[0040] Figure 20 The diagram shown is a structural schematic of a deep learning-based matter decomposition device based on a physical model, according to one embodiment of this application. Detailed Implementation

[0041] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0042] Furthermore, as used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of the stated feature, operation, element, component, item, kind, and / or group, but do not preclude the presence, occurrence, or addition of one or more other features, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are interpreted as inclusive, or mean any one or any combination thereof. Thus, “A, B, or C” or “A, B, and / or C” means “any one of: A; B; C; A and B; A and C; B and C; A, B, and C.” Exceptions to this definition arise only when combinations of elements, functions, or operations are inherently mutually exclusive in some manner.

[0043] To address the problems described in the background section, this invention provides a deep learning-based matter decomposition method, apparatus, terminal, and storage medium based on a physical model, aiming to solve the problems of computational complexity and cumbersome calibration in existing matter decomposition methods. Meanwhile, to make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention are further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention.

[0044] Before providing a further detailed description of the present invention, the nouns and terms used in the embodiments of the present invention are explained, and the nouns and terms used in the embodiments of the present invention are subject to the following interpretations:

[0045] <1> Computed tomography (CT), also known as computed tomography, is an imaging diagnostic procedure. This technique was formerly called computed axial tomography. It uses computer processing to combine numerous X-ray measurements taken from different angles to generate cross-sectional (tomographic) images (virtual "slices") of a specific scanned area, allowing users to see the interior of an object without cutting it.

[0046] <2> The photon counting detector (PCD) is crucial in energy-dispersive CT imaging technology, which has seen rapid development over the past decade. It can detect the attenuation characteristics of X-rays at different energies, overcoming the limitations of traditional X-ray CT imaging and facilitating qualitative and quantitative analysis of the detected object. The PCD can detect continuous multicolor X-ray spectra separately according to different energy ranges, obtaining X-ray photons within specific energy ranges through energy thresholds, and then reconstructing information-rich CT images based on the attenuation characteristics of X-rays at different energies.

[0047] <3> Artificial Neural Networks (ANNs), also known as Neural Networks (NNs) or neural network-like systems, are mathematical or computational models in machine learning and cognitive science that mimic the structure and function of biological neural networks (the central nervous system of animals, especially the brain). They are used to estimate or approximate functions. Neural networks consist of a large number of interconnected artificial neurons performing computations. In most cases, artificial neural networks can change their internal structure based on external information; they are adaptive systems, meaning they possess learning capabilities.

[0048] This invention provides a physical model-based deep learning matter decomposition method, an apparatus for implementing the physical model-based deep learning matter decomposition method, a storage medium, and an electronic terminal storing an executable program for implementing the physical model-based deep learning matter decomposition method. Regarding the implementation of the physical model-based deep learning matter decomposition method, this invention will describe exemplary implementation scenarios.

[0049] like Figure 1The diagram illustrates a flowchart of a deep learning-based matter decomposition method using a physics model, as shown in this embodiment of the invention. The deep learning-based matter decomposition method using a physics model in this embodiment mainly includes the following steps:

[0050] Step S11: Obtain the response energy spectrum and physical parameters of the CT detector; wherein, the physical parameters of the CT detector are obtained by reducing the error between the experimental response energy spectrum and the predicted response energy spectrum of the CT detector through a preset calibration experiment.

[0051] Specifically, steps S11 to S12 describe the practical application process of the material decomposition method of the present invention. The response energy spectrum of the CT detector is the final response energy spectrum of the detector chip in actual application, and is a known quantity. The physical parameters of the CT detector include, but are not limited to, the detector thickness, detection efficiency, detector electric field strength, charge sharing coefficient, chip dead time, chip counting threshold, chip threshold drift with count rate, and detector energy resolution listed in this embodiment. The physical parameters of the CT detector are also known quantities in step S11, and their calibration method is described in steps S21 to S24 below.

[0052] Step S12: Based on the response energy spectrum and physical parameters of the CT detector, the thickness information of the detected material is calculated using a pre-trained neural network; the pre-trained neural network is used to map the functional relationship between the physical parameters of the CT detector, the thickness information of the detected material, and the response energy spectrum of the CT detector.

[0053] Specifically, the pre-trained neural network described in this embodiment can accurately map the functional relationship between the physical parameters of the CT detector, the thickness information of the detected material, and the response energy spectrum of the CT detector. The construction and training process of this pre-trained neural network is described in steps S31 to S33 below. Therefore, given that the physical parameters and response energy spectrum of the CT detector are known, another unknown parameter, namely the thickness information of the detected material, can be obtained based on the pre-trained neural network, thereby completing the material decomposition.

[0054] For example, when a neural network fits the functional relationship y = f(x, a, b, c) between the physical parameters of a CT detector, the thickness information of the material being detected, and the response energy spectrum of the CT detector, then by fixing the physical parameters a, b, c of the CT detector and the response energy spectrum y of the CT detector, the corresponding x can be derived, where x is the thickness information of the material being detected.

[0055] It should be noted that the application of neural networks in this invention is not a traditional input-output method, but rather utilizes the characteristic of neural networks to map nonlinear relationships between arbitrary data. Therefore, given the other known parameters in the neural network, the value of the remaining unknown parameter can be deduced from the trained neural network.

[0056] like Figure 2 The diagram shown is a flowchart of the physical parameter calibration experiment for a CT detector in one embodiment of this application. In some implementations of this embodiment, the physical parameters of the CT detector in step S11 are obtained based on the following calibration experiment:

[0057] Step S21: Obtain the predicted values ​​of the physical parameters of the CT detector, the thickness information of the experimental material, and the experimental response energy spectrum of the CT detector.

[0058] It should be noted that the initial values ​​of the predicted physical parameters of the CT detector are preferably empirical values, while the thickness information of the experimental material and the experimental response energy spectrum of the CT detector are known parameters in the calibration experiment.

[0059] Step S22: Input the thickness information of the experimental material and the predicted values ​​of the physical parameters of the CT detector into the pre-trained neural network to obtain the predicted response energy spectrum of the CT detector.

[0060] Specifically, the pre-trained neural network can accurately map the functional relationship between the physical parameters of the CT detector, the thickness information of the material being detected, and the response energy spectrum of the CT detector. Therefore, when the thickness information of the experimental material and the predicted values ​​of the physical parameters of the CT detector are known, the predicted response energy spectrum of the CT detector output by the neural network can be obtained. The construction and training process of this pre-trained neural network is described in steps S31 to S33 below.

[0061] Step S23: Calculate the error between the predicted response energy spectrum of the CT detector and the experimental response energy spectrum of the CT detector. If the error is greater than a preset error threshold, update the predicted value of the physical parameters of the CT detector and repeat the above steps until the error is less than the preset error threshold.

[0062] Specifically, after obtaining the predicted response energy spectrum of the CT detector from the neural network output, the error between the predicted response energy spectrum and the experimental response energy spectrum of the CT detector can be calculated. By pre-setting an error threshold, if the error between the predicted and experimental response energy spectra of the CT detector is greater than the error threshold, the predicted values ​​of the physical parameters of the CT detector are updated using the backpropagation algorithm, and the above steps S21 to S23 are repeated until the error between the predicted and experimental response energy spectra of the CT detector is less than the error threshold.

[0063] Step S24: When the error is less than a preset error threshold, set the physical parameters of the CT detector to the predicted values ​​of the current physical parameters of the CT detector.

[0064] Specifically, when the error between the predicted response energy spectrum of the CT detector and the experimental response energy spectrum of the CT detector is less than the error threshold, it indicates that the predicted value of the physical parameters of the CT detector has reached the ideal range, and it is used as the physical parameters of the CT detector in step S11 for the actual material decomposition process.

[0065] like Figure 3 The diagram shown illustrates a flowchart of a training method for a pre-trained neural network in one embodiment of this application. In some implementations of this embodiment, the pre-trained neural network described in step S12 is obtained based on the following steps:

[0066] Step S31: Build a physical model, which is used to map the functional relationship between the physical parameters of the CT detector, the thickness information of the detected material, and the response energy spectrum of the CT detector under a fixed initial energy spectrum;

[0067] Specifically, the physical model can be built using simulation software. Its essence lies in mapping the functional relationship between the physical parameters of the CT detector, the thickness information of the detected material, and the response energy spectrum of the CT detector under a fixed initial energy spectrum.

[0068] In the process of X-rays changing from the initial energy spectrum to the response energy spectrum, X-rays undergo multiple attenuation processes, mainly including attenuation by matter, interaction between X-rays and the detector, charge transport, and chip response. The attenuation principle of X-rays in each process and the corresponding physical model design method are explained below:

[0069] 1) Attenuation model of X-rays and matter

[0070] After generating an X-ray energy spectrum, the X-rays pass through the human body and attenuate. The main interactions between X-rays and matter are: photoelectric effect, Compton scattering, pair production, and Rayleigh scattering.

[0071] Regarding the attenuation model of X-rays and matter, an initial energy spectrum was generated using Spektr, and then filtered to remove the low-energy portion of the spectrum. The attenuated X-ray energy spectrum was calculated using the matter attenuation coefficients provided by the NISTXCOM database. Figure 4 As shown, the initial energy spectrum is displayed; as... Figure 5 As shown, the attenuation spectrum is the result of X-rays interacting with matter.

[0072] 2) Interaction model between X-rays and detectors

[0073] After X-rays are attenuated by matter, they interact with the detector through photoelectric effect, Compton scattering, electron-pair effect, and Rayleigh scattering.

[0074] The interaction model between X-rays and the detector was simulated using the GATE tool. This simulation depicted the position and energy deposition of X-rays of different energies after they entered the detector, as well as the escape, reabsorption, and charge sharing effects of characteristic X-rays. Figure 6 As shown, this is the energy spectrum after the X-rays interact with the detector.

[0075] 3) Charge transport model

[0076] When X-rays pass through matter and reach the detector, the detector's semiconductor material converts the photons into free electron-hole pairs. Due to the presence of a strong electric field, the electrons and holes move towards their respective cathodes; this movement is the transfer of electric charge.

[0077] During the transport process, charge is affected by drift, self-repulsion, thermal diffusion, and capture.

[0078] The entire transmission process can be expressed by the following formula:

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] Where n f n represents the spatial distribution of free electrons. t j represents the spatial distribution of the trapped electrons. n D represents electron flux. n Represents the electron drift constant, μ n Represents electron mobility. Represents voltage. p f p represents the spatial distribution of free holes t j represents the spatial distribution of captured holes pD represents hole flux. p Represents the hole drift constant, μ p This represents hole mobility.

[0086] Equations 1, 2, and 3 represent the electron transport equations, while equations 4, 5, and 6 represent the hole transport equations. Equations 1 and 4 express the spatial distribution of free electrons and free holes over time; the rate of change of the spatial distribution of free electrons over time is equal to the divergence of the electron flux minus the electron capture coefficient. Equations 3 and 6 express the spatial distribution of captured electrons over time; the electron capture rate also increases with time. Equations 2 and 5 represent the electron flux and hole flux. The electron flux consists of two parts: thermal drift and diffusion of electrons under an electric field.

[0087] Regarding the charge transport model, since calculating formulas 1-6 is very time-consuming, this invention uses empirical formulas for approximation to calculate the collecting charge q at the anode. ie For example.

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] Among them, Q pair It is the amount of electron-hole pair charge generated during the entire detector response process, σ t It is the size of the electron cloud when electrons reach the anode, σ0 is the size of the initial charge cloud, and D n It is the electron drift coefficient.

[0095] The electron cloud of the charge is approximated as a two-dimensional Gaussian distribution, whose standard deviation changes over time. Equation 8 represents the distribution of the electron cloud, Equation 9 represents the change of standard deviation over time, Equation 10 represents the charge quantity, Equation 11 represents the initial standard deviation, Equation 12 represents the approximate time for the charge to reach the anode, and Equation 7 represents the collection charge q at the anode obtained by integrating over space. ie .

[0096] 4) Signal generation model

[0097] As the charge moves towards the two ends of the electrode, a signal is generated. Based on the Shockley-Ramo theory, the weighting potential is calculated. By combining the weighting potential with the charge, the detector's response signal can be obtained.

[0098] In the Shockley-Ramon theorem, the signal generated by the electron cloud can be detected by the electrodes of the detector. The generation of the signal can be divided into two parts. Taking electrons as an example, the charges are divided into two categories. The first category is the charge q that moves directly to the anode. c The second type is the induced charge q generated during electron transport. i :

[0099] q t =q c +q i (Formula 13)

[0100]

[0101] q i =∫∫∫ V qψd 3 r (Formula 15)

[0102] Formula 13 represents a total charge of q i With q c The sum of the charges. Formula 14 represents the charge moving to the anode, and Formula 15 represents the induced charge during the transfer process.

[0103]

[0104]

[0105] Where ψ represents the weighting potential. Equation 17 indicates that at a specific anode, its weighting potential is 1, and all others are 0. Equation 16 indicates that the second derivative of the weighting potential is 0.

[0106] Since the above formula is too complicated, the present invention adopts the following empirical formula:

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113] Based on the above empirical formulas 18 to 23, the computational load can be significantly reduced and the calculation result error is small. Furthermore, by combining this with the Shockley-Ramo theorem, the weighted potential ψ(x, y, z) at any position of the detector can be calculated.

[0114] By combining the Shockley-Ramo theorem and the charge transport model described in 3), the charges generated by the anode and cathode can be calculated.

[0115] The induced charge Q generated by an electron at time t e for:

[0116] Q e =q ie ψ(x(t),y(t),z(t))(Formula 24)

[0117] Since the size of a detector pixel is much larger than the size of an electron, formula (Formula 24) can be simplified to:

[0118] Q e =q ie ψ(0, 0, z(t))(Formula 25)

[0119] The induced charge Q generated by the hole at time t h for:

[0120] Q h =q ih ψ(x, y, z(t))(Formula 26)

[0121] Since the size of a hole is much larger than the size of an electron, a similar approximation cannot be made here. However, for the sake of calculation convenience, x and y can be regarded as approximate values.

[0122] Combining formulas (Formula 25) and (Formula 26), the total induced charge Q at the anode can be obtained. s :

[0123] Q s (t)=Q e (t)-Q h (t)=q ie ψ(0, 0, z) e (t))-q ih ψ(x, y, z) h (t))(Formula 27)

[0124] z e (t)=z0-E f μ e t(Formula 28)

[0125] z h (t)=z0+E f μ h t(Formula 29)

[0126] Among them, z e (t) and z h (t) represents the positions of the free electron and free charge in the z-direction, respectively, and μ e μ represents electron mobility. h E represents the mobility of holes. f It represents a strong electric field.

[0127] After obtaining the induced charge, the induced current I can be obtained by taking the time derivative of the induced charge. i (t):

[0128]

[0129] The induced current is the signal received by the detector.

[0130] Regarding the signal generation model, this invention employs the following simulation steps: Based on the Shockley-Ramo theorem, the weighted potentials at each location of the detector are pre-calculated. Then, the locations of electrons and holes, and the induced charges generated at those locations, are calculated separately. Multiplying these by the weighted potentials yields the total induced charge of the detector. Differentiating the induced charge with respect to time yields the induced current, which is the final response signal of the detector. The entire detector response process is as follows: Figure 7 As shown. Figure 8 As shown, this represents the detector signal generated over a period of time.

[0131] 5) Chip Response Model

[0132] After the detector generates a response signal, the signal enters the chip module. The chip analyzes and counts the signal to obtain the final count of the photon counting CT. The chip response steps include current integration, high-pass filtering, timing judgment, difference calculation, and comparison counting.

[0133] like Figure 9The diagram shows the internal structure of an ASIC. The ASIC chip consists of several modules: current integration, high-pass filtering, timing determination, difference calculation, and comparison counting. When the detector generates a signal, the ASIC passes it through a preamplifier to obtain a voltage signal. This voltage is then high-pass filtered, allowing only signals above a certain threshold to pass through, thus removing noise. When the ASIC is in an active state, it counts the energy range of the signal, completing one cycle. The chip then enters a paralyzed state for approximately 16 ns to prevent signal saturation. During this paralyzed state, all signals arriving at the ASIC are not counted, leading to energy spectrum distortion. Furthermore, when two signals arrive simultaneously, the ASIC struggles to distinguish them, treating them as a single signal, resulting in energy spectrum stacking. Additionally, errors in threshold determination can also cause errors in the final energy spectrum count.

[0134] When the induced current of the detector passes through the current integration module, the current is integrated to output a voltage signal:

[0135]

[0136] Where I(t) is the detector induced current, C f This is the value of the feedback capacitor.

[0137] The obtained voltage signal is then subjected to high-pass filtering.

[0138] Suppose that at time T0, the high-pass filtered signal exceeds a set threshold voltage, and the effective signal of the voltage signal at this time is V0. Here, T1, T2, and V2 are defined as follows:

[0139] T1 = T0 + ΔT deadtime (Formula 32)

[0140] T2 = T1 + ΔT reset =T0+ΔT deadtime +ΔT reset (Formula 33)

[0141] V1=V(T1) (Formula 34)

[0142] ΔV=V1-V0 (Formula 35)

[0143] ΔV will be used for photon energy comparison.

[0144] Before the ASIC actually starts working, it undergoes a calibration process. This involves calculating the relationship between photons at different energy levels and the threshold voltage V. threshold The relationship between them.

[0145] After energy calibration, the theoretical voltage of the photon at each energy level can be obtained. Based on this theoretical voltage, the ASIC can classify the obtained ΔV. For example, if ΔV falls within the voltage energy range corresponding to a 30keV photon, the count in this range is incremented by one.

[0146] Thus, the active reset circuit model was completed, a signal accumulation model was established, and photon counting was performed on X-rays in different energy ranges.

[0147] The response process of the above chip was simulated, and the results are as follows: Figure 10 The chip response energy spectrum is shown.

[0148] like Figure 11 The diagram shows the change from the initial energy spectrum to the response energy spectrum. It is evident that the energy spectrum undergoes significant changes from the initial energy spectrum to the final chip response energy spectrum. Therefore, a method is needed to map the relationship between the initial and response energy spectra under different material thicknesses, different charge sharing coefficients, different chip dead times, and different chip counting thresholds.

[0149] The entire physical model, from the initial energy spectrum to the final response energy spectrum of the detector chip, can be divided into the following steps:

[0150]

[0151]

[0152] S3(E)=DS2(E)(Formula 38)

[0153] S4(E)=f1(S3(E),σ)(Formula 39)

[0154] S5(E)=f2(S4(E), Nτ, Q) trigger )(Formula 40)

[0155]

[0156] Where S0 is the initial energy spectrum, S1 is the energy spectrum after low-energy filtering, S2 is the energy spectrum after material attenuation, S3 is the energy spectrum of the ideal detector response, S4 is the energy spectrum of the actual detector response, S5 is the ASIC response energy spectrum, N represents the total count, and μ a μ b The attenuation coefficient representing low-energy X-ray filtration, l a l bμ1 and μ2 represent the thickness of the filter material, l1 and l2 represent the attenuation coefficients of the two substances, D represents the theoretical response matrix of the detector, σ represents the charge sharing coefficient, and f represents the attenuation coefficient of the two substances. i (:) represents the neural network, ΔE represents the energy spectrum shift, τ represents the detector dead time, and Q trigger N represents the chip counting threshold. i ′ represents the detector's count in the i-th energy range, E i This represents the starting energy of the i-th energy range.

[0157] In this equation, Equation 36 represents the initial energy spectrum after low-energy filtering, Equation 37 represents the energy spectrum after human attenuation, Equation 38 represents the ideal response of the detector, Equation 39 represents the neural network fitting the non-ideal situation, Equation 40 represents the neural network simulating the chip response, and Equation 41 represents the actual count of the detector. Using Equations 36 to 41, a deep learning-based model can be established to fit the detector response.

[0158] In some implementations of this embodiment, the physical model includes a first physical model and a second physical model; wherein, the first physical model is used to map the functional relationship between the initial energy spectrum, detector thickness, detection efficiency, electric field strength, charge sharing coefficient, thickness information of the detected material, and the intermediate energy spectrum of the CT detector, and the second physical model is used to map the functional relationship between the initial energy spectrum, chip dead time, chip count threshold, chip threshold drift coefficient with count rate, energy resolution, thickness information of the detected material, and the final response energy spectrum of the CT detector.

[0159] Specifically, the physical model can be composed of a first physical model and a second physical model. The first physical model, namely the detector response model, consists of the aforementioned 1) X-ray and matter attenuation model, 2) X-ray and detector interaction model, 3) charge transport model, and 4) signal generation model. It is used to map the functional relationship between the initial energy spectrum, detector thickness, detection efficiency, electric field strength, charge sharing coefficient, thickness information of the detected material, and the intermediate energy spectrum of the CT detector. The second physical model, namely the chip response model, corresponds to the aforementioned 5) chip response model. It is used to map the functional relationship between the initial energy spectrum, chip dead time, chip count threshold, chip threshold drift coefficient with count rate, energy resolution, thickness information of the detected material, and the final response energy spectrum of the CT detector.

[0160] Step S32: Generate a dataset based on the physical model. The dataset includes the physical parameters of different CT detectors under a fixed initial energy spectrum, the thickness information of different detected materials, and the response energy spectrum of the CT detectors corresponding to the two.

[0161] Based on the above physical model, a large dataset can be generated, which includes the physical parameters of different CT detectors under a fixed initial energy spectrum, the thickness information of different detected materials, and the response energy spectrum of the CT detectors corresponding to both.

[0162] In some implementations of this embodiment, a first dataset and a second dataset are generated using a first physical model and a second physical model, respectively. The first dataset includes detector response data for a single-energy spectrum and detector response data for a full-energy spectrum; the second dataset includes chip response data for a full-energy spectrum. The specific generation process of the datasets is described in step S33 below.

[0163] Step S33: Input the dataset into an untrained neural network to train the untrained neural network until it converges.

[0164] In some implementations of this embodiment, the untrained neural network includes an untrained first neural network and an untrained second neural network. When training the untrained neural network, the input to the untrained first neural network is an ideal energy spectrum, detector thickness, detection efficiency, electric field strength, and charge sharing coefficient, and the output is the detector's response energy spectrum under the given detector thickness, detection efficiency, electric field strength, and charge sharing coefficient. The input to the untrained second neural network is the output of the untrained first neural network, chip dead time, chip counting threshold, chip threshold drift coefficient with count rate, and energy resolution, and the output is the chip's actual response energy spectrum.

[0165] Specifically, the inputs to the first neural network are the ideal energy spectrum, the charge sharing coefficient σ, the detector thickness, the detection efficiency, and the detector electric field strength. The output is the detector response when the charge sharing coefficient is σ, the detector thickness, the detection efficiency, and the detector electric field strength. In short, the goal of the first neural network is to map the ideal energy spectrum to the corresponding energy spectrum when the charge sharing coefficient is given, detector thickness, detection efficiency, and detector electric field strength.

[0166] like Figure 12 The diagram shows the structure of the first neural network. In terms of structure, the first neural network uses a three-layer fully connected neural network with dimensions of 242×200, 200×200, and 200×241, and uses tanh(x) as the activation function. The loss function used is mean squared error.

[0167] Regarding the creation of the first neural network dataset, based on the physical model in step S31, 4000 sets of detector response data with monoenergetic spectra and 4720 sets of detector response data with omnienergetic spectra were generated; each set of data simulated the interaction between the particle and the detector (over 1 million times). Among them, 6104 sets of data were used as training data, 1308 sets of data were used as validation data, and 1308 sets of data were used as test data.

[0168] Regarding the model performance of the first neural network, the following table shows the fitting results of the first neural network to the detector response.

[0169]

[0170] Table 1

[0171] like Figure 13 The diagram shows the fitting results of the first neural network to the detector response. (a) and (b) show the fitting results of the neural network for the full-energy spectrum, and (c) shows the fitting results of the neural network for the single-energy spectrum. In (a), (b), and (c), the top energy spectrum is the input energy spectrum, the middle is the neural network output energy spectrum, and the bottom is the ideal energy spectrum. (From Table 1 and...) Figure 13 It can be observed that the neural network model can fit the physical model of the detector response very well.

[0172] The inputs to the second neural network are the output energy spectrum of the detector response model of the first neural network, the detector dead time τ, and the chip counting threshold Q. trigger The chip threshold drift coefficient with count rate and detector energy resolution are considered, and the output is the chip's response at that time. In short, the goal of the second neural network is to achieve the chip's response given τ and Q. trigger Given the chip threshold drift coefficient with count rate and detector energy resolution, the ideal energy spectrum is mapped to the corresponding energy spectrum at this time.

[0173] like Figure 14 The diagram shows the structure of the second neural network. In terms of structure, the second neural network uses a four-layer fully connected neural network with dimensions of 243×300, 300×300, 300×300, and 300×241, and uses tanh(x) as the activation function. The loss function used is mean squared error.

[0174] Regarding the creation of the second neural network dataset, based on the physical model in step S31, 1245 sets of chip response data with full-energy spectra were generated. The dataset simulated different X-ray operating currents (10mA, 50mA, and 200mA), different material thicknesses, different chip dead times (13ns to 21ns), and different chip counting thresholds (16keV to 24keV). Of these, 873 sets were used as training data, 186 sets as validation data, and 186 sets as test data.

[0175] Regarding the model performance of the second neural network, the following table shows the fitting results of the second neural network to the detector response.

[0176]

[0177] Table 2

[0178] like Figure 15 As shown, this is a schematic diagram illustrating the fitting results of the second neural network to the chip response. (b) shows the fitting results of the neural network to the full-energy spectrum. (a) is the input energy spectrum, the middle one is the neural network output energy spectrum, and (c) is the ideal energy spectrum. From Table 2 and... Figure 15 It can be observed that the second neural network can fit the chip response physical model very well.

[0179] It should be noted that the reason for using multiple physical models and multiple neural networks in this embodiment to map the nonlinear relationship between different physical parameters is that building neural networks for each of the many physical parameters of the CT detector allows the neural networks to learn the characteristics and relationships between each physical parameter more accurately and efficiently. Furthermore, since the principles by which each physical parameter causes changes in the energy spectrum are different, training multiple neural networks separately and then integrating them to map the nonlinear relationship between the thickness information of the material, the various physical parameters of the CT detector, and the response energy spectrum can enhance the generalization ability of the neural network and enable it to learn the intrinsic relationship between each parameter more accurately.

[0180] like Figure 16 The diagram shows the overall structure of the first and second neural networks. It should be noted that in the diagram, super-parameter group 1 refers to detector thickness, detection efficiency, detector electric field strength, and charge sharing coefficient; super-parameter group 2 refers to chip dead time, chip counting threshold, chip threshold drift with count rate, and detector energy resolution; the input energy spectrum is the initial energy spectrum of the CT detector; and the output energy spectrum is the final response energy spectrum of the CT detector. These concepts have a one-to-one correspondence and should not be ambiguous.

[0181] The loss function expression for the first neural network is:

[0182]

[0183] The loss function expression for the second neural network is:

[0184]

[0185] like Figure 17 This displays the overall fitting effect of the neural network on the physical model. The fitting results are as follows: Figure 18 As shown in the figure, (a) is the prediction result of the neural network, and (b) is the actual value of the system. The overall fitting result of the neural network basically meets expectations and can effectively capture useful information.

[0186] In summary, the complete process of this application can be described as follows: Figure 18 As shown.

[0187] Step S181: Build a physical model to map the functional relationship between the initial energy spectrum of the CT detector, physical parameters, thickness information of the detected material, and response energy spectrum.

[0188] Specifically, a full-process simulation of the photon counting detector is performed, which can simulate phenomena such as charge sharing and signal stacking. After the simulation program is completed, a large number of simulations are conducted to obtain different system response results with different charge sharing coefficients, different detector dead times, and different chip start-up energies.

[0189] Step S182: Generate a dataset based on the physical model, and build and train a neural network to fit the physical model.

[0190] This neural network incorporates multiple physical parameters: detector thickness, detection efficiency, electric field strength, charge sharing coefficient, chip dead time, chip counting threshold, chip threshold drift coefficient with count rate, and energy resolution, and uses deep learning to fit the response of the entire system.

[0191] Step S183: Set up several calibration experiments to determine the physical parameters of the CT detector by reducing the error between the neural network prediction value and the experimental response value.

[0192] Once the neural network model can obtain relatively accurate fitting results, the physical parameters of the CT detector can be deduced by fixing the neural network parameters and the thickness information of the detected material through several sets of calibration experiments.

[0193] Step S184: During imaging, the thickness information of the detected material is calculated based on the physical parameters of the CT detector and the actual response of the CT detector using a neural network.

[0194] Once the physical parameters of the CT detector are determined and the neural network is trained, the most suitable material thickness information can be obtained based on the physical parameters of the CT detector, the neural network, and the actual response energy spectrum, thus completing the material decomposition.

[0195] The advantages of this method compared to traditional material decomposition methods are as follows: traditional methods, such as calibration methods, require fitting a large number of free variables, while this method only needs to fit a small number of physical parameters after determining the neural network parameters; traditional methods involve extremely complex calibration operations, while this method only requires a few calibration operations to determine the physical parameters; traditional methods, such as model methods, involve complex model calculations, while this method does not require complex calculations to obtain the thickness information of the material.

[0196] Although the steps in the above embodiments are described in the above order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not need to be executed in such order. They can be executed simultaneously (in parallel) or in reverse order. These simple variations are all within the protection scope of this invention.

[0197] The physical model-based deep learning matter decomposition method provided in this invention can be implemented on the terminal side or the server side. For the hardware structure of the physical model-based deep learning matter decomposition terminal, please refer to [link to relevant documentation]. Figure 19 This is a schematic diagram of an optional hardware structure of a physical model-based deep learning matter decomposition terminal 1900 provided in an embodiment of the present invention. The terminal 1900 can be a mobile phone, computer device, tablet device, personal digital processing device, factory back-end processing device, etc. The physical model-based deep learning matter decomposition terminal 1900 includes: at least one processor 1901, a memory 1902, at least one network interface 1904, and a user interface 1906. The various components in the device are coupled together through a bus system 1905. It is understood that the bus system 1905 is used to realize the connection and communication between these components. In addition to a data bus, the bus system 1905 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 19 The general will label all buses as bus systems.

[0198] The user interface 1906 may include a monitor, keyboard, mouse, trackball, clicker, button, touchpad, or touch screen.

[0199] It is understood that memory 1902 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM) or programmable read-only memory (PROM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memories described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable categories of memory.

[0200] In this embodiment of the invention, the memory 1902 is used to store various types of data to support the operation of the physics-based deep learning matter decomposition terminal 1900. Examples of this data include: any executable program for operation on the physics-based deep learning matter decomposition terminal 1900, such as operating system 19021 and application program 19022; operating system 19021 includes various system programs, such as framework layer, core library layer, driver layer, etc., for implementing various basic services and handling hardware-based tasks. Application program 19022 may include various applications, such as media player, browser, etc., for implementing various application services. The implementation of the physics-based deep learning matter decomposition method provided in this embodiment of the invention can be included in application program 19022.

[0201] The methods disclosed in the above embodiments of the present invention can be applied to or implemented by processor 1901. Processor 1901 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in processor 1901 or by instructions in the form of software. The processor 1901 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 1901 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. General-purpose processor 1901 may be a microprocessor or any conventional processor, etc. The steps of the accessory optimization method provided in the embodiments of the present invention can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software module may be located in a storage medium, which is located in memory. The processor reads the information in the memory and combines it with its hardware to complete the steps of the aforementioned method.

[0202] In an exemplary embodiment, the physical model-based deep learning matter decomposition terminal 1900 may be used by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), or complex programmable logic devices (CPLDs) to perform the aforementioned method.

[0203] like Figure 20 The diagram illustrates a structural schematic of a physical model-based deep learning matter decomposition device according to an embodiment of the present invention. In this embodiment, the physical model-based deep learning matter decomposition device 2000 includes: a parameter acquisition module 2001, used to acquire the response energy spectrum of a CT detector and the physical parameters of the CT detector; wherein the physical parameters of the CT detector are obtained by reducing the error between the experimental response energy spectrum and the predicted response energy spectrum of the CT detector through a preset calibration experiment; and a thickness calculation module 2002, which calculates the thickness information of the detected material based on the response energy spectrum and the physical parameters of the CT detector using a pre-trained neural network; the pre-trained neural network is used to map the functional relationship between the physical parameters of the CT detector, the thickness information of the detected material, and the response energy spectrum of the CT detector.

[0204] It should be noted that the physical model-based deep learning matter decomposition device provided in the above embodiments is only illustrated by the division of the above-described program modules when performing physical model-based deep learning matter decomposition. In practical applications, the above processing can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program modules to complete all or part of the processing described above. In addition, the physical model-based deep learning matter decomposition device and the physical model-based deep learning matter decomposition method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0205] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented using computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0206] In the embodiments provided in this application, the computer-readable and writable storage medium may include read-only memory, random access memory, EEPROM, CD-ROM or other optical disc storage devices, disk storage devices or other magnetic storage devices, flash memory, USB flash drive, portable hard drive, or any other medium capable of storing desired program code in the form of instructions or data structures and accessible by a computer. Additionally, any connection may be appropriately referred to as a computer-readable medium. For example, if instructions are transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of the medium. However, it should be understood that computer-readable and writable storage media and data storage media do not include connections, carrier waves, signals, or other transient media, but are intended for non-transient, tangible storage media. The disks and optical discs used in the application include compact discs (CDs), laser discs, optical discs, digital multifunction discs (DVDs), floppy disks, and Blu-ray discs, where disks typically copy data magnetically, while optical discs use lasers to copy data optically.

[0207] In summary, this application provides a physical model-based deep learning method, apparatus, terminal, and medium for matter decomposition. This invention offers a method to improve the efficiency of physical model-based deep learning matter decomposition, addressing the problems of computational complexity and cumbersome calibration in existing matter decomposition methods. Therefore, this application effectively overcomes the various shortcomings of the prior art and has high industrial applicability.

[0208] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A deep learning material decomposition method based on a physical model, characterized in that, include: The response energy spectrum and physical parameters of the CT detector are obtained; wherein, the physical parameters of the CT detector are obtained by reducing the error between the experimental response energy spectrum and the predicted response energy spectrum of the CT detector through a preset calibration experiment. Based on the response energy spectrum and physical parameters of the CT detector, the thickness information of the detected material is calculated using a pre-trained neural network. The pre-trained neural network is used to fit a physical model, which maps the functional relationship between the initial energy spectrum, the physical parameters of the CT detector, the thickness information of the detected material, and the response energy spectrum of the CT detector. The physical model includes: ； in, For the initial energy spectrum, The energy spectrum is after low-energy filtering. The energy spectrum after material decay is shown. The energy spectrum represents the ideal response of the detector. This represents the energy spectrum of the detector's true response. For the ASIC response energy spectrum, Indicates the total count. , Represents the attenuation coefficient for low-energy X-ray filtration. , This represents the thickness of the filter material. , Represents the attenuation coefficients of the two substances. , Represents the thickness of the two substances. The theoretical response matrix representing the detector. Represents the charge sharing coefficient. Represents neural networks, Represents the shift in the energy spectrum. Represents the detector's dead time. Represents the chip counting threshold. Indicates the detector at the Counting of energy ranges, Indicates the first The starting energy of each energy range.

2. The deep learning-based matter decomposition method based on a physical model according to claim 1, characterized in that, The physical parameters of the CT detector were obtained based on the following calibration experiment: The predicted values ​​of the physical parameters of the CT detector, the thickness information of the experimental material, and the experimental response energy spectrum of the CT detector were obtained. The thickness information of the experimental material and the predicted values ​​of the physical parameters of the CT detector are input into the pre-trained neural network to obtain the predicted response energy spectrum of the CT detector. Calculate the error between the predicted response energy spectrum of the CT detector and the experimental response energy spectrum of the CT detector. If the error is greater than a preset error threshold, update the predicted value of the physical parameters of the CT detector and repeat the above steps until the error is less than the preset error threshold. When the error is less than a preset error threshold, the physical parameters of the CT detector are set to equal the predicted values ​​of the current physical parameters of the CT detector.

3. The deep learning-based matter decomposition method based on a physical model according to claim 1, characterized in that, The pre-trained neural network is obtained based on the following steps: A physical model is constructed, which is used to map the functional relationship between the initial energy spectrum, the physical parameters of the CT detector, the thickness information of the detected material, and the response energy spectrum of the CT detector. A dataset is generated based on the physical model. The dataset includes physical parameters of different CT detectors under a fixed initial energy spectrum, thickness information of different detected materials, and the response energy spectrum of the CT detectors corresponding to the two. The dataset is input into an untrained neural network to train the untrained neural network until it converges.

4. The deep learning-based matter decomposition method based on a physical model according to claim 3, characterized in that, The physical parameters of the CT detector include detector thickness, detection efficiency, electric field strength, charge sharing coefficient, chip dead time, chip counting threshold, chip threshold drift coefficient with count rate, and energy resolution.

5. The deep learning-based matter decomposition method based on a physical model according to claim 4, characterized in that, The pre-trained neural network includes a pre-trained first neural network and a pre-trained second neural network; wherein, the pre-trained first neural network is used to map the functional relationship between detector thickness, detection efficiency, electric field strength, charge sharing coefficient, thickness information of the detected material, and intermediate energy spectrum of the CT detector, and the pre-trained second neural network is used to map the functional relationship between chip dead time, chip count threshold, chip threshold drift coefficient with count rate, energy resolution, thickness information of the detected material, and final response energy spectrum of the CT detector.

6. The deep learning-based matter decomposition method based on a physical model according to claim 4, characterized in that, The physical model includes a first physical model and a second physical model; wherein, the first physical model is used to map the functional relationship between the initial energy spectrum, detector thickness, detection efficiency, electric field strength, charge sharing coefficient, thickness information of the detected material, and the intermediate energy spectrum of the CT detector, and the second physical model is used to map the functional relationship between the initial energy spectrum, chip dead time, chip count threshold, chip threshold drift coefficient with count rate, energy resolution, thickness information of the detected material, and the final response energy spectrum of the CT detector.

7. The deep learning-based matter decomposition method based on a physical model according to claim 4, characterized in that, The untrained neural network includes an untrained first neural network and an untrained second neural network; wherein, when training the untrained neural network, the input of the untrained first neural network is the ideal energy spectrum and detector thickness, detection efficiency, electric field strength, and charge sharing coefficient, and the output is the response energy spectrum of the detector under the detector thickness, detection efficiency, electric field strength, and charge sharing coefficient; the input of the untrained second neural network is the output of the untrained first neural network, chip dead time, chip count threshold, chip threshold drift coefficient with count rate, and energy resolution, and the output is the actual response energy spectrum of the chip.

8. A deep learning-based matter decomposition device based on a physical model, characterized in that, include: Parameter acquisition module: used to acquire the response energy spectrum and physical parameters of the CT detector; wherein, the physical parameters of the CT detector are obtained by reducing the error between the experimental response energy spectrum and the predicted response energy spectrum of the CT detector through a preset calibration experiment; Thickness calculation module: used to calculate the thickness information of the detected material based on the response energy spectrum and physical parameters of the CT detector, using a pre-trained neural network; wherein, the pre-trained neural network is used to fit a physical model, the physical model being used to map the functional relationship between the initial energy spectrum, the physical parameters of the CT detector, the thickness information of the detected material, and the response energy spectrum of the CT detector; wherein, the physical model includes: ； in, For the initial energy spectrum, The energy spectrum is after low-energy filtering. The energy spectrum after material decay is shown. The energy spectrum represents the ideal response of the detector. This represents the energy spectrum of the detector's true response. For the ASIC response energy spectrum, Indicates the total count. , Represents the attenuation coefficient for low-energy X-ray filtration. , This represents the thickness of the filter material. , Represents the attenuation coefficients of the two substances. , Represents the thickness of the two substances. The theoretical response matrix representing the detector. Represents the charge sharing coefficient. Represents neural networks, Represents the shift in the energy spectrum. Represents the detector's dead time. Represents the chip counting threshold. Indicates the detector at the Counting of energy ranges, Indicates the first The starting energy of each energy range.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. An electronic terminal, characterized in that, include: Processor and memory; The memory is used to store computer programs; The processor is configured to execute a computer program stored in the memory to cause the terminal to perform the method as described in any one of claims 1 to 7.

Images