Method and apparatus for multi-contrast x-ray imaging with single exposure at arbitrary grating positions

Abstract

The application relates to the technical field of grating phase contrast imaging, and provides a single-exposure multi-contrast X-ray imaging method and device for arbitrary analysis grating position. The method constructs a neural network structure composed of a phase alignment front-end module and a conditional diffusion model, the former encodes single-exposure projection into orthogonal features aligned with the analysis grating phase; the multi-contrast reference data obtained by conventional stepping is used as supervision to train the conditional diffusion model to generate corresponding absorption, phase and dark-field contrast images from the phase alignment features in the latent space; the trained diffusion model is used to directly reconstruct high-quality multi-contrast X-ray images from single-exposure projection; the method effectively eliminates the multi-solution caused by the arbitrary grating position by explicitly modeling the analysis grating phase and introducing a diffusion generation model, realizes stable acquisition of high-quality multi-contrast imaging under the conditions of no stepping and low dose, reduces the mechanical complexity of the system, and improves the robustness and application potential of imaging.

Description

Technical Field

[0001] This application relates to the field of grating phase contrast imaging technology, and in particular to a method and apparatus for single-exposure multi-contrast X-ray imaging at arbitrary analytical grating positions. Background Technology

[0002] Compared to traditional absorption contrast imaging, X-ray grating interferometry based on the Talbot-Lau effect can obtain absorption contrast, phase contrast, and dark-field contrast images of an object in a single scan. Absorption contrast images have excellent imaging capabilities for high-density or high-atomic-number tissues such as metals, bones, and teeth; differential phase contrast images can significantly enhance the imaging contrast of light-element materials such as soft tissues; dark-field contrast images are most sensitive to low-density tissues containing microstructures, such as alveoli, and help reveal sub-pixel-scale structural information. The Talbot-Lau effect is a self-imaging interferometry phenomenon based on grating diffraction, combining the grating self-imaging (Talbot) effect with the double-grating incoherent interference (Lau) effect.

[0003] X-ray grating interferometry imaging systems based on the Talbot-Lau effect typically consist of a source grating G0, a phase grating G1, and an analysis grating G2. A conventional X-ray source, modulated by the source grating G0, forms a spatially coherent beam array. After passing through the phase grating G1 and propagating freely for a certain distance, it forms interference fringes on the plane of the analysis grating G2, which are then acquired by a two-dimensional detector. To separate absorption, phase, and dark-field contrast signals from the interference fringes, current techniques generally employ stepped scanning of the analysis grating G2. At each projection angle, G2 is shifted multiple times sub-periodically (typically 4–8 steps) to obtain a set of stepped projection sequences. The three contrast components are then extracted using methods such as Fourier analysis or phase-locked demodulation.

[0004] However, because the analysis grating absorbs most of the X-rays and multiple exposures are required for each projection angle to separate the contrast signal, the traditional step-scan method inevitably increases the radiation dose and scan time. Simultaneously, the multiple precise movements of the analysis grating also increase the complexity of the system structure and motion control. Directly reducing the tube current to lower the dose introduces stronger quantum noise into the projection data. Considering that Fourier analysis performs global calculations on the entire set of step sequences, once a local pixel in a frame of the sequence is contaminated with noise, this noise will propagate to the phase contrast image at the corresponding pixel location, resulting in significant image degradation; fine structures may be masked by noise, or even the overall anatomical structure may disappear.

[0005] Related technologies offer deep learning-based multi-contrast reconstruction methods, but most of these methods still rely on step-by-step acquisition at a fixed analysis grating position. They use convolutional neural networks to denoise and enhance low-dose step-by-step projections or reconstructed images, or learn a mapping from low-dose multi-step data to high-quality multi-step data. While these methods can alleviate noise issues to some extent, they still require multiple step-by-step exposures at each projection angle, making it difficult to fundamentally eliminate the dose and mechanical burden caused by step-by-step scanning. Furthermore, the phase position of the analysis grating within the field of view generally changes with geometric arrangement, temperature drift, and system bias. If the grating phase is not explicitly modeled and analyzed during network training and actual imaging, directly inferring the multi-contrast image from a single frame interferometric projection at any G2 position will result in multiple solutions. The reconstruction results are sensitive to the analysis grating position, exhibiting insufficient stability and generalization ability. Summary of the Invention

[0006] In view of this, the present application provides a method and apparatus for single-exposure multi-contrast X-ray imaging at arbitrary analysis grating positions, in order to solve the problems of high cost, complex operation and insufficient stability and generalization ability of multi-contrast X-ray imaging in the prior art.

[0007] A first aspect of this application provides a method for single-exposure multi-contrast X-ray imaging at arbitrary analysis grating positions, comprising:

[0008] Acquire multi-step interferometric projection and multi-contrast reference data; multi-step interferometric projection includes a multi-frame interferometric projection sequence of the sample under test and the background, and multi-contrast reference data includes absorption contrast image, phase contrast image and dark field contrast image;

[0009] A phase alignment front-end module is designed based on the analysis grating parameters. The phase alignment front-end module is used to perform interference phase alignment processing on multiple frames of interferometric projection sequences to obtain multi-channel features that are phase aligned with the analysis grating.

[0010] Construct a pre-trained conditional diffusion generator network; the pre-trained conditional diffusion generator network takes multi-channel features as input and outputs the absorption contrast image, phase contrast image and dark field contrast image of each frame of interferometric projection corresponding to the multi-channel features;

[0011] From the multi-frame interferometric projection sequence, single-frame interferometric projection is selected sequentially as the single exposure input, and the absorption contrast image, phase contrast image and dark field contrast image corresponding to the projection angle of the single-frame interferometric projection are used as supervision labels to construct a training sample set.

[0012] The training sample set is input into the pre-trained conditional diffusion generation network. The features obtained by the single-exposure input phase alignment front-end module are used as the conditional input. The phase alignment front-end module and the pre-trained conditional diffusion generation network are jointly trained to obtain the trained phase alignment front-end module and conditional diffusion generation network.

[0013] The single-exposure interferometric projection of the sample under test at any analytical grating position is obtained. The single-exposure interferometric projection is then sequentially input into the trained phase alignment front-end module and conditional diffusion generation network to obtain the absorption contrast image, phase contrast image and dark field contrast image of the sample under test.

[0014] A second aspect of this application provides a single-exposure multi-contrast X-ray imaging apparatus with arbitrary analysis grating positions, comprising:

[0015] The acquisition unit is configured to acquire multi-step interferometric projection and multi-contrast reference data; the multi-step interferometric projection includes a multi-frame interferometric projection sequence of the sample under test and the background, and the multi-contrast reference data includes absorption contrast image, phase contrast image and dark field contrast image.

[0016] The phase alignment unit is configured to design a phase alignment front-end module based on the analysis grating parameters. The phase alignment front-end module is used to perform interference phase alignment processing on multiple frames of interferometric projection sequences to obtain multi-channel features that are phase aligned with the analysis grating.

[0017] The building unit is configured to build a pre-trained conditional diffusion generation network; the pre-trained conditional diffusion generation network takes multi-channel features as input and outputs the absorption contrast image, phase contrast image and dark field contrast image of each frame of interferometric projection corresponding to the multi-channel features.

[0018] The building unit is also configured to sequentially select a single frame of interferometric projection from the multi-frame interferometric projection sequence as the single exposure input, and use the absorption contrast image, phase contrast image and dark field contrast image corresponding to the projection angle of the single frame interferometric projection as supervision labels to build a training sample set;

[0019] The training unit is configured to input the training sample set into the pre-trained conditional diffusion generation network, use the features obtained by the single-exposure input phase alignment front-end module as the conditional input, and jointly train the phase alignment front-end module and the pre-trained conditional diffusion generation network to obtain the trained phase alignment front-end module and conditional diffusion generation network.

[0020] The imaging unit is configured to acquire the single-exposure interferometric projection of the sample under test at any analysis grating position. The single-exposure interferometric projection is then sequentially input into the trained phase alignment front-end module and conditional diffusion generation network to obtain the absorption contrast image, phase contrast image and dark field contrast image of the sample under test.

[0021] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0022] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0023] The beneficial effects of this application embodiment compared with the prior art are as follows: This application embodiment obtains a single-exposure projection at any analysis grating position; constructs a neural network structure composed of a phase alignment front-end module and a conditional diffusion model, whereby the phase alignment front-end module explicitly encodes the single-exposure projection as an orthogonal feature aligned with the phase of the analysis grating; using a multi-contrast image obtained by conventional step acquisition as supervision, the conditional diffusion model is trained to generate corresponding absorption, phase, and dark-field contrast images from the phase alignment features in the latent space; and using the trained diffusion model, high-quality multi-contrast X-ray images can be directly reconstructed from the single-exposure projection. This invention, by explicitly modeling and analyzing the grating phase and introducing a diffusion generation model, effectively eliminates the ambiguity caused by arbitrary grating positions, achieving stable acquisition of high-quality multi-contrast imaging under stepless and low-dose conditions, reducing system mechanical complexity, and improving imaging robustness and application potential. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic flowchart of a single-exposure multi-contrast X-ray imaging method with arbitrary analysis grating position provided in an embodiment of this application.

[0026] Figure 2 This is a schematic diagram of the structure of the X-ray grating interferometric imaging device based on the Talbot-Lau effect provided in the embodiments of this application.

[0027] Figure 3 This is a schematic diagram of the structure of the joint model of the phase alignment front-end module and the pre-trained conditional diffusion generator network provided in the embodiments of this application.

[0028] Figure 4 This is a schematic flowchart of another single-exposure multi-contrast X-ray imaging method with arbitrary analysis grating position provided in the embodiments of this application.

[0029] Figure 5 This is a comparison image of multi-contrast imaging results under laboratory X-ray tube conditions provided in the embodiments of this application.

[0030] Figure 6 This is a schematic diagram of a single-exposure multi-contrast X-ray imaging device with arbitrary analysis grating position provided in an embodiment of this application.

[0031] Figure 7 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0032] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0033] The following describes in detail, with reference to the accompanying drawings, a single-exposure multi-contrast X-ray imaging method and apparatus for arbitrary analysis grating positions according to embodiments of this application.

[0034] As mentioned above, existing multi-contrast X-ray imaging suffers from high cost, complex operation, and insufficient stability and generalization ability. In view of this, embodiments of this application provide a single-exposure multi-contrast X-ray imaging method with arbitrary analysis grating positions. This method can stably reconstruct absorption, phase, and dark field contrast from a single-exposure projection at any analysis grating position. While reducing radiation dose and simplifying mechanical structure, it explicitly considers the phase information of the analysis grating, overcoming the ambiguity problem under single-exposure, and improving the robustness and practical application potential of multi-contrast X-ray imaging.

[0035] Figure 1 This is a schematic flowchart of a single-exposure multi-contrast X-ray imaging method with arbitrary analysis grating position provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0036] In step S101, multi-step interferometric projection and multi-contrast reference data are obtained.

[0037] Among them, the multi-step interferometric projection includes a multi-frame interferometric projection sequence of the sample under test and the background, and the multi-contrast reference data includes absorption contrast image, phase contrast image and dark field contrast image.

[0038] In step S102, a phase alignment front-end module is designed based on the analysis grating parameters. The phase alignment front-end module is used to perform interference phase alignment processing on the multi-frame interferometric projection sequence to obtain multi-channel features that are phase aligned with the analysis grating.

[0039] In step S103, a pre-trained conditional diffusion generator network is constructed.

[0040] The pre-trained conditional diffusion generation network takes multi-channel features as input and outputs the absorption contrast image, phase contrast image, and dark field contrast image of each frame of interferometric projection corresponding to the multi-channel features.

[0041] In step S104, a single frame of interferometric projection is selected sequentially from the multi-frame interferometric projection sequence as the single exposure input, and the absorption contrast image, phase contrast image and dark field contrast image corresponding to the projection angle of the single frame interferometric projection are used as supervision labels to construct a training sample set.

[0042] In step S105, the training sample set is input into the pre-trained conditional diffusion generation network. The features obtained by the single-exposure input phase alignment front-end module are used as conditional inputs. The phase alignment front-end module and the pre-trained conditional diffusion generation network are jointly trained to obtain the trained phase alignment front-end module and conditional diffusion generation network.

[0043] In step S106, the single-exposure interference projection of the sample under test at any analysis grating position is obtained. The single-exposure interference projection is then sequentially input into the trained phase alignment front-end module and conditional diffusion generation network to obtain the absorption contrast image, phase contrast image and dark field contrast image of the sample under test.

[0044] In some embodiments of this application, the method may be executed by a server or by a terminal device with certain processing capabilities.

[0045] In some embodiments of this application, multi-step interferometric projection and multi-contrast reference data can be obtained first; the multi-step interferometric projection includes a multi-frame interferometric projection sequence of the sample under test and the background, and the multi-contrast reference data includes absorption contrast image, phase contrast image and dark field contrast image.

[0046] Then, a phase alignment front-end module is designed based on the analysis grating parameters. The phase alignment front-end module is used to perform interference phase alignment processing on multiple frames of interferometric projection sequences to obtain multi-channel features that are phase aligned with the analysis grating.

[0047] Simultaneously, a pre-trained conditional diffusion generation network is constructed. This pre-trained conditional diffusion generation network takes multi-channel features as input and outputs the absorption contrast image, phase contrast image, and dark field contrast image of each frame of interferometric projection corresponding to the multi-channel features.

[0048] A single-frame interferometric projection can be selected sequentially from a multi-frame interferometric projection sequence as the input for a single exposure, and the absorption contrast image, phase contrast image, and dark field contrast image corresponding to the projection angle of the single-frame interferometric projection can be used as supervision labels to construct a training sample set.

[0049] Furthermore, the training sample set is input into the pre-trained conditional diffusion generation network, and the features obtained by the single-exposure input phase alignment front-end module are used as conditional inputs. The phase alignment front-end module and the pre-trained conditional diffusion generation network are jointly trained to obtain the trained phase alignment front-end module and conditional diffusion generation network.

[0050] After completing the joint training of the phase alignment front-end module and the pre-trained conditional diffusion generation network, the single-exposure interference projection of the sample under test at any analytical grating position can be obtained. The single-exposure interference projection is then sequentially input into the trained phase alignment front-end module and conditional diffusion generation network to obtain the absorption contrast image, phase contrast image and dark field contrast image of the sample under test.

[0051] According to the technical solution provided in the embodiments of this application, a single-exposure projection at any analysis grating position is obtained; a neural network structure consisting of a phase alignment front-end module and a conditional diffusion model is constructed. The phase alignment front-end module explicitly encodes the single-exposure projection into orthogonal features aligned with the phase of the analysis grating; using a multi-contrast image obtained by conventional step acquisition as supervision, the conditional diffusion model is trained to generate corresponding absorption, phase, and dark-field contrast images from the phase alignment features in the latent space; using the trained diffusion model, a high-quality multi-contrast X-ray image can be directly reconstructed from the single-exposure projection. This invention effectively eliminates the ambiguity caused by arbitrary grating positions by explicitly modeling and analyzing the grating phase and introducing a diffusion generation model, achieving stable acquisition of high-quality multi-contrast imaging under stepless and low-dose conditions, reducing the mechanical complexity of the system and improving the robustness and application potential of the imaging.

[0052] In some embodiments of this application, obtaining multi-step interferometric projection and multi-contrast reference data may include: using an X-ray grating interferometer based on the Talbot-Lau effect to control the analysis grating to perform multi-step scanning at multiple preset projection angles with predetermined step sizes, thereby obtaining a multi-frame interferometric projection sequence of the sample under test and the background; obtaining multi-contrast reference data through Fourier analysis based on the corresponding reference step sequence; the step of obtaining a single-exposure interferometric projection of the sample under test at any analysis grating position includes using an X-ray grating interferometer based on the Talbot-Lau effect to fix the analysis grating at an arbitrary phase position for scanning, thereby obtaining a single-exposure interferometric projection.

[0053] The corresponding reference step sequence refers to the reference step sequence associated with the multi-contrast reference data. Each set of multi-contrast reference data corresponds to one set of reference step sequences. The reference step sequence can be set according to actual needs and is not restricted here.

[0054] The Talbot-Lau effect is an optical interference phenomenon that combines the characteristics of the Talbot and Lau effects. It is typically achieved using a grating system under incoherent spatial illumination to generate interference fringes. The Talbot effect refers to the self-image formation of a periodic object at a specific location in the Fresnel diffraction region under monochromatic plane wave illumination. The Lau effect refers to the high-contrast bright and dark fringes observed at the rear of two parallel diffraction gratings with the same grating constant when illuminated by an extended light source.

[0055] Such as Figure 2 The X-ray grating interferometric imaging device based on the Talbot-Lau effect shown acquires multi-step interferometric projection and multi-contrast reference data. The device includes:

[0056] An X-ray source is used to emit incoherent or partially coherent X-rays.

[0057] The source grating G0 is used to split the broadband radiation from the X-ray source into multiple coherent sub-beams.

[0058] A phase grating G1 is arranged between the source grating G0 and the analysis grating G2 to modulate the phase of the incident coherent sub-beam, thereby forming the Talbot self-imaging effect.

[0059] Analysis grating G2, positioned in front of the detector, is used to convert the interference fringes generated by G1 into intensity modulation.

[0060] A two-dimensional X-ray detector is used to receive the intensity distribution of X-rays after passing through a sample and three gratings and output a digital projection image.

[0061] The grating motion mechanism is used to drive the analysis grating G2 to perform sub-periodic stepping along its periodic direction, realize multi-step interferometric acquisition, and fix the analysis grating G2 at any phase position to obtain single-exposure interferometric projection.

[0062] In the Talbot-Lau system, the source grating G0, phase grating G1, and analysis grating G2 are arranged sequentially along the optical path, with the sample placed between G1 and G2. The periods and spacing of the three gratings satisfy the Talbot-Lau geometric conditions, resulting in stable interference fringes forming on the plane of the analysis grating G2. By changing the translation of the analysis grating G2 in the periodic direction, interference fringe images at multiple phase positions can be obtained at each projection angle.

[0063] In some embodiments of this application, the phase alignment front-end module includes a normalization processing submodule, an orthogonal demodulation submodule, and a phase rotation submodule.

[0064] Using a phase alignment front-end module to perform interference phase alignment processing on a multi-frame interferometric projection sequence includes using the phase alignment front-end module to perform interference phase alignment processing on each single frame of the multi-frame interferometric projection sequence.

[0065] Furthermore, the phase alignment front-end module can be used to perform interferometric phase alignment processing on the target single-frame interferometric projection, which may include: using a normalization processing submodule to perform intensity normalization processing on the target single-frame interferometric projection to obtain a normalized interferometric fringe function; using an orthogonal demodulation submodule to perform phase-locked demodulation on the normalized interferometric fringes based on a preset reference fringe phase to obtain a pair of orthogonal components; using a phase rotation submodule to estimate the global phase shift of the analysis grating on the complex plane, and performing rotation alignment on the orthogonal classification based on the global phase shift to obtain the phase-aligned features of the target single-frame interferometric projection; wherein, the target single-frame interferometric projection is any single-frame interferometric projection in a multi-frame interferometric projection sequence.

[0066] In other words, the phase alignment front-end module is used to preprocess and phase normalize single-frame interferometric projections.

[0067] Specifically, the normalization processing submodule is used to perform intensity normalization processing on the target single-frame interferometric projection to obtain the normalized interferometric fringe function, which can be the intensity distribution of the interferometric projection containing the sample to be measured. and corresponding background projection intensity distribution Intensity normalization is performed to obtain the normalized interference fringe function. ;in, These are the pixel coordinates of the detector. Used to characterize the relative intensity change with respect to the background This is to determine the background or sample-free interference projection intensity distribution under the same system conditions, in order to reduce the effects of brightness drift and gain non-uniformity.

[0068] Using the orthogonal demodulation submodule to perform phase-locked demodulation of the normalized interference fringes based on a preset reference fringe phase can be achieved by combining the reference fringe phase obtained from system calibration with phase-locked demodulation of the normalized interference fringes to obtain a pair of orthogonal components:

[0069] ;in, This indicates the interference regulation envelope. The phase introduced for the sample; The reference fringe phase obtained for system calibration; To analyze the nominal phase of the grating at its current mechanical position.

[0070] Next, we can construct complex interference signals. In the preset area Internal Summation is used to estimate the global phase shift of the analytical grating. ;in, It is a complex argument function. It can be selected as a relatively uniform region far from the sample boundary.

[0071] Subsequently, for vectors Planar rotation is performed to align the phase of the analysis grating.

[0072] After the above processing, orthogonal components aligned with the phase of the analysis grating are obtained. and Furthermore, , With normalized interference fringes By splicing channels together, phase alignment conditional features are formed. This serves as the conditional input for the subsequent diffusion generation network.

[0073] In some embodiments of this application, the pre-trained conditional diffusion generation network includes at least an encoder, a diffusion denoising network, and a decoder. The encoder maps multi-contrast reference data to a low-dimensional latent representation; the diffusion denoising network denoises the forward diffusion noise latent vector corresponding to the low-dimensional latent representation to obtain a denoised latent vector; and the decoder decodes the denoised latent vector into reconstructed multi-contrast reference data.

[0074] In some embodiments of this application, joint training of the phase alignment front-end module and the pre-trained conditional diffusion generation network may include: for target training samples in the training sample set, determining the conditional features of the target single-frame interferometric projection through the phase alignment front-end module, and determining the low-dimensional latent representation of the target multi-contrast reference data through the encoder of the pre-trained conditional diffusion generation network; the target training sample is any training sample in the training sample set, the target training sample includes the target single-frame interferometric projection and the target multi-contrast reference data, and the conditional features of the target single-frame interferometric projection are the features of the target single-frame interferometric projection after phase alignment; randomly selecting from a preset discrete time step set A time step is selected, and a noise latent vector for that time step is generated through forward diffusion. A diffusion denoising network is used to predict injected noise based on noise regression loss, and the predicted injected noise is used to denoise the noise latent vector, resulting in a denoised latent vector. The denoised latent vector is then decoded into reconstructed target multi-contrast reference data, and the image domain reconstruction loss of the reconstructed target multi-contrast reference data and the target multi-contrast reference data is calculated. The sum of the noise regression loss and the image domain reconstruction loss is determined as the total loss. The phase alignment front-end module and the pre-trained conditional diffusion generation network are jointly iteratively optimized with the goal of minimizing the total loss until the iteration termination condition is met.

[0075] The iteration termination condition can be the convergence of the joint model consisting of the phase-aligned front-end module and the pre-trained conditional diffusion generation network, or the traversal of all single-frame interferometric projections and all supervision labels in the sample set, or the number of iterations exceeding a preset threshold, or other conditions.

[0076] In some implementations, the decoder is denoted as... The diffusion denoising network is denoted as The decoder is denoted as Meanwhile, the acquired target multi-contrast reference data is denoted as... The low-dimensional latent representation obtained by mapping the target's multi-contrast reference data into the encoder can be denoted as: .in, These are absorption contrast images, phase contrast images, and dark field contrast images, respectively, taken from the same viewpoint. It is a low-dimensional latent representation, which corresponds to a low-dimensional or multi-channel latent feature tensor.

[0077] The set of discrete time steps can be defined in the latent space of a pre-trained conditional diffusion generative network. , It is a positive integer greater than 1.

[0078] Time steps can be randomly selected from this set of discrete time steps. Time steps are generated through forward diffusion. The noise latent vector. Wherein, at time step generated by forward diffusion... Noise latent vector At that time, it is possible to represent low-dimensional potentials. Perform a forward diffusion process, gradually adding Gaussian noise during the process to obtain the potential noise vector at each time step. ;in, , To be compatible with low-dimensional latent representations Standard normal noise vectors of the same dimension satisfy , Standard Gaussian noise, It is the identity matrix; , is a preset noise attenuation coefficient sequence used to control the signal-to-noise ratio at each time step, and ∏ is the multiplication symbol.

[0079] The injected noise can then be predicted using the diffusion denoising network based on noise regression loss. This diffusion denoising network... A time-conditional U-Net structure can be adopted, by adjusting the time steps... Potential noise vector With conditional features The injected noise is obtained based on noise regression loss prediction as input. .in, For trainable parameters, For time step The noise latent vector. Internally, the network encodes time-step information through a time embedding module and introduces conditional features through feature concatenation or modulation mechanisms. This enables adaptive modeling of arbitrary analysis grating positions.

[0080] In some implementations, the noise regression loss can be ;in, The L2 norm represents a vector or feature tensor.

[0081] Can be time step Iterate to time step 0 and calculate the final denoised latent vector at time step 0. Then, regarding the Decoding yields the reconstructed image and compared with the reference multi-contrast image Calculate image domain reconstruction loss .in, , , These are the non-negative weighting coefficients corresponding to the three contrast levels. This represents the L1 norm.

[0082] In some examples, the total loss can be defined as By minimizing the total loss Parameters of diffusion denoising network And the parameters of the encoder and decoder are jointly optimized.

[0083] In some embodiments of this application, inputting the single-exposure interferometric projection sequentially into the trained phase-aligned front-end module and the conditional diffusion generation network to obtain the absorption contrast image, phase contrast image, and dark-field contrast image of the sample under test may include: using the trained phase-aligned front-end module to determine the conditional features of the single-exposure interferometric projection; the conditional features of the single-exposure interferometric projection are the features after phase alignment of the single-exposure interferometric projection; and defining a set of discrete time steps in the latent space of the trained conditional diffusion generation network. , A positive integer greater than 1; the initial latent vector is sampled from a standard Gaussian distribution. Since time step Iterate up to time step 0, at each time step The predicted injected noise for the current time step is calculated using the diffusion denoising network in the trained conditional diffusion generative network. And estimate the denoised latent vector at this time step as follows: And update the potential noise vector for the next time step: When the iteration reaches time step 0, the final denoised latent vector is obtained. ;Will The input decoder obtains the reconstructed multi-contrast reference data, which serves as the absorption contrast image, phase contrast image, and dark field contrast image of the single-exposure interferometric projection corresponding to the sample under test, thereby realizing single-exposure multi-contrast X-ray imaging at any analysis grating position.

[0084] Figure 3 This is a schematic diagram of the structure of the joint model of the phase alignment front-end module and the pre-trained conditional diffusion generative network provided in the embodiments of this application. Figure 3 As shown, a single-frame interferometric projection can be selected from the sample set as the input for a single exposure, and its corresponding absorption contrast image, phase contrast image, and dark field contrast image can be used as supervision labels. On the one hand, the single-frame interferometric projection can be input into the phase alignment module to obtain the alignment features of the single-frame interferometric projection; on the other hand, the absorption contrast image, phase contrast image, and dark field contrast image can be input into the pixel space of the pre-trained conditional diffusion generation network, encoded by the encoder to obtain a low-dimensional latent representation, and a forward diffusion process is performed in the latent space to obtain the noise latent vectors corresponding to different time steps.

[0085] Injected noise can be predicted using a noise regression loss based on the noise latent vector and alignment features. Furthermore, the predicted injected noise can be used to denoise the noise latent vector, yielding a denoised latent vector. Then, a sampling process is performed in the latent space based on this denoised latent vector to calculate the latent noise vector for the next time step.

[0086] The latent vector decoded at the next time step can be used as the reconstructed multi-contrast reference data, and the image domain reconstruction loss of the reconstructed multi-contrast reference data and the original multi-contrast reference data can be calculated to obtain the image loss.

[0087] During training, the sum of image loss and noise regression loss can be determined as the total loss. The phase alignment front-end module and the pre-trained conditional diffusion generation network are jointly iteratively optimized with the goal of minimizing the total loss until the iteration termination condition is met.

[0088] During the imaging process, time steps can be used to... Start iterating up to time step 0, and calculate the final denoised latent vector. And then through the decoder Decoding is performed to obtain reconstructed multi-contrast reference data, which is used as a single-exposure multi-contrast X-ray image of the sample under test.

[0089] Figure 4 This is a schematic flowchart of another single-exposure multi-contrast X-ray imaging method with arbitrary analysis grating position provided in an embodiment of this application. For example... Figure 4 As shown, the embodiments of this application perform the following steps in the training phase and the generation phase respectively:

[0090] Training sample construction involves selecting a single frame of interferometric projection from the corresponding multi-step interferometric projection sequence for each projection angle as the input for a single exposure, and then calculating the conditional features using the phase alignment front-end module. .

[0091] Absorption, phase, and dark-field contrast images obtained using multi-step Fourier analysis As a label, input encoder Obtaining latent representations , constitute training sample pairs .

[0092] In each training iteration, a time step is randomly selected. Generate according to the forward diffusion formula And enable the diffusion denoising network to predict noise. Among them, the noise regression loss is .

[0093] Simultaneously, the denoised latent representation is decoded to obtain the reconstructed image. The image domain reconstruction loss is obtained by using weighted L1 loss to constrain the image reconstruction quality of three contrasts. .

[0094] The total loss is determined to be By minimizing The parameters of the phase alignment front-end module, encoder, diffusion denoising network, and decoder are jointly trained.

[0095] In actual imaging, at any analysis grating position, only one frame of interferometric projection of the sample under test can be acquired as the single exposure input, and the conditional features can be calculated by the phase alignment front-end module. .

[0096] In the latent space, initial latent vectors are sampled from a standard Gaussian distribution. Since time step Iterate up to time step 0, and at each time step, use the trained diffusion denoising network to calculate the noise prediction. And estimate and update the noise-free latent representation as follows: , .

[0097] When iterates to At that time, the final denoised latent representation is obtained. ;Will Input Decoder The reconstructed multi-contrast image is obtained. This refers to absorption contrast images, phase contrast images, and dark field contrast images under single-exposure conditions, enabling single-exposure multi-contrast X-ray imaging at any analytical grating position.

[0098] To verify the technical effectiveness of the technical solution provided in the embodiments of this application, the following experiment was designed: A laboratory X-ray tube, a matching Talbot-Lau grating interferometer system, and small animal samples were used to verify the technical solution provided in the embodiments of this application. High-quality multi-contrast reference images were obtained through multi-step scanning and Fourier analysis as labels, and the conditional diffusion generation network was trained using the above training process. After training, at multiple different analysis grating positions, only a single frame of interferometric projection was acquired and input into the phase alignment front-end module and conditional diffusion generation network of the embodiments of this application to obtain the single-exposure multi-contrast reconstruction result.

[0099] Figure 5 These are comparison images of multi-contrast imaging results under laboratory X-ray tube conditions, provided in the embodiments of this application. Figure 5 As shown, the left column is the "true value" multi-contrast image obtained by multi-step high-dose Fourier analysis, the middle column is the reconstruction result under single-exposure conditions using the technical solution provided in the embodiments of this application, and the right column is the corresponding single-exposure low-dose input projection (the position of the analysis grating used during training is shown below).

[0100] A comparison of the three rows of absorption, phase, and dark field data shows that the middle column "Results" and the left column "True Values" are basically consistent in terms of structure, such as bone contours, lung textures, and dark field scattering details. Compared with the low-dose input on the right, noise particles are significantly reduced and grayscale levels are smoother and more continuous. Therefore, it can be intuitively reflected that this application has restored image quality close to that of a high-dose reference under single exposure and low-dose conditions, and significantly suppressed quantum noise.

[0101] In addition, the phase position of the interference fringes can be seen in the two single-exposure projections on the right, "Input" and "Training", which means that the position of the analysis grating is different during testing and training. However, the reconstruction results in the middle column are still in good agreement with the "True Value" in the left column. This shows that the phase alignment front end and the conditional diffusion model can compensate for the differences caused by different analysis grating phases and have good stability and robustness for arbitrary grating positions.

[0102] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0103] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0104] Figure 6 This is a schematic diagram of a single-exposure multi-contrast X-ray imaging device with arbitrary analysis grating position provided in an embodiment of this application. Figure 6 As shown, the device includes:

[0105] The acquisition unit 601 is configured to acquire multi-step interferometric projection and multi-contrast reference data; the multi-step interferometric projection includes a multi-frame interferometric projection sequence of the sample under test and the background, and the multi-contrast reference data includes an absorption contrast image, a phase contrast image and a dark field contrast image.

[0106] The phase alignment unit 602 is configured to design a phase alignment front-end module based on the analysis grating parameters. The phase alignment front-end module is used to perform interference phase alignment processing on a multi-frame interferometric projection sequence to obtain multi-channel features that are phase aligned with the analysis grating.

[0107] The building unit 603 is configured to build a pre-trained conditional diffusion generation network. The pre-trained conditional diffusion generation network takes multi-channel features as input and outputs the absorption contrast image, phase contrast image and dark field contrast image of each frame of interferometric projection corresponding to the multi-channel features.

[0108] The construction unit 603 is also configured to sequentially select single-frame interferometric projections from the multi-frame interferometric projection sequence as single-exposure inputs, and use the absorption contrast image, phase contrast image and dark field contrast image corresponding to the projection angle of the single-frame interferometric projection as supervision labels to construct a training sample set.

[0109] Training unit 604 is configured to input the training sample set into the pre-trained conditional diffusion generation network, use the features obtained by the single-exposure input phase alignment front-end module as the conditional input, and jointly train the phase alignment front-end module and the pre-trained conditional diffusion generation network to obtain the trained phase alignment front-end module and conditional diffusion generation network.

[0110] Imaging unit 605 is configured to acquire single-exposure interference projection of the sample under test at any analysis grating position, and input the single-exposure interference projection sequentially into the trained phase alignment front-end module and conditional diffusion generation network to obtain absorption contrast image, phase contrast image and dark field contrast image of the sample under test.

[0111] According to the technical solution provided in the embodiments of this application, a single-exposure projection at any analysis grating position is obtained; a neural network structure consisting of a phase alignment front-end module and a conditional diffusion model is constructed. The phase alignment front-end module explicitly encodes the single-exposure projection into orthogonal features aligned with the phase of the analysis grating; using a multi-contrast image obtained by conventional step acquisition as supervision, the conditional diffusion model is trained to generate corresponding absorption, phase, and dark-field contrast images from the phase alignment features in the latent space; using the trained diffusion model, a high-quality multi-contrast X-ray image can be directly reconstructed from the single-exposure projection. This invention effectively eliminates the ambiguity caused by arbitrary grating positions by explicitly modeling and analyzing the grating phase and introducing a diffusion generation model, achieving stable acquisition of high-quality multi-contrast imaging under stepless and low-dose conditions, reducing the mechanical complexity of the system and improving the robustness and application potential of the imaging.

[0112] In some implementations, acquiring multi-step interferometric projection and multi-contrast reference data includes: using an X-ray grating interferometer based on the Talbot-Lau effect to control the analysis grating to perform multi-step scanning at multiple preset projection angles with predetermined step sizes, thereby acquiring a multi-frame interferometric projection sequence of the sample under test and the background; obtaining multi-contrast reference data through Fourier analysis based on the corresponding reference step sequence; and acquiring a single-exposure interferometric projection of the sample under test at any analysis grating position by using an X-ray grating interferometer based on the Talbot-Lau effect to fix the analysis grating at an arbitrary phase position for scanning, thereby acquiring a single-exposure interferometric projection.

[0113] In some implementations, the phase alignment front-end module includes a normalization processing submodule, an orthogonal demodulation submodule, and a phase rotation submodule. Performing interference phase alignment processing on a multi-frame interferometric projection sequence using the phase alignment front-end module includes performing interference phase alignment processing on each single-frame interferometric projection in the multi-frame interferometric projection sequence separately. Performing interference phase alignment processing on a target single-frame interferometric projection using the phase alignment front-end module includes: performing intensity normalization processing on the target single-frame interferometric projection using the normalization processing submodule to obtain a normalized interference fringe function; performing phase-locked demodulation on the normalized interference fringes based on a preset reference fringe phase using the orthogonal demodulation submodule to obtain a pair of orthogonal components; and estimating the global phase shift of the analysis grating on the complex plane using the phase rotation submodule, and rotating and aligning the orthogonal classification based on the global phase shift to obtain the phase-aligned features of the target single-frame interferometric projection. Wherein, the target single-frame interferometric projection is any single-frame interferometric projection in the multi-frame interferometric projection sequence.

[0114] In some implementations, the pre-trained conditional diffusion generation network includes at least an encoder, a diffusion denoising network, and a decoder; the encoder is used to map multi-contrast reference data to a low-dimensional latent representation; the diffusion denoising network is used to denoise the forward diffusion noise latent vector corresponding to the low-dimensional latent representation to obtain a denoised latent vector; and the decoder is used to decode the denoised latent vector into reconstructed multi-contrast reference data.

[0115] In some implementations, the phase alignment front-end module and the pre-trained conditional diffusion generation network are jointly trained, including: for target training samples in the training sample set, the phase alignment front-end module determines the conditional features of the target single-frame interferometric projection, and the encoder of the pre-trained conditional diffusion generation network determines the low-dimensional latent representation of the target multi-contrast reference data; the target training sample is any training sample in the training sample set, the target training sample includes the target single-frame interferometric projection and the target multi-contrast reference data, and the conditional features of the target single-frame interferometric projection are the features of the target single-frame interferometric projection after phase alignment; and the time step is randomly selected from a preset discrete time step set. In the intermediate step, a noise latent vector is generated at the time step through forward diffusion; the injected noise is predicted using a diffusion denoising network based on noise regression loss, and the predicted injected noise is used to denoise the noise latent vector to obtain a denoised latent vector; the denoised latent vector is decoded into reconstructed target multi-contrast reference data, and the image domain reconstruction loss of the reconstructed target multi-contrast reference data and the target multi-contrast reference data is calculated; the sum of the noise regression loss and the image domain reconstruction loss is determined as the total loss, and the phase alignment front-end module and the pre-trained conditional diffusion generation network are jointly iteratively optimized with the goal of minimizing the total loss until the iteration termination condition is met.

[0116] In some implementations, generating noisy latent vectors for time steps via forward diffusion includes defining a set of discrete time steps in the latent space of a pre-trained conditional diffusion generation network. , It is a positive integer greater than 1; for low-dimensional latent representations Perform a forward diffusion process, gradually adding Gaussian noise during the process to obtain the potential noise vector at each time step. ;in, , To be compatible with low-dimensional latent representations Standard normal noise vectors of the same dimension satisfy , Standard Gaussian noise, It is the identity matrix; , The preset noise attenuation coefficient sequence is used to control the signal-to-noise ratio at each time step, where ∏ represents the multiplication sign; the noise regression loss is... ;in, The L2 norm of a vector or feature tensor; For time steps The corresponding predicted injection noise, For the trainable parameters of the diffusion denoising network, For time steps The noise latent vector; The conditional features of the target single-frame interferometric projection.

[0117] In some implementations, the single-exposure interferometric projection is sequentially input into a trained phase-aligned front-end module and a conditional diffusion generation network to obtain an absorption contrast image, a phase contrast image, and a dark-field contrast image of the sample under test. This includes: using the trained phase-aligned front-end module to determine the conditional features of the single-exposure interferometric projection; the conditional features of the single-exposure interferometric projection are the features after phase alignment of the single-exposure interferometric projection; and defining a set of discrete time steps in the latent space of the trained conditional diffusion generation network. , A positive integer greater than 1; the initial latent vector is sampled from a standard Gaussian distribution. Since time step Iterate up to time step 0, at each time step The predicted injected noise for the current time step is calculated using the diffusion denoising network in the trained conditional diffusion generative network. And estimate the denoised latent vector at this time step as follows: And update the potential noise vector for the next time step: When the iteration reaches time step 0, the final denoised latent vector is obtained. ;Will The input decoder obtains the reconstructed multi-contrast reference data, which serves as the absorption contrast image, phase contrast image, and dark field contrast image of the single-exposure interferometric projection of the sample under test.

[0118] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0119] Figure 7 This is a schematic diagram of the electronic device provided in an embodiment of this application. For example... Figure 7 As shown, the electronic device 7 of this embodiment includes a processor 701, a memory 702, and a computer program 703 stored in the memory 702 and executable on the processor 701. When the processor 701 executes the computer program 703, it implements the steps in the various method embodiments described above. Alternatively, when the processor 701 executes the computer program 703, it implements the functions of each module / unit in the various device embodiments described above.

[0120] Electronic device 7 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 7 may include, but is not limited to, processor 701 and memory 702. Those skilled in the art will understand that... Figure 7 This is merely an example of electronic device 7 and does not constitute a limitation on electronic device 7. It may include more or fewer components than shown, or different components.

[0121] The processor 701 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0122] The memory 702 can be an internal storage unit of the electronic device 7, such as a hard disk or RAM of the electronic device 7. The memory 702 can also be an external storage device of the electronic device 7, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the electronic device 7. The memory 702 can also include both internal and external storage units of the electronic device 7. The memory 702 is used to store computer programs and other programs and data required by the electronic device.

[0123] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0124] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0125] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for single-exposure multi-contrast X-ray imaging at arbitrary analysis grating positions, characterized in that, include: Acquire multi-step interferometric projection and multi-contrast reference data; The multi-step interferometric projection includes a multi-frame interferometric projection sequence of the sample under test and the background, and the multi-contrast reference data includes absorption contrast image, phase contrast image and dark field contrast image; A phase alignment front-end module is designed based on the analysis grating parameters. The phase alignment front-end module is used to perform interference phase alignment processing on the multi-frame interference projection sequence to obtain multi-channel features that are phase aligned with the analysis grating. Construct a pre-trained conditional diffusion generation network; the pre-trained conditional diffusion generation network takes the multi-channel features as input and outputs the absorption contrast image, phase contrast image and dark field contrast image of each frame of interferometric projection corresponding to the multi-channel features; From the multi-frame interferometric projection sequence, single-frame interferometric projections are selected sequentially as single-exposure inputs, and the absorption contrast image, phase contrast image, and dark field contrast image corresponding to the projection angle of the single-frame interferometric projection are used as supervision labels to construct a training sample set. The training sample set is input into the pre-trained conditional diffusion generation network. The features obtained by the single-exposure input phase alignment front-end module are used as conditional inputs. The phase alignment front-end module and the pre-trained conditional diffusion generation network are jointly trained to obtain the trained phase alignment front-end module and conditional diffusion generation network. The single-exposure interference projection of the sample under test at any analytical grating position is obtained. The single-exposure interference projection is then sequentially input into the trained phase alignment front-end module and the conditional diffusion generation network to obtain the absorption contrast image, phase contrast image and dark field contrast image of the sample under test.

2. The method according to claim 1, characterized in that, Acquire multi-step interferometric projection and multi-contrast reference data, including: Using an X-ray grating interferometer based on the Talbot-Lau effect, the analysis grating is controlled to perform multi-step scanning at a predetermined step size at multiple preset projection angles to obtain a multi-frame interferometric projection sequence of the sample and background. The multi-contrast reference data is obtained through Fourier analysis based on the corresponding reference step sequence; The steps for obtaining the single-exposure interference projection of the sample under test at any analytical grating position include using an X-ray grating interferometer based on the Talbot-Lau effect to scan the analytical grating at an arbitrary phase position to obtain the single-exposure interference projection.

3. The method according to claim 1, characterized in that, The phase alignment front-end module includes a normalization processing submodule, an orthogonal demodulation submodule, and a phase rotation submodule; Using the phase alignment front-end module to perform interference phase alignment processing on the multi-frame interference projection sequence includes using the phase alignment front-end module to perform interference phase alignment processing on each single frame interference projection in the multi-frame interference projection sequence. The phase alignment front-end module is used to perform interferometric phase alignment processing on the target single-frame interferometric projection, including: The normalization processing submodule is used to perform intensity normalization processing on the target single-frame interferometric projection to obtain the normalized interferometric fringe function; The orthogonal demodulation submodule is used to perform phase-locked demodulation on the normalized interference fringes based on a preset reference fringe phase to obtain a pair of orthogonal components; The phase rotation submodule is used to estimate the global phase shift of the analysis grating in the complex plane, and the orthogonal classification is rotated and aligned based on the global phase shift to obtain the phase-aligned features of the target single-frame interferometric projection. The target single-frame interferometric projection is any single-frame interferometric projection in the multi-frame interferometric projection sequence.

4. The method according to claim 1, characterized in that, The pre-trained conditional diffusion generation network includes at least an encoder, a diffusion denoising network, and a decoder; The encoder is used to map the multi-contrast reference data into a low-dimensional latent representation; The diffusion denoising network is used to denoise the forward diffusion noise latent vector corresponding to the low-dimensional latent representation to obtain the denoised latent vector. The decoder is used to decode the denoised latent vectors into reconstructed multi-contrast reference data.

5. The method according to claim 4, characterized in that, Joint training of the phase-aligned front-end module and the pre-trained conditional diffusion generation network includes: For the target training samples in the training sample set, the conditional features of the target single-frame interferometric projection are determined by the phase alignment front-end module, and the low-dimensional latent representation of the target multi-contrast reference data is determined by the encoder of the pre-trained conditional diffusion generation network; the target training sample is any training sample in the training sample set, the target training sample includes the target single-frame interferometric projection and the target multi-contrast reference data, and the conditional features of the target single-frame interferometric projection are the features of the target single-frame interferometric projection after phase alignment; A time step is randomly selected from a preset set of discrete time steps, and a noise potential vector for the time step is generated through forward diffusion. The diffusion denoising network is used to predict injected noise based on noise regression loss, and the predicted injected noise is used to denoise the noise latent vector to obtain the denoised latent vector. The denoised latent vector is decoded into reconstructed target multicontrast reference data, and the image domain reconstruction loss of the reconstructed target multicontrast reference data and the target multicontrast reference data is calculated. The sum of the noise regression loss and the image domain reconstruction loss is determined as the total loss. The phase alignment front-end module and the pre-trained conditional diffusion generation network are jointly iteratively optimized with the goal of minimizing the total loss until the iteration termination condition is met.

6. The method according to claim 5, characterized in that, Noise latent vectors for each time step are generated through forward diffusion, including: Define a set of discrete time steps in the latent space of the pre-trained conditional diffusion generative network. , It is a positive integer greater than 1; For the low-dimensional latent representation Perform a forward diffusion process, gradually adding Gaussian noise during the process to obtain the potential noise vector at each time step. ;in, , To be compatible with low-dimensional latent representations Standard normal noise vectors of the same dimension satisfy , Standard Gaussian noise, It is the identity matrix; , is a preset noise attenuation coefficient sequence used to control the signal-to-noise ratio at each time step, and ∏ is the multiplication symbol; The noise regression loss is ;in, The L2 norm of a vector or feature tensor; For time steps The corresponding predicted injection noise, For the trainable parameters of the diffusion denoising network, For time steps The noise latent vector; These are the conditional features of the target single-frame interferometric projection.

7. The method according to claim 6, characterized in that, The single-exposure interferometric projection is sequentially input into the trained phase-aligned front-end module and the conditional diffusion generation network to obtain the absorption contrast image, phase contrast image, and dark-field contrast image of the sample under test, including: The conditional features of the single-exposure interferometric projection are determined using the trained phase-aligned front-end module; the conditional features of the single-exposure interferometric projection are the features after phase alignment of the single-exposure interferometric projection. Define a set of discrete time steps in the latent space of the trained conditional diffusion generative network. , It is a positive integer greater than 1; Initial latent vectors are sampled from a standard Gaussian distribution. Since time step Iterate up to time step 0, at each time step The predicted injected noise for the current time step is calculated using the diffusion denoising network in the trained conditional diffusion generative network. And estimate the denoised latent vector at this time step as follows: And update the potential noise vector for the next time step: ; When the iteration reaches time step 0, the final denoised latent vector is obtained. ; Will The input decoder obtains the reconstructed multi-contrast reference data, which serves as the absorption contrast image, phase contrast image, and dark field contrast image of the single-exposure interference projection corresponding to the sample under test.

8. A single-exposure multi-contrast X-ray imaging device with arbitrary analysis grating position, characterized in that, include: The acquisition unit is configured to acquire multi-step interferometric projection and multi-contrast reference data; The multi-step interferometric projection includes a multi-frame interferometric projection sequence of the sample under test and the background, and the multi-contrast reference data includes absorption contrast image, phase contrast image and dark field contrast image; The phase alignment unit is configured to design a phase alignment front-end module based on the analysis grating parameters, and use the phase alignment front-end module to perform interference phase alignment processing on the multi-frame interferometric projection sequence to obtain multi-channel features that are phase aligned with the analysis grating. The building unit is configured to build a pre-trained conditional diffusion generation network; the pre-trained conditional diffusion generation network takes the multi-channel features as input and outputs the absorption contrast image, phase contrast image and dark field contrast image of each frame of interferometric projection corresponding to the multi-channel features; The construction unit is further configured to sequentially select a single frame of interferometric projection from the multi-frame interferometric projection sequence as a single exposure input, and use the absorption contrast image, phase contrast image and dark field contrast image corresponding to the projection angle of the single frame of interferometric projection as supervision labels to construct a training sample set; The training unit is configured to input the training sample set into the pre-trained conditional diffusion generation network, use the features obtained by the single-exposure input phase alignment front-end module as conditional input, and jointly train the phase alignment front-end module and the pre-trained conditional diffusion generation network to obtain the trained phase alignment front-end module and conditional diffusion generation network. The imaging unit is configured to acquire a single-exposure interferometric projection of the sample under test at any analytical grating position, and to sequentially input the single-exposure interferometric projection into the trained phase alignment front-end module and the conditional diffusion generation network to obtain the absorption contrast image, phase contrast image and dark field contrast image of the sample under test.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images