Data correction model training and correction method for photon counting energy spectrum CT

Abstract

The invention provides a data correction model training and correction method for photon counting energy spectrum CT, and the method comprises the steps: training a first neural network model based on a plurality of first multi-channel distortion data and a plurality of corresponding first single-channel reference data under a single-group physical parameter condition, the first photon energy data correction model is used for correspondingly outputting first single-channel correction data according to the first multi-channel distortion data and the corresponding first single-channel reference data; training a second neural network model based on a plurality of second multi-channel distortion data, a plurality of corresponding second single-channel reference data and a plurality of groups of physical parameters under the condition of the plurality of groups of physical parameters, the second photon energy data correction model is used for correspondingly outputting second single-channel correction data according to second multi-channel distortion data, corresponding second single-channel reference data and physical parameters, and the second neural network model is formed by adding a parameter fusion network in the first neural network model in advance.

Description

Technical Field

[0001] This invention relates to the field of photon counting spectral CT imaging technology, and in particular to a method for training a photon energy data correction model, a correction method, and an image reconstruction method for photon counting spectral CT. Background Technology

[0002] Computed tomography (CT) is the foundation of radiological diagnosis and a crucial tool for major surgeries and important medical decisions. CT primarily utilizes highly penetrating X-ray beams that pass through specific cross-sections of the patient's body and are received by detectors. Based on the varying degrees of X-ray absorption by different tissues, specific algorithms are used to reconstruct the cross-sectional information of the body, resulting in tomographic images. In a CT system, detector technology is one of the most critical technologies; the detector array is responsible for converting the received X-ray signals into electrical signals. Traditional energy integrating detectors (EIDs) use scintillation crystals to absorb X-ray photons and emit visible light of corresponding intensities. This visible light is detected by photodiodes and converted into electrical signals. During this conversion process, the energy of all X-rays is accumulated, and the detector cannot collect information about the energy of individual photons. Therefore, EIDs have certain drawbacks, such as low energy resolution, significant susceptibility to electronic and scattering noise in image quality, and a tendency to produce artifacts.

[0003] With the continuous development of computed tomography (CT) technology, direct photon counting detectors (PCD) have received widespread attention and application. Unlike energy integration detectors, PCD directly converts X-ray photons into electron-hole pairs. Driven by a bias voltage, electrons move to the anode and are absorbed, allowing for direct counting of X-ray photons. The pulse height generated by charge deposition on the anode is proportional to the absorbed photon energy. By setting a pulse detection threshold, low-energy noise can be effectively filtered out, improving the signal-to-noise ratio. However, PCD also has inherent limitations. When an X-ray photon falls at the edge of a detector pixel, the resulting charge cloud may be recorded by adjacent pixels. Each pixel records only a portion of the actual photon energy, leading to errors in photon energy recording—a phenomenon known as charge sharing. This results in spectral distortion of the measured X-rays, thus affecting the final tomographic reconstruction image. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method for training a photon energy data correction model, a correction method, and an image reconstruction method for photon counting spectral CT, in order to eliminate or improve one or more defects existing in the prior art.

[0005] The first aspect of the present invention provides a method for training a photon energy data correction model for photon counting spectral CT, the method comprising the following steps:

[0006] A first neural network model is trained based on multiple first multi-channel photon energy distortion data and corresponding multiple first single-channel photon energy reference data obtained under a single set of physical parameters. This first neural network model is then trained to become a first photon energy data correction model that outputs corresponding first single-channel photon energy correction data for photon counting spectral CT based on the first multi-channel photon energy distortion data and the corresponding first single-channel photon energy reference data.

[0007] Based on multiple second multi-channel photon energy distortion data and corresponding multiple second single-channel photon energy reference data obtained under multiple sets of physical parameters, a second neural network model is trained to train the second neural network model into a second photon energy data correction model for outputting corresponding second single-channel photon energy correction data for photon counting energy spectrum CT based on the second multi-channel photon energy distortion data and corresponding second single-channel photon energy reference data and physical parameters. The second neural network model is formed by adding a parameter fusion network to the first neural network model in advance.

[0008] In some embodiments of the present invention, the multi-channel photon energy distortion data is formed by merging the photon energy distortion data of the central pixel channel generated under one or more sets of physical parameters with the photon energy distortion data of multiple channels near the central pixel channel.

[0009] In some embodiments of the present invention, the first neural network model includes a first feature extraction network, a three-dimensional filtering network, a dimensionality reduction processing network, a two-dimensional filtering network, and a second feature extraction network;

[0010] The first feature extraction network includes multiple feature extraction networks connected in sequence, used to extract energy features from the input first multi-channel photon energy distortion data and output the first multi-channel photon energy features accordingly. Each feature extraction network includes a three-dimensional convolutional layer, a normalization layer and a non-linear activation function connected in sequence.

[0011] The three-dimensional filtering network includes multiple residual blocks and a first channel injection layer connected in sequence, used to filter and denoise the extracted first multi-channel photon energy features. Each residual block includes a three-dimensional convolutional layer, a normalization layer, and a nonlinear activation function connected in sequence. The first channel injection layer includes a pooling layer connected in sequence, as well as multiple sets of three-dimensional convolutional layers and nonlinear activation functions connected in sequence. Along the input-to-output direction of the three-dimensional filtering network, the input of the first three-dimensional convolutional layer in the first residual block and the output of the last nonlinear activation function in the first channel injection layer are added together to obtain the output of the three-dimensional filtering network.

[0012] The dimensionality reduction processing network includes a three-dimensional convolutional layer, a normalization layer, a nonlinear activation function, and a dimensionality reduction layer connected in sequence. It is used to reduce the channel dimension of the first multi-channel photon energy feature after filtering and denoising, and outputs the first single-channel photon energy feature accordingly.

[0013] The two-dimensional filtering network includes multiple residual blocks and a second channel injection layer connected in sequence, used to filter and denoise the output first single-channel photon energy features. Each residual block includes a two-dimensional convolutional layer, a normalization layer, and a nonlinear activation function connected in sequence. The second channel injection layer includes a pooling layer connected in sequence, as well as multiple sets of two-dimensional convolutional layers and nonlinear activation functions connected in sequence. Along the input-to-output direction of the two-dimensional filtering network, the input of the first two-dimensional convolutional layer in the first residual block and the output of the last nonlinear activation function in the second channel injection layer are added together to obtain the output of the two-dimensional filtering network.

[0014] The second feature extraction network includes multiple two-dimensional convolutional layers connected in sequence, used to extract deeper energy features from the filtered and denoised first single-channel photon energy features, and output the first single-channel photon energy correction data accordingly.

[0015] In some embodiments of the present invention, the second neural network model includes a first feature extraction network, a first parameter fusion network, a three-dimensional filtering network, a second parameter fusion network, a dimensionality reduction processing network, a third parameter fusion network, a two-dimensional filtering network, a fourth parameter fusion network, a second feature extraction network, and a fifth parameter fusion network, wherein each parameter fusion network includes a nonlinear activation function and a fully connected layer connected in sequence.

[0016] The first feature extraction network includes multiple feature extraction networks connected in sequence, used to extract energy features from the input second multi-channel photon energy distortion data and output the second multi-channel photon energy features accordingly. Each feature extraction network includes a three-dimensional convolutional layer, a normalization layer and a non-linear activation function connected in sequence.

[0017] The first parameter fusion network is used to introduce multiple sets of physical parameters, fuse the second multi-channel photon energy features extracted by the first feature extraction network with the corresponding set of physical parameters, and output the first multi-channel fusion result accordingly.

[0018] The three-dimensional filtering network includes multiple residual blocks and a first channel injection layer connected in sequence, used to filter and denoise the output first multi-channel fusion result. Each residual block includes a three-dimensional convolutional layer, a normalization layer and a nonlinear activation function connected in sequence. The first channel injection layer includes a pooling layer connected in sequence, and multiple sets of three-dimensional convolutional layers and nonlinear activation functions connected in sequence. Along the input to output direction of the three-dimensional filtering network, the input of the first three-dimensional convolutional layer in the first residual block and the output of the last nonlinear activation function in the first channel injection layer are added together to obtain the output of the three-dimensional filtering network.

[0019] The second parameter fusion network is used to introduce multiple sets of physical parameters, and fuse the first multi-channel fusion result after filtering and denoising output by the three-dimensional filtering network with the corresponding set of physical parameters to output the second multi-channel fusion result.

[0020] The dimensionality reduction processing network includes a three-dimensional convolutional layer, a normalization layer, a nonlinear activation function, and a dimensionality reduction layer connected in sequence, which is used to reduce the channel dimension of the output second multi-channel fusion result, corresponding to the output first single-channel fusion result;

[0021] The third parameter fusion network is used to introduce multiple sets of physical parameters, and fuse the first single-channel fusion result output by the dimensionality reduction processing network with the corresponding set of physical parameters to output the second single-channel fusion result.

[0022] The two-dimensional filtering network includes multiple residual blocks and a second channel injection layer connected in sequence, used to filter and denoise the output second single-channel fusion result. Each residual block includes a two-dimensional convolutional layer, a normalization layer, and a nonlinear activation function connected in sequence. The second channel injection layer includes a pooling layer connected in sequence, as well as multiple sets of two-dimensional convolutional layers and nonlinear activation functions connected in sequence. Along the input to output direction of the two-dimensional filtering network, the input of the first two-dimensional convolutional layer in the first residual block and the output of the last nonlinear activation function in the second channel injection layer are added together to obtain the output of the two-dimensional filtering network.

[0023] The fourth parameter fusion network is used to introduce multiple sets of physical parameters, and fuse the filtered and denoised second single-channel fusion result output by the two-dimensional filtering network with the corresponding set of physical parameters to output the third single-channel fusion result.

[0024] The second feature extraction network includes multiple two-dimensional convolutional layers connected in sequence, used to extract deeper energy features from the output third single-channel fusion result;

[0025] The fifth parameter fusion network is used to introduce multiple sets of physical parameters, fuse the deeper energy features extracted by the second feature extraction network with the corresponding sets of physical parameters, and output the second single-channel photon energy correction data.

[0026] In some embodiments of the present invention, the physical parameters include some or all of the following: the thickness of the photon counting detector, pixels, bias voltage, and the number of photons emitted by the X-ray emitting end.

[0027] In some embodiments of the present invention, before training the second neural network model, the method further includes: encoding the multiple sets of physical parameters by position encoding.

[0028] A second aspect of the present invention provides a method for correcting photon energy data in photon counting spectral CT, the method comprising the following steps:

[0029] Acquire multi-channel photon energy distortion data of the target;

[0030] If the target multi-channel photon energy distortion data is obtained under a single set of physical parameters, then the target multi-channel photon energy distortion data is input into the first photon energy data correction model so that the first photon energy data correction model outputs the corresponding target single-channel photon energy correction data.

[0031] If the target multi-channel photon energy distortion data is obtained under multiple sets of physical parameters, then the target multi-channel photon energy distortion data is input into the second photon energy data correction model so that the second photon energy data correction model outputs the corresponding target single-channel photon energy correction data.

[0032] The first photon energy data correction model and the second photon energy data correction model are obtained in advance based on the photon energy data correction model training method for photon counting spectral CT as described in the first aspect above.

[0033] A third aspect of the present invention provides a method for reconstructing photon-counted CT images, the method comprising the following steps:

[0034] Image reconstruction is performed on the single-channel photon energy correction data of the target obtained by the photon energy data correction method for photon counting spectral CT as described in the second aspect above, to obtain the photon counting CT reconstructed image corresponding to the multi-channel photon energy distortion data of the target.

[0035] A fourth aspect of the present invention provides an electronic device comprising: a computer device including a processor and a memory, the memory storing computer instructions, the processor including a multi-core CPU and a multi-core GPU, for executing the computer instructions stored in the memory, wherein when the computer instructions are executed by the processor, the device implements the steps of the photon energy data correction model training method for photon counting spectral CT as described in the first aspect above, or implements the steps of the photon energy data correction method for photon counting spectral CT as described in the second aspect above, or implements the steps of the photon counting CT image reconstruction method as described in the third aspect above.

[0036] A fifth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the photon energy data correction model training method for photon counting spectral CT as described in the first aspect above, or implements the steps of the photon energy data correction method for photon counting spectral CT as described in the second aspect above, or implements the steps of the photon counting CT image reconstruction method as described in the third aspect above.

[0037] A sixth aspect of the present invention provides a computer program product comprising computer instructions which, when executed by a processor, implement the steps of the photon energy data correction model training method for photon counting spectral CT as described in the first aspect above, or implement the steps of the photon energy data correction method for photon counting spectral CT as described in the second aspect above, or implement the steps of the photon counting CT image reconstruction method as described in the third aspect above.

[0038] The present invention provides a photon energy data correction model training method, correction method, and image reconstruction method for photon counting spectral CT. Addressing the charge-sharing effect, it trains corresponding neural network models based on single-set physical parameters and multiple sets of multi-channel photon energy distortion data under different physical parameter conditions, along with their respective single-channel photon energy reference data. Furthermore, it employs a parameter fusion training method under multiple physical parameter conditions. This approach not only corrects the nonlinear distortion spectrum caused by the charge-sharing effect but also enables the neural network model to adapt to various detector physical parameter conditions, improving the model's robustness and ensuring good correction results. Moreover, it enables the image reconstruction obtained from the photon energy data after energy spectrum correction to achieve high image clarity.

[0039] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0040] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0041] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0042] Figure 1 This is a flowchart illustrating a method for training a photon energy data correction model for photon counting spectral CT according to an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram of the structure of the first neural network model corresponding to a single set of physical parameter conditions in one embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the structure of a second neural network model corresponding to multiple sets of physical parameter conditions in one embodiment of the present invention;

[0045] Figure 4 This is a flowchart illustrating a photon energy data correction method for photon counting spectral CT according to an embodiment of the present invention;

[0046] Figure 5 This is a schematic flowchart of a photon counting CT image reconstruction method according to an embodiment of the present invention;

[0047] Figure 6 This is a schematic diagram comparing the effects of reconstructed images obtained under a single-parameter mode in one embodiment of the present invention.

[0048] Figure 7 This is a schematic diagram comparing the effects of reconstructed images obtained under multiple parameter modes in one embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0050] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0051] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0052] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0053] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0054] The charge-sharing effect of photon counting detectors can distort the measured X-ray spectrum (energy spectrum), thus affecting the reconstructed image from the final computed tomography scan. To address this charge-sharing effect, this invention provides a photon energy data correction model training method, correction method, and image reconstruction method for photon-counted energy spectrum CT. By using a pre-designed neural network model to correct the photon energy data distorted by the charge-sharing effect, the nonlinear distortion of the energy spectrum caused by the charge-sharing effect can be corrected, thereby enabling the image reconstructed based on the corrected photon energy data to achieve higher clarity.

[0055] Figure 1 This is a flowchart illustrating a method for training a photon energy data correction model for photon counting spectral CT according to an embodiment of the present invention. Figure 1 As shown, the method includes the following steps:

[0056] Step S110: Based on multiple first multi-channel photon energy distortion data and corresponding multiple first single-channel photon energy reference data obtained under a single set of physical parameters, train a first neural network model to train the first neural network model into a first photon energy data correction model for outputting corresponding first single-channel photon energy correction data for photon counting energy spectrum CT according to the first multi-channel photon energy distortion data and the corresponding first single-channel photon energy reference data; and

[0057] Step S120: Based on multiple second multi-channel photon energy distortion data and corresponding multiple second single-channel photon energy reference data obtained under multiple sets of physical parameters, a second neural network model is trained to train the second neural network model into a second photon energy data correction model for outputting corresponding second single-channel photon energy correction data for photon counting energy spectrum CT according to the second multi-channel photon energy distortion data and corresponding second single-channel photon energy reference data and physical parameters. The second neural network model is formed by adding a parameter fusion network to the first neural network model in advance.

[0058] In some embodiments, the multi-channel photon energy distortion data is formed by merging the photon energy distortion data of the central pixel channel generated under one or more sets of physical parameters with the photon energy distortion data of multiple channels near the central pixel channel.

[0059] Specifically, the photon energy distortion data of the central pixel channel and its surrounding channels generated under one or more sets of physical parameters can be obtained by software simulation of the charge-sharing effect generated by the photon counting detector under one or more sets of physical parameters. Here, a channel can refer to a row or column of the pixel array in the detector; the central pixel channel is the row or column containing the central pixel unit in the pixel array, and each central pixel channel contains multiple central pixel units. When the charge-sharing effect occurs, because photons fall on the edge of the central pixel unit of the detector and are recorded by multiple surrounding pixel units, the actual number of photons received by the central pixel unit decreases, and the photon energy decreases. Therefore, the number of photons in the high-energy region decreases, and the number of photons in the low-energy region increases, leading to distortion of the incident X-ray energy spectrum. Therefore, in the ideal state where the detector does not produce the charge-sharing effect, the true photon energy received by the central pixel unit should be the sum of the photon energy received by the central pixel unit and its surrounding multiple pixel units under the state where the detector produces the charge-sharing effect. In this embodiment, based on the principle that the charge-sharing effect in photon-counting CT causes spectral distortion, under the physical parameters of one or more sets of detectors, software simulation is used to obtain distorted photon energy data, i.e., photon energy distortion data, received or recorded by pixel units in each row and column of the pixel array at multiple different angles. In the specific implementation, an open-source photon-counting detector toolkit is used to simulate a real photon-counting CT scanner. By introducing the charge-sharing effect generated by the interaction between incident photons and detector materials, and simultaneously introducing a certain amount of noise, photon energy distortion data is obtained. Based on the energy of the incident X-rays, this distorted photon energy data is divided into multiple energy windows, corresponding to multiple energy thresholds of the detector, such as 20, 70, and 120. The photon energy distortion data received or recorded by all pixel units in each row or column are grouped together, and multiple groups of photon energy distortion data form a photon energy distortion dataset. This dataset contains multiple single-channel photon energy distortion data, each representing the photon energy distortion data of different energy windows at different angles where charge sharing effects occur, recorded by all pixel units in each channel. Therefore, each single-channel photon energy distortion data in the dataset contains four dimensions, and its size can be represented as (R, W, A, C) = (1, W, A, C), where R represents the number of rows or columns (channels) of the photon counting detector, W represents the number of energy windows into which the photons received by each pixel unit are divided according to their energy magnitude (i.e., the number of energy thresholds), A represents the number of angles within the detector's detection range, and C represents the number of pixel units in each channel.

[0060] In a specific example, the charge sharing effect considers the case where, when the central pixel receives an X-ray photon, the photon falls on the edge of the central pixel and is received by a 3×3 pixel array centered on the central pixel (that is, the central pixel is located in the second row and second column of the 3×3 pixel array). In other words, it considers the charge sharing of the 3×3 pixel array containing the central pixel and its surrounding pixels, where the number of photons and photon energy of the central pixel, which is a high-energy region, decreases, while the number of photons and photon energy of the surrounding pixels of the central pixel, which is a low-energy region, increase.

[0061] Therefore, the number of channels in multi-channel photon energy distortion data formed by merging and expanding single-channel photon energy distortion data is related to the number of pixel units that receive or record distorted photons when the incident photons generate a charge-sharing effect. In a specific example, for the charge-sharing case of a 3×3 pixel array centered on the center pixel, the photon energy distortion data of the channel containing the center pixel that generates the charge-sharing effect is expanded into three-channel photon energy distortion data. After the photons received by the center pixel generate a charge-sharing effect, specifically, all the photon energy distortion data recorded in the channel containing the center pixel and the channels of the eight pixel units surrounding the center pixel in the 3×3 pixel array that record the photon energy are merged to form three-channel photon energy distortion data. The size of the three-channel photon energy distortion data is (R,W,A,C)=(3,W,A,C).

[0062] The single-channel photon energy reference data in step S110 or S120 is a collection of photon energy reference data received or recorded by all pixel units in each row or column at various different angles, acquired under an ideal state where the detector does not produce charge sharing effects, under one or more sets of physical parameters of the photon counting detector. In acquiring this photon energy reference data, an open-source photon counting detector toolkit is used to simulate a real photon counting CT scanner. One or more sets of detector physical parameters are set, and no charge sharing effects or noise are introduced. Under this ideal state, photon energy reference data received or recorded by pixel units in each row and column at various different angles is acquired. In a specific example, under the ideal state, the photon is received and recorded by the central pixel in a 3×3 pixel array. It is assumed that, apart from the central pixel, multiple other pixels around the central pixel do not receive the incident photon; that is, the photon is not recorded by any other pixel, and the photon energy spectrum is not distorted.

[0063] In some embodiments, the physical parameters include some or all of the following: the thickness of the photon counting detector, pixels, bias voltage, and the number of photons emitted by the X-ray emitting end.

[0064] Specifically, the number of photons emitted by the aforementioned transmitter refers to the number of photons emitted by the transmitter per unit time. A single set of physical parameters can be a condition formed by taking a set of corresponding values ​​for the three parameters: the thickness of the photon counting detector, the number of pixels, and the number of photons emitted by the transmitter. For example, a thickness of 'a', a pixel value of 'b', and a photon emission value of 'e' can be represented by the vector (a, b, e). Alternatively, it can be a condition formed by taking a set of corresponding values ​​for the four parameters: the thickness of the photon counting detector, the number of pixels, the bias voltage, and the photon emission value. For example, a thickness of 'a', a pixel value of 'b', a photon emission value of 'e', ​​and a bias voltage value of 'd' can be represented by the vector (a, b, e, d). There are multiple possibilities. Multiple sets of physical parameter conditions can be conditions formed by taking multiple corresponding values ​​of the three parameters of the photon counting detector: thickness, pixel count, and number of photons emitted by the transmitter. For example, the thickness value is a1, the pixel count is b1, and the number of photons emitted by the transmitter is e1; and the thickness value is a2, the pixel count is b2, and the number of photons emitted by the transmitter is e2, etc., i.e., vectors (a1,b1,e1) and (a2,b2,e2), etc.; or conditions formed by taking multiple corresponding values ​​of the four parameters of the photon counting detector: thickness, pixel count, bias voltage, and number of photons emitted by the transmitter. For example, the thickness value is a1, the pixel count is b1, the number of photons emitted by the transmitter is e1, and the bias voltage is d1; and the thickness value is a2, the pixel count is b2, the number of photons emitted by the transmitter is e2, and the bias voltage is d2, etc., i.e., vectors (a1,b1,e1,d1) and (a2,b2,e2,d2), etc.; and there are also many different cases.

[0065] In some embodiments, the first neural network model includes: a first feature extraction network, a three-dimensional filtering network, a dimensionality reduction processing network, a two-dimensional filtering network, and a second feature extraction network connected in sequence;

[0066] The first feature extraction network includes multiple feature extraction networks connected in sequence, used to extract energy features from the input first multi-channel photon energy distortion data and output the first multi-channel photon energy features accordingly. Each feature extraction network includes a three-dimensional convolutional layer, a normalization layer and a non-linear activation function connected in sequence.

[0067] The three-dimensional filtering network includes multiple residual blocks and a first channel injection layer connected in sequence, used to filter and denoise the extracted first multi-channel photon energy features. Each residual block includes a three-dimensional convolutional layer, a normalization layer, and a nonlinear activation function connected in sequence. The first channel injection layer includes a pooling layer connected in sequence, as well as multiple sets of three-dimensional convolutional layers and nonlinear activation functions connected in sequence. Along the input-to-output direction of the three-dimensional filtering network, the input of the first three-dimensional convolutional layer in the first residual block and the output of the last nonlinear activation function in the first channel injection layer are added together to obtain the output of the three-dimensional filtering network.

[0068] The dimensionality reduction processing network includes a three-dimensional convolutional layer, a normalization layer, a nonlinear activation function, and a dimensionality reduction layer connected in sequence. It is used to reduce the channel dimension of the first multi-channel photon energy feature after filtering and denoising, and outputs the first single-channel photon energy feature accordingly.

[0069] The two-dimensional filtering network includes multiple residual blocks and a second channel injection layer connected in sequence, used to filter and denoise the output first single-channel photon energy features. Each residual block includes a two-dimensional convolutional layer, a normalization layer, and a nonlinear activation function connected in sequence. The second channel injection layer includes a pooling layer connected in sequence, as well as multiple sets of two-dimensional convolutional layers and nonlinear activation functions connected in sequence. Along the input-to-output direction of the two-dimensional filtering network, the input of the first two-dimensional convolutional layer in the first residual block and the output of the last nonlinear activation function in the second channel injection layer are added together to obtain the output of the two-dimensional filtering network.

[0070] The second feature extraction network includes multiple two-dimensional convolutional layers connected in sequence, used to extract deeper energy features from the filtered and denoised first single-channel photon energy features, and output the first single-channel photon energy correction data accordingly.

[0071] like Figure 2As shown, in a specific example, the first feature extraction network includes two feature extraction networks. Three-channel photon energy distortion data (3,3,360,1632) obtained under a set of physical parameters is input into the first feature extraction network. The energy features in the three-dimensional convolutional layer of the first feature extraction network are extracted for the first time, and then passed through a normalization layer and a nonlinear activation function for group normalization and nonlinearity introduction. The second feature extraction network performs a second feature extraction using a three-dimensional convolutional layer, and then passes through a normalization layer and a nonlinear activation function to obtain the output three-channel energy feature data (64,3,360,1632). A 3D filtering network can alleviate the vanishing gradient problem in deep networks. It consists of two residual blocks. The three-channel energy feature data (64,3,360,1632) is input into the 3D filtering network for filtering and noise removal. The pooling layer adaptively compresses the features. The input three-channel energy feature data (64,3,360,1632) is added to the outputs of the two residual blocks and the channel injection layer, resulting in the final filtered three-channel energy feature data (64,3,360,1632). Then, the filtered three-channel energy feature data (64,3,360,1632) is input into a dimensionality reduction network for channel dimensionality reduction, converting it into single-channel energy feature data (64,1,360,1632) and removing the dimension with a dimension of 1, outputting the single-channel energy feature data (64,360,1632). A two-dimensional filtering network can alleviate the gradient vanishing problem in deep networks. It consists of two residual blocks. The single-channel energy feature data (64,360,1632) is input into the two-dimensional filtering network for filtering and noise removal. The pooling layer adaptively compresses the features. The input single-channel energy feature data (64,360,1632) is added to the outputs of the two residual blocks and the channel injection layer, resulting in the final filtered single-channel energy feature data (64,360,1632). Finally, the filtered single-channel energy feature data (64,360,1632) is input into a second feature extraction network for deeper energy feature extraction. The second feature extraction network outputs the final single-channel energy feature data (3,360,1632), which is the corrected single-channel photon energy data (1,3,360,1632).

[0072] In some embodiments, the photon energy data correction model training method further includes encoding the multiple sets of physical parameters by position encoding before training the second neural network model.

[0073] Specifically, the position of multiple sets of physical parameters of the detector is encoded using the following formulas:

[0074]

[0075]

[0076] posA = sin(X) extend ⊙InvF)

[0077] posB=cos(X extend ⊙InvF)

[0078] pos = [posA, posB]

[0079] Where, n represents the number of physical parameters in a set of photon counting detector physical parameters; c represents the dimension size of the encoded output data, which is consistent with the size of the first dimension in the output data of each network (first feature extraction network, three-dimensional filtering network, dimension reduction processing network, two-dimensional filtering network, or second feature extraction network) to be fused in the neural network model with parameter fusion (i.e., related to the size of the convolution kernel of each convolutional layer in each network); A is a constructed vector, the value of the elements in the vector is 2k, where k∈[0,2,4,…,c×n / 2); InvF is the inverse frequency matrix in position encoding; X is a physical parameter vector composed of the physical parameters of a set of photon counting detectors, the length of which is n; the physical parameter vector X is expanded to obtain Xn. extend X extend The length of posA is the same as the length of the inverse frequency matrix InvF; posA is the sine coding matrix, posB is the cosine coding matrix, ⊙ represents the operation of multiplying corresponding elements of the matrix, and pos is the physical parameter coding matrix obtained after coding.

[0080] In some embodiments, the second neural network model includes: a first feature extraction network, a first parameter fusion network, a three-dimensional filtering network, a second parameter fusion network, a dimensionality reduction processing network, a third parameter fusion network, a two-dimensional filtering network, a fourth parameter fusion network, a second feature extraction network, and a fifth parameter fusion network. Each parameter fusion network includes a non-linear activation function and a fully connected layer connected sequentially.

[0081] The first feature extraction network includes multiple feature extraction networks connected in sequence, used to extract energy features from the input second multi-channel photon energy distortion data and output the second multi-channel photon energy features accordingly. Each feature extraction network includes a three-dimensional convolutional layer, a normalization layer and a non-linear activation function connected in sequence.

[0082] The first parameter fusion network is used to introduce multiple sets of physical parameters, fuse the second multi-channel photon energy features extracted by the first feature extraction network with the corresponding set of physical parameters, and output the first multi-channel fusion result accordingly.

[0083] The three-dimensional filtering network includes multiple residual blocks and a first channel injection layer connected in sequence, used to filter and denoise the output first multi-channel fusion result. Each residual block includes a three-dimensional convolutional layer, a normalization layer and a nonlinear activation function connected in sequence. The first channel injection layer includes a pooling layer connected in sequence, and multiple sets of three-dimensional convolutional layers and nonlinear activation functions connected in sequence. Along the input to output direction of the three-dimensional filtering network, the input of the first three-dimensional convolutional layer in the first residual block and the output of the last nonlinear activation function in the first channel injection layer are added together to obtain the output of the three-dimensional filtering network.

[0084] The second parameter fusion network is used to introduce multiple sets of physical parameters, and fuse the first multi-channel fusion result after filtering and denoising output by the three-dimensional filtering network with the corresponding set of physical parameters to output the second multi-channel fusion result.

[0085] The dimensionality reduction processing network includes a three-dimensional convolutional layer, a normalization layer, a nonlinear activation function, and a dimensionality reduction layer connected in sequence, which is used to reduce the channel dimension of the output second multi-channel fusion result, corresponding to the output first single-channel fusion result;

[0086] The third parameter fusion network is used to introduce multiple sets of physical parameters, and fuse the first single-channel fusion result output by the dimensionality reduction processing network with the corresponding set of physical parameters to output the second single-channel fusion result.

[0087] The two-dimensional filtering network includes multiple residual blocks and a second channel injection layer connected in sequence, used to filter and denoise the output second single-channel fusion result. Each residual block includes a two-dimensional convolutional layer, a normalization layer, and a nonlinear activation function connected in sequence. The second channel injection layer includes a pooling layer connected in sequence, as well as multiple sets of two-dimensional convolutional layers and nonlinear activation functions connected in sequence. Along the input to output direction of the two-dimensional filtering network, the input of the first two-dimensional convolutional layer in the first residual block and the output of the last nonlinear activation function in the second channel injection layer are added together to obtain the output of the two-dimensional filtering network.

[0088] The fourth parameter fusion network is used to introduce multiple sets of physical parameters, and fuse the filtered and denoised second single-channel fusion result output by the two-dimensional filtering network with the corresponding set of physical parameters to output the third single-channel fusion result.

[0089] The second feature extraction network includes multiple two-dimensional convolutional layers connected in sequence, used to extract deeper energy features from the output third single-channel fusion result;

[0090] The fifth parameter fusion network is used to introduce multiple sets of physical parameters, fuse the deeper energy features extracted by the second feature extraction network with the corresponding sets of physical parameters, and output the second single-channel photon energy correction data.

[0091] like Figure 3 As shown, in a specific example, in Figure 2 Based on the network structure shown, a fusion network of the first to fifth parameters is added, namely... Figure 2 The outputs of each network in the first step are added to the outputs of the respective parameter fusion networks and then fed into the next network. Each parameter fusion network consists of a non-linear activation function and a fully connected layer connected sequentially. The non-linear activation function uses the SiLU function, and the data dimension of the output of the fully connected layer is consistent with the data dimension of the output of the corresponding network being added. That is, in Figure 3In the neural network model shown, based on the three-channel photon energy distortion data (3,3,360,1632) obtained under multiple sets of physical parameters, pre-encoded sets of physical parameters under corresponding conditions are also input. The three-channel photon energy distortion data (3,3,360,1632) is input into the first feature extraction network for two feature extractions, which outputs three-channel energy feature data (64,3,360,1632). The physical parameter encoding matrix (1,3×64) obtained by pre-encoding a set of physical parameters under corresponding conditions is input into the first and second parameter fusion networks respectively, both outputting a physical parameter encoding matrix (1,64), and (1,64) is transformed into (64,3,360,1632). Alternatively, it can be transformed by the fully connected components in the first and second parameter fusion networks. The data dimension is transformed layer by layer, transforming (1,64) into (64,3,360,1632). After passing through the physical parameter encoding matrix (64,3,360,1632) of the first parameter fusion network and the corresponding three-channel energy feature data (64,3,360,1632), the first three-channel fusion result is obtained and input into the three-dimensional filtering network. The three-dimensional filtering network outputs the filtered first three-channel fusion result (64,3,360,1632). This process is repeated, passing through the physical parameter encoding matrix (64,3,360,1632) of the second parameter fusion network and the corresponding three-channel energy feature data (64,3,360,1632). The filtered first three-channel fusion result (64, 3, 360, 1632) is summed to obtain the second three-channel fusion result, which is then input into the dimensionality reduction processing network for dimensionality reduction. The dimensionality reduction processing network outputs the first single-channel fusion result (64, 1, 360, 1632), omitting dimension 1. The physical parameter encoding matrix (1, 3×64) obtained by precoding the set of physical parameters under the above conditions is then input into the third and fourth parameter fusion networks, respectively, and both output the physical parameter encoding matrix (1, 64). ), and (1,64) is transformed into (64,360,1632), or the data dimension can be transformed by the fully connected layer in the third and fourth parameter fusion network, transforming (1,64) into (64,360,1632). The physical parameter encoding matrix (64,360,1632) after passing through the third parameter fusion network and the corresponding first single-channel fusion result (64,360,1632) are added to obtain the second single-channel fusion result and input into the two-dimensional filtering network. The two-dimensional filtering network outputs the filtered second single-channel fusion result (64,360,1632).The physical parameter encoding matrix (64,360,1632) obtained through the fourth parameter fusion network is added to the corresponding filtered second single-channel fusion result (64,360,1632) to obtain the third single-channel fusion result, which is then input into the second feature extraction network. The second feature extraction network outputs the third single-channel fusion result (3,360,1632) after extracting energy features. The physical parameter encoding matrix (1,3×64) obtained by pre-encoding this set of physical parameters under the above conditions is then input into the fifth parameter fusion network to output the physical parameter encoding matrix (64,360,1632). The physical parameter encoding matrix (1,3) is transformed into (3,360,1632). Alternatively, the data dimension can be transformed by the fully connected layer in the fifth parameter fusion network, changing (1,3) into (3,360,1632). The result of adding the physical parameter encoding matrix (3,360,1632) from the fifth parameter fusion network to the corresponding third single-channel fusion result (3,360,1632) after extracting energy features is the corrected single-channel photon energy data (1,3,360,1632).

[0092] In steps S110 and S120, the photon energy dataset, consisting of multiple multi-channel photon energy distortion data and corresponding multiple first single-channel photon energy reference data, is divided into a training dataset and a test dataset according to a certain ratio. Based on the training dataset and the test dataset, the neural network model is iteratively trained by minimizing the loss function until the loss function converges, enabling the model to correct photon energy distortion data and energy spectrum. During the iterative training process, the model performance under different initial learning rates is compared to select a suitable initial learning rate, and the learning rate is dynamically adjusted on demand.

[0093] During the iterative training of the parameter-fused second neural network model, multi-channel photon energy distortion data, corresponding single-channel photon energy reference data, and pre-encoded physical parameters of the corresponding conditions (sets) are input into the second neural network model with parameter fusion. A loss function is calculated and minimized using backpropagation. This loss function specifically uses the MSE function, as shown in the following formula:

[0094]

[0095] Among them, L MSELet X be the loss value, S be the physical parameters of the photon counting detector, f be the corrected single-channel photon energy data output by the model (single-channel photon energy correction data), y be the single-channel photon energy reference data corresponding to S, and i be the index of the input multi-channel photon energy distortion data, the corresponding single-channel photon energy reference data, or a set of physical parameters, i = 1, 2, ..., n, where n is the total number of parameters involved in the mean square error calculation. Then, the gradient values ​​of the current model parameters are calculated through backpropagation. In each iteration of training, the gradient descent method is used, combined with the model's learning rate, and the Adam optimization algorithm is used to adjust the model's parameters and gradually reduce the gradient value until the model's loss function converges, so that the model's loss function value reaches its minimum value. This allows the model's output to infinitely approach the corresponding single-channel photon energy reference data. Then, the training of the neural network model is stopped, and the trained parameter fusion neural network model for energy spectrum correction of photon counting CT imaging is finally obtained.

[0096] Figure 4 This is a flowchart illustrating a method for correcting photon energy data in photon counting spectral CT according to an embodiment of the present invention. Figure 4 As shown, this embodiment of the invention also provides a method for correcting photon energy data in photon counting spectral CT, the method comprising the following steps:

[0097] Step S410: Obtain target multi-channel photon energy distortion data.

[0098] The target multi-channel photon energy distortion data in this step is obtained by acquiring the real photon energy distortion data generated by the charge sharing effect of the photon counting detector, and then merging the photon energy distortion data of each central pixel channel with the photon energy distortion data of multiple channels near the corresponding central pixel channel.

[0099] Step S420: If the target multi-channel photon energy distortion data is obtained under a single set of physical parameters, then the target multi-channel photon energy distortion data is input into the first photon energy data correction model so that the first photon energy data correction model outputs the corresponding target single-channel photon energy correction data for photon counting spectral CT.

[0100] Step S430: If the target multi-channel photon energy distortion data is obtained under multiple sets of physical parameters, then the target multi-channel photon energy distortion data is input into the second photon energy data correction model so that the second photon energy data correction model outputs the corresponding target single-channel photon energy correction data for photon counting spectral CT.

[0101] The first photon energy data correction model and the second photon energy data correction model are obtained by training in advance based on the photon energy data correction model training method for photon counting energy spectrum CT in any of the foregoing embodiments.

[0102] The single-channel photon energy correction data output by the trained neural network model, i.e., the aforementioned photon energy data correction model, is the corrected photon energy data of the central pixel channel contained in the input multi-channel photon energy distortion data after correction. Correspondingly, the energy spectrum of the distorted X-ray photons is also corrected, gathering the incorrectly counted lower-energy photons into higher-energy regions. In a specific example, the corrected single-channel photon energy data is the photon energy correction data of the central pixel channel in a 3×3 matrix, i.e., the corrected photon energy data of all pixels in the second row or column of the 3×3 matrix. Accordingly, by gathering the photon counts of the eight surrounding pixels of the incorrectly counted central pixel in the 3×3 matrix into that central pixel, the distorted energy spectrum can be effectively corrected.

[0103] Figure 5 This is a schematic flowchart of a photon counting CT image reconstruction method according to an embodiment of the present invention. Figure 5 As shown, this embodiment of the invention also provides a photon counting CT image reconstruction method, which includes the following steps:

[0104] Step S510: Perform image reconstruction on the target single-channel photon energy correction data obtained by the photon energy data correction method for photon counting spectral CT in any of the foregoing embodiments to obtain the photon counting CT reconstructed image corresponding to the target multi-channel photon energy distortion data.

[0105] In this embodiment, a noise-removed reconstructed image for photon counting CT is directly obtained from multiple corrected single-channel photon energy data obtained by the photon energy data correction method in any of the foregoing embodiments using a filtered back-projection algorithm. The pixel size of the detector, the number of rows and columns of the detector array, and other parameters used in this algorithm are consistent with the physical parameters of the detector set when acquiring the photon energy distortion data corresponding to the reconstructed image.

[0106] The performance of the neural network model trained using the aforementioned image reconstruction method, namely the photon energy data correction model, was tested and evaluated. Specifically, the first and second photon energy data correction models were tested using test datasets obtained under single and multiple sets of physical parameters. First, the first and second photon energy data correction models corrected the multi-channel photon energy distortion data and energy spectrum in the test datasets under single and multiple sets of physical parameters, respectively. Then, image reconstruction was performed on the single-channel photon energy correction data output by the first and second photon energy data correction models, respectively, to obtain the first and second photon count CT reconstructed images. Next, the image reconstruction method was used to directly reconstruct the ideal photon energy reference data in the test datasets under single and multiple sets of physical parameters, respectively, to obtain the first and second reconstructed reference images. Finally, the performance of the first and second photon energy data correction models was evaluated by comparing the first and second photon count CT reconstructed images with the first and second reconstructed reference images, respectively. Specifically, several evaluation metrics are used, including the peak signal-to-noise ratio (PSNR), normalized mean square error (NMSE), and structural similarity index (SSIM), as shown in the following formulas:

[0107]

[0108]

[0109]

[0110]

[0111] Among them, F MSE Let I(k,j) and K(k,j) represent a monochrome image of size m×p, where I(k,j) is the photon-counted CT reconstructed image, and K(k,j) is the reconstructed reference image. k and j represent the k-th row and j-th column in the image data matrix, respectively, k = 0, 2, ..., m-1, j = 0, 2, ..., p-1; F NMSE Normalized mean square error (MSE) is a normalized form of MSE; F PSNR Peak signal-to-noise ratio (PSNR), MAX represents the maximum value of each pixel in a monochrome image; F SSIM μ is the structural similarity index. I and μ K δ represents the average values ​​of images I and K, respectively. I 2 and δ K 2The variances of images I and K are δ, respectively. Ik Let F be the covariance of images I and K, where C1 and C2 are constants. NMSE The value ranges between 0 and 1, F NMSE The smaller the value of F, the better the performance of the photon energy data correction model; PSNR The larger the value, the closer the reconstructed image from photon-count CT is to the reconstructed reference image; that is, the higher the quality of the reconstructed image from photon-count CT. SSIM The value ranges from 0 to 1 and is used to measure the similarity between a photon-counted CT reconstructed image and a reconstructed reference image. SSIM The larger the value, the higher the similarity and the smaller the difference between the photon count CT reconstructed image and the reconstructed reference image.

[0112] In a specific test case, the test results are shown in Tables 1-3. In Tables 1-3, the evaluation index data corresponding to noise are based on the distorted images obtained after image reconstruction from photon energy distortion data obtained under noisy conditions and with charge-sharing effects (e.g., ...). Figure 6 or Figure 7 The distorted image in the image is obtained; the parameter fusion mode or multiple parameter modes correspond to the second photon energy data correction model; the single parameter mode corresponds to the first photon energy data correction model.

[0113] Table 1

[0114]

[0115] Table 2

[0116]

[0117] Table 3

[0118]

[0119] As shown in Tables 1-3, under the same input conditions, the correction performance of the second photon energy data correction model in the parameter fusion mode is basically consistent with that of the first photon energy data correction model in the single-parameter mode, and all evaluation indicators are significantly improved. After energy spectrum correction, the average value of each indicator for multiple energy windows is taken. It can be seen that in both modes, the PSNR value can reach above 30, the NMSE value is close to 0, and the SSIM value is close to 1. This also indicates that in both modes, the photon-counted CT reconstructed image obtained through energy spectrum correction is very close to the reconstructed reference image obtained under ideal conditions, and the image similarity is very high.

[0120] Correspondingly, the reconstructed images obtained in the two modes are as follows: Figures 6-7As shown, the distorted image is the one directly reconstructed from the photon energy distortion data, the corrected image is the photon-counted CT reconstructed image obtained by reconstructing the corrected single-channel photon energy data (single-channel photon energy corrected data), and the reference image is the reconstructed reference image obtained by reconstructing the photon energy reference data under ideal conditions. By selecting a region of interest (ROI) and measuring the average pixel value, it can be seen that the ROI map (region of interest map) in the photon-counted CT reconstructed image obtained after energy spectrum correction is quite close to the mean of the reconstructed reference image. Therefore, it can be concluded that both modes can effectively improve the quality of the reconstructed image. It should be noted that the second photon energy data correction model in parameter fusion mode or multi-parameter mode showed superior correction performance for photon energy distortion data obtained under multiple conditions with different sets of physical parameters, and the quality of the corresponding reconstructed images was effectively improved. In contrast, the first photon energy data correction model in single-parameter mode showed superior correction performance, which was basically consistent with the above performance, only for photon energy distortion data obtained under a single condition (single set of physical parameters). The correction effect of photon energy distortion data obtained under other conditions decreased. Therefore, the second photon energy data correction model in parameter fusion mode or multi-parameter mode has better and wider applicability.

[0121] Corresponding to the above method, the present invention also provides an electronic device, which includes a computer device, the computer device including a processor and a memory, the memory storing computer instructions, and the processor including a multi-core CPU and a multi-core GPU, for executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the device implements the steps of the aforementioned method.

[0122] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned method. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0123] This invention also provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the aforementioned method.

[0124] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0125] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0126] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0127] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A photon energy data correction model training method for photon counting spectral CT, characterized in that: The method comprises: Training a first neural network model based on a plurality of first multi-channel photon energy distortion data and a corresponding plurality of first single-channel photon energy reference data obtained under a single set of physical parameter conditions, so as to train the first neural network model as a first photon energy data correction model for outputting corresponding first single-channel photon energy correction data for photon counting spectral CT according to the first multi-channel photon energy distortion data and the corresponding first single-channel photon energy reference data; and A second neural network model is trained based on multiple second multi-channel photon energy distortion data and corresponding multiple second single-channel photon energy reference data obtained under multiple groups of physical parameter conditions, and multiple groups of physical parameters, so that the second neural network model is trained as a second photon energy data correction model for outputting corresponding second single-channel photon energy correction data for photon counting spectral CT according to the second multi-channel photon energy distortion data and corresponding second single-channel photon energy reference data and physical parameters, wherein the second neural network model is formed by adding a parameter fusion network to the first neural network model in advance.

2. The method according to claim 1, characterized in that The multi-channel photon energy distortion data is formed by combining the photon energy distortion data of a central pixel channel generated under a single set or multiple sets of physical parameter conditions and the photon energy distortion data of multiple channels near the central pixel channel.

3. The method according to claim 1, characterized in that: The first neural network model includes a first feature extraction network, a three-dimensional filtering network, a dimensionality reduction processing network, a two-dimensional filtering network, and a second feature extraction network; The first feature extraction network includes a plurality of feature extraction networks connected in sequence, and is used to extract energy features from the input first multi-channel photon energy distortion data, and correspondingly output first multi-channel photon energy features, wherein each feature extraction network includes a three-dimensional convolution layer, a normalization layer, and a nonlinear activation function connected in sequence; The three-dimensional filtering network includes a plurality of residual blocks and a first channel injection layer connected in sequence, which are used to filter and denoise the extracted first multi-channel photon energy features, wherein each residual block includes a three-dimensional convolution layer, a normalization layer and a nonlinear activation function connected in sequence, the first channel injection layer includes a pooling layer connected in sequence, and a plurality of groups of three-dimensional convolution layers and nonlinear activation functions connected in sequence, and along the direction from input to output of the three-dimensional filtering network, the input of the first three-dimensional convolution layer in the first residual block and the output of the last nonlinear activation function in the first channel injection layer are added as the output of the three-dimensional filtering network; The dimension reduction processing network includes a three-dimensional convolution layer, a normalization layer, a nonlinear activation function and a dimension reduction layer connected in sequence, which is used to reduce the channel dimension of the first multi-channel photon energy feature after filtering and denoising, and correspondingly output the first single-channel photon energy feature; The two-dimensional filtering network includes a plurality of residual blocks and a second channel injection layer connected in sequence, and is used to filter and denoise the output first single-channel photon energy feature, wherein each residual block includes a two-dimensional convolution layer, a normalization layer and a nonlinear activation function connected in sequence, and the second channel injection layer includes a pooling layer connected in sequence, and a plurality of groups of two-dimensional convolution layers and nonlinear activation functions connected in sequence, and along the direction from input to output of the two-dimensional filtering network, the input of the first two-dimensional convolution layer in the first residual block and the output of the last nonlinear activation function in the second channel injection layer are added as the output of the two-dimensional filtering network; The second feature extraction network includes a plurality of two-dimensional convolutional layers connected in sequence, which are used to extract deeper energy features from the filtered and denoised first single-channel photon energy features, and correspondingly output first single-channel photon energy correction data.

4. The method according to claim 1, characterized in that The second neural network model includes a first feature extraction network, a first parameter fusion network, a three-dimensional filtering network, a second parameter fusion network, a dimensionality reduction processing network, a third parameter fusion network, a two-dimensional filtering network, a fourth parameter fusion network, a second feature extraction network and a fifth parameter fusion network, wherein each parameter fusion network includes a nonlinear activation function and a fully connected layer connected in sequence; The first feature extraction network includes a plurality of feature extraction networks connected in sequence, and is used to extract energy features from the input second multi-channel photon energy distortion data, and correspondingly output second multi-channel photon energy features, wherein each feature extraction network includes a three-dimensional convolution layer, a normalization layer, and a nonlinear activation function connected in sequence; The first parameter fusion network is used to introduce multiple groups of physical parameters, fuse the second multi-channel photon energy features extracted by the first feature extraction network with the physical parameters of the corresponding groups, and output a first multi-channel fusion result accordingly; The three-dimensional filtering network includes a plurality of residual blocks and a first channel injection layer connected in sequence, which are used to filter and denoise the output first multi-channel fusion result, wherein each residual block includes a three-dimensional convolution layer, a normalization layer and a nonlinear activation function connected in sequence, and the first channel injection layer includes a pooling layer connected in sequence, and a plurality of groups of three-dimensional convolution layers and nonlinear activation functions connected in sequence, and along the direction from input to output of the three-dimensional filtering network, the input of the first three-dimensional convolution layer in the first residual block and the output of the last nonlinear activation function in the first channel injection layer are added as the output of the three-dimensional filtering network; The second parameter fusion network is used to introduce multiple groups of physical parameters, fuse the first multi-channel fusion result after filtering and denoising output by the three-dimensional filtering network with the physical parameters of the corresponding groups, and output a second multi-channel fusion result accordingly; The dimension reduction processing network includes a three-dimensional convolution layer, a normalization layer, a nonlinear activation function and a dimension reduction layer connected in sequence, which is used to reduce the channel dimension of the output second multi-channel fusion result and correspondingly output the first single-channel fusion result; The third parameter fusion network is used to introduce multiple groups of physical parameters, fuse the first single-channel fusion result output by the dimensionality reduction processing network with the physical parameters of the corresponding group, and output a second single-channel fusion result accordingly; The two-dimensional filtering network includes a plurality of residual blocks and a second channel injection layer connected in sequence, which are used to filter and denoise the output second single-channel fusion result, wherein each residual block includes a two-dimensional convolution layer, a normalization layer and a nonlinear activation function connected in sequence, and the second channel injection layer includes a pooling layer connected in sequence, and a plurality of groups of two-dimensional convolution layers and nonlinear activation functions connected in sequence, and along the direction from input to output of the two-dimensional filtering network, the input of the first two-dimensional convolution layer in the first residual block and the output of the last nonlinear activation function in the second channel injection layer are added as the output of the two-dimensional filtering network; The fourth parameter fusion network is used to introduce multiple groups of physical parameters, fuse the second single-channel fusion result after filtering and denoising output by the two-dimensional filtering network with the physical parameters of the corresponding group, and output a third single-channel fusion result accordingly; The second feature extraction network includes a plurality of two-dimensional convolutional layers connected in sequence, and is used to extract deeper energy features in the output third single-channel fusion result; The fifth parameter fusion network is used to introduce multiple groups of physical parameters, fuse the deeper energy features extracted by the second feature extraction network with the physical parameters of the corresponding groups, and output the second single-channel photon energy correction data accordingly.

5. The method according to any one of claims 1 to 4, characterized in that The physical parameters include part or all of the following: thickness, pixels, bias voltage and number of photons emitted by the X-ray emitting end of the photon counting detector; Before training the second neural network model, the method further includes: encoding the multiple groups of physical parameters by position encoding.

6. A method for correcting photon energy data for photon counting spectral CT, characterized in that: The method comprises: Acquire target multi-channel photon energy distortion data; If the target multi-channel photon energy distortion data is obtained under a single set of physical parameter conditions, the target multi-channel photon energy distortion data is input into a first photon energy data correction model, so that the first photon energy data correction model outputs corresponding target single-channel photon energy correction data for photon counting spectral CT; If the target multi-channel photon energy distortion data is obtained under multiple sets of physical parameter conditions, the target multi-channel photon energy distortion data is input into a second photon energy data correction model, so that the second photon energy data correction model outputs corresponding target single-channel photon energy correction data for photon counting spectral CT; The first photon energy data correction model and the second photon energy data correction model are pre-trained based on the photon energy data correction model training method for photon counting spectral CT according to any one of claims 1 to 5.

7. A photon counting CT image reconstruction method, characterized in that: The method comprises: The target single-channel photon energy correction data obtained by the photon energy data correction method for photon counting spectral CT as claimed in claim 6 is subjected to image reconstruction to obtain a photon counting CT reconstructed image corresponding to the target multi-channel photon energy distortion data.

8. An electronic device comprising a processor, a memory and computer instructions stored in the memory, characterized in that: The processor includes a multi-core CPU and a multi-core GPU, which are used to execute the computer instructions. When the computer instructions are executed, the device implements the steps of the method according to any one of claims 1 to 7.

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

10. A computer program product comprising computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

Images