Static x-ray multi-contrast imaging method based on a physical sampling process diffusion model

Abstract

The application discloses a static X-ray multi-contrast imaging method based on a physical sampling process diffusion model, and belongs to the technical field of deep learning and X-ray grating interference imaging. The method comprises the following steps: acquiring a phase stepping projection graph of X-ray grating interference imaging to construct a training data set; constructing a diffusion model based on a physical sampling process, wherein the model takes a single-exposure projection graph as input and outputs multiple phase stepping images; training the diffusion model by using the training data set, so that the diffusion model fits the real physical sampling process in the diffusion process, and the physical consistency and accuracy of the generated result are enhanced by using a loss function combining image information and physical information; and finally, the single X-ray exposure projection graph is input into the trained model to generate the remaining phase stepping projection graphs, and three kinds of contrast images, namely absorption, phase and dark field images, are parsed. The application can realize high-quality multi-contrast imaging through single exposure, and significantly reduces the radiation dose and imaging time.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and X-ray grating interferometric imaging technology, specifically involving a static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process. Background Technology

[0002] In X-ray imaging, grating interferometry has attracted widespread attention due to its ability to achieve multi-contrast imaging using conventional X-ray sources. Based on the Talbot-Lau interferometry principle, this method analyzes the attenuation, phase shift, and small-angle scattering effects produced by the interaction of X-rays with matter to extract three types of contrast information from the sample: absorption, phase, and dark field. Absorption contrast is highly effective in distinguishing tissues of different densities, but has limited contrast for low-density materials; phase contrast, on the other hand, has extremely high sensitivity to low-density structures and can effectively improve imaging contrast; while dark field contrast, based on scattering mechanisms, has excellent imaging capabilities for the microstructure and edge features within the sample, thus achieving multi-dimensional and complementary characterization of the examined object.

[0003] However, traditional X-ray grating interferometry imaging methods typically require multiple step-by-step movements and exposure acquisitions within one cycle of the analysis grating to extract multi-contrast information; this is known as "phase-stepping" technology. This multi-step acquisition mechanism inevitably leads to prolonged imaging time and a significant increase in radiation dose. In applications such as clinical diagnosis and industrial non-destructive testing, where time efficiency and radiation safety are paramount, these drawbacks severely restrict the widespread adoption and application of this technology.

[0004] To address the aforementioned issues, resolving multi-contrast information in a single exposure has become a crucial research direction in this field. Currently, some techniques attempt to circumvent the need for multi-step sampling by improving imaging system hardware, such as employing specially designed gratings or detectors. However, these hardware improvements often come at the cost of increased system cost and structural complexity, and may even lead to a decrease in image quality, thus limiting their universality and practicality.

[0005] Therefore, developing a computational method that can accurately and efficiently resolve multicontrast information using only a single exposure without altering existing imaging hardware is crucial for promoting the practical application of low-dose, fast X-ray grating imaging technology. Currently, no research has introduced diffusion models based on physical sampling processes into this field to utilize the physical information contained in the phase step curve as constraints to guide high-quality reconstruction of multicontrast images from single-exposure images. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process. This method achieves high-quality multi-contrast imaging with only a single exposure, without requiring changes to the imaging hardware. It significantly reduces radiation dose and imaging time, and has good interpretability and generalization.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A static X-ray multicontrast imaging method based on a diffusion model of the physical sampling process, the method comprising:

[0009] Step S1: Obtain the phase-stepped projection map of X-ray grating interferometry imaging and construct the training dataset;

[0010] Step S2: Construct a diffusion model based on the physical sampling process. The diffusion model includes a backbone network for converting a single X-ray exposure projection image into multiple phase-stepped images.

[0011] Step S3: Train the diffusion model using the training dataset. The training process includes multi-step diffusion that simulates the physical sampling process and loss function calculation that combines image information and physical information.

[0012] Step S4: Input the single X-ray exposure projection image of the object under test into the trained diffusion model to generate multiple corresponding phase step projection images, and analyze the X-ray absorption, phase and dark field contrast images.

[0013] Furthermore, in step S2, the backbone network adopts the Unet architecture, with its input being a single-channel single X-ray exposure projection image and its output being a k-channel image, where k is the total number of phase steps.

[0014] Furthermore, the backbone network of the Unet architecture is composed of multiple convolutional layers connected sequentially, and residual connections are set between at least two sets of non-adjacent convolutional layers.

[0015] Furthermore, in step S3, the multi-step diffusion process includes:

[0016] Single-step diffusion step: Input the single X-ray exposure projection image into the backbone network to obtain multiple initial phase-stepped images; perform a fast Fourier transform on the multiple phase-stepped images to obtain an initial phase-stepped exposure curve; perform a virtual phase-stepping operation on the phase-stepped exposure curve according to the physical sampling rules to generate a single X-ray exposure projection image representing the next sampling position; use the newly generated single X-ray exposure projection image as input and re-input it into the backbone network to generate a series of intermediate projection images in sequence;

[0017] Repeat the single-step diffusion step t times to finally output the complete multi-step projection map predicted by the model.

[0018] Furthermore, in step S3, the loss function is composed of the image information loss and physical information loss in each diffusion process. The image information loss calculates the difference in pixel values ​​between the output image of the diffusion model and the real image, while the physical information loss calculates the difference in physical parameters between the output image of the diffusion model and the real image after Fourier transform.

[0019] Furthermore, the image information loss is the mean square error of the pixel values ​​at corresponding phase positions between the phase step projection map generated by the model and the real phase step projection map.

[0020] Furthermore, the physical information loss is the mean square error between the 0th-order amplitude, 1st-order amplitude, and 1st-order phase obtained from the phase step image generated by the model after fast Fourier transform, and the corresponding physical parameters obtained from the real phase step projection image after fast Fourier transform.

[0021] Furthermore, step S3 also includes updating the parameters of the backbone network based on the loss function calculation results after the loss function calculation is completed, using a stochastic gradient descent optimization algorithm.

[0022] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned static X-ray multicontrast imaging method based on a physical sampling process diffusion model.

[0023] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned static X-ray multicontrast imaging method based on a diffusion model of a physical sampling process.

[0024] The beneficial effects of this invention are as follows:

[0025] True single-exposure multi-contrast imaging is achieved: Compared with the prior art, the present invention does not require any modification to hardware such as grating interferometers or detectors. Only one X-ray exposure is needed to generate a complete phase-stepped projection sequence through a trained diffusion model, and the absorption, phase and dark field contrast images can be resolved with high precision, which greatly reduces the radiation dose and time cost required for imaging.

[0026] The model improves the physical accuracy and interpretability of the results: its core innovation lies in embedding the physical sampling process into the diffusion model. By constructing a loss function that integrates image and physical information, the network-generated results are strictly constrained by the physical laws of the real phase step curve, ensuring the physical rationality and reliability of the generated images and overcoming the shortcomings of poor generalization and weak interpretability of purely data-driven methods.

[0027] It possesses excellent versatility and application flexibility: After training, this method can seamlessly switch between conventional phase-stepping imaging mode and low-dose rapid single-exposure mode, which can flexibly adapt to the needs of different imaging scenarios (such as conventional detection and dynamic or dose-sensitive applications), and powerfully promote the practical application and popularization of X-ray multi-contrast imaging technology. Attached Figure Description

[0028] Figure 1 This is a flowchart of the static X-ray multi-contrast imaging method based on the diffusion model of the physical sampling process of the present invention;

[0029] Figure 2 This is a schematic diagram of the imaging principle of an X-ray grating multi-contrast imaging system.

[0030] Figure 3 This is a schematic diagram of the training process of the diffusion model based on the physical sampling process of the present invention;

[0031] Figure 4 This is a schematic diagram illustrating the process of obtaining a multi-contrast image from a single-exposure X-ray projection image based on a diffusion model of the physical sampling process according to the present invention.

[0032] Figure 5 This refers to a set of data obtained by using the method provided by this invention to collect a dataset;

[0033] Figure 6 Comparative experimental results of static X-ray multicontrast imaging using a trained diffusion model based on a physical sampling process, provided by the present invention. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0035] Figure 1This invention provides a flowchart of a static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process. Compared to existing single-exposure X-ray multi-contrast imaging methods, which mostly require modifications to hardware such as grating interferometers and detectors and additional debugging, the method provided by this invention can be implemented on a common X-ray three-grating interferometer, and the results conform to the actual physical sampling process. A single exposure of the object under test can resolve the absorption, phase, and dark field contrast signals. Essentially, this method is an underdetermined signal analysis method operating in the image domain. It uses a diffusion model to fit the contrast signal, resulting in higher interpretability and accuracy of physical information. Compared to traditional diffusion models, this method only borrows the idea of ​​diffusion. It is completely different from traditional diffusion models based on Markov chains or bridge models and other probabilistic models. This method combines the physically-based grating stepping process with diffusion, starting from a projection and diffusing to all phase steps, then using physical loss to ensure the accuracy of the diffusion process. The specific steps of this method are as follows:

[0036] Step S1: Obtain the phase-stepped projection map of X-ray grating interferometry imaging and construct the training dataset.

[0037] Figure 2 This is a schematic diagram of an X-ray grating multi-contrast imaging system; the X-ray grating multi-contrast imaging system, based on a Talbot-Lau interferometer, comprises six parts: an X-ray source 20, a source grating 21, an object under inspection 22, a phase grating 23, an absorption grating 24, and a detector 25. The parameters of the Talbot-Lau interferometer should satisfy the following formulas (1)-(4):

[0038] (1)

[0039] (2)

[0040] (3)

[0041] (4)

[0042] in, This indicates the distance between the phase grating 23 and the absorption grating 24; For magnification ratio, The distance between the source grating 21 and the phase grating 23; Indicates the first Talbot distance of order fraction; For the period of phase grating 23, The wavelength of X-rays. For the period of absorption grating 24, For the period of source grating 21, The width of the source grating 21 that allows X-rays to pass through in each cycle.

[0043] During data acquisition, multiple phase stepping projection images are obtained using phase stepping scans to form the labels of the training dataset. Then, the first stepping position of each stepping scan is taken as a single projection image in the training dataset to form the training dataset based on the diffusion model of the physical sampling process.

[0044] Step S2: Construct a diffusion model based on the physical sampling process. The diffusion model includes a backbone network for converting a single X-ray exposure projection image into multiple phase-stepped images.

[0045] The diffusion model based on the physical sampling process consists of a backbone network. This backbone network receives one input channel and outputs multiple channels, converting a single X-ray exposure projection image into multiple phase-stepped images. This process can be represented as follows: Where X represents a single X-ray exposure projection image, and Y represents multiple phase-stepped images output by the backbone network. This represents all parameters in the backbone network. In this embodiment, the Unet architecture is used as the backbone network. The Unet backbone network consists of five sequentially connected convolutional layers. Each convolutional layer uses a 3×3 kernel with a stride of 1, and maintains the spatial resolution of the feature maps unchanged (i.e., no downsampling is used). To enhance feature reuse and information flow, two residual connections are introduced into the network: 1. The output feature map of the second convolutional layer is superimposed on the output feature map of the fifth convolutional layer by element-wise addition. 2. The output feature map of the third convolutional layer is superimposed on the output feature map of the fourth convolutional layer by element-wise addition. This design allows the network to directly utilize fine-grained features from the middle layers (the second and third convolutional layers) in deeper layers (the fourth and fifth convolutional layers), alleviating the gradient vanishing problem, enhancing feature propagation ability, and retaining more spatial detail information.

[0046] Step S3: Train a diffusion model based on the physical sampling process using the training dataset. The training process includes simulating multi-step diffusion of the physical sampling process and calculating a loss function that combines image information and physical information; specifically:

[0047] The multi-step diffusion process based on physical sampling, such as Figure 3 As shown, this constitutes the core mechanism for training the model in this invention. This process aims to iteratively recover complete physical sampling information from a single exposure image using a learnable backbone network. Specifically, the process involves t diffusion iterations: in the initial state, a single X-ray exposure projection image is... The input is fed into the backbone network, which outputs a set of initial phase-stepped images. Subsequently, a Fast Fourier Transform (FFT) is performed on these images to obtain an initial X-ray phase-stepped exposure curve. On this curve, based on physical sampling rules, a virtual phase step is performed on the corresponding X-ray exposure projection image, generating a new single-exposure projection image representing the next sampling position. This process can be represented as .

[0048] The above single-step process is repeated t times, generating sequentially. , , ..., This process involves generating a series of intermediate single projection images and finally outputting a complete set of multi-step projection images predicted by the model. This series of operations simulates the entire process in the real physical world where the grating is moved step by step within one cycle and exposure sampling is performed sequentially. It simulates the information of multiple sampling points in one cycle of the real physical sampling step exposure curve.

[0049] The loss function calculation involves comparing multiple sampling points on the simulated real physical sampling step exposure curve during the multi-step diffusion process with the actual physical sampling step exposure curve, and calculating the loss using the loss function. This ensures that the output of the backbone network is constrained by physical information, making it closer to the actual phase step sampling process. Figure 3 As shown, during the t sampling process, t loss terms can be obtained ( Each loss term includes two parts: image information loss and physical information loss, which can be represented by formula (5), where This represents the loss term in the i-th diffusion sampling process during the t-th sampling diffusion process. and These represent the loss of image information and the loss of physical information during this sampling process, respectively:

[0050] (5)

[0051] The image information loss refers to the loss of pixel values ​​between the multi-step phase-stepped projection image output by the backbone network after each diffusion process and the step projection image at the same phase position on the real phase-stepped sampling curve, which can be expressed by formula (6):

[0052] (6)

[0053] in, These represent the pixel positions in the horizontal and vertical directions of the image, respectively. These represent the dimensions of the image in the horizontal and vertical directions, respectively. and These represent the multi-step phase step projection image output by the backbone network during the i-th diffusion sampling process and the step projection image at the same phase position on the real phase step sampling curve, respectively. This represents the number of phase steps performed, such as... Figure 3 As shown in this embodiment .

[0054] The physical information loss refers to the loss of amplitude, phase, and other physical information obtained after the fast Fourier transform of the multi-step phase-step projection image output by the backbone network after each diffusion process, compared with the amplitude, phase, and other physical information resolved from the real phase-step projection image. It can be expressed by formula (7):

[0055] (7)

[0056] in, These represent the pixel positions in the horizontal and vertical directions of the image, respectively. These represent the dimensions of the image in the horizontal and vertical directions, respectively. These represent the 0th-order amplitude, 1st-order amplitude, and 1st-order phase information of the image obtained by step sampling of the real physical phase after fast Fourier transform. These represent the 0th-order amplitude, 1st-order amplitude, and 1st-order phase information obtained by performing a Fast Fourier Transform on the phase step image obtained in the i-th step of the t-step diffusion process. This design internalizes the discrete physical sampling constraints into the continuous iteration of the model, ensuring the physical rationality and accuracy of the generated results.

[0057] After calculating the loss function, the parameters of the backbone network are updated based on the results using a suitable optimization algorithm. In this embodiment, stochastic gradient descent is used to perform backpropagation.

[0058] Step S4: Input the single X-ray exposure projection image of the object under test into the trained diffusion model to generate multiple corresponding phase step projection images, and analyze the X-ray absorption, phase, and dark field contrast images.

[0059] like Figure 4As shown, a single X-ray exposure projection image of the object being measured at a fixed grating position is input into the trained model. This means the single projection image is input into the trained backbone network. Upon receiving this single input image, the network, leveraging its deep understanding of the physical sampling process learned during training, directly performs forward inference to generate a complete set of multiple projection images covering a full phase step cycle. This set of generated projection images is then combined with the background phase step projection image acquired under conditions without the object being measured (i.e., an empty field). For each pixel in the image, harmonic analysis is performed on the phase step curve formed by the object projection data and the background projection data (usually by extracting the zeroth and first harmonic components using Fast Fourier Transform). Finally, by calculating the ratio of the zeroth harmonic amplitudes of the object and the background, the absorption contrast image is resolved; by calculating the difference in the first harmonic phases of the two, the phase contrast image is resolved; and by calculating the ratio of the first harmonic amplitudes of the two, the dark field contrast image is resolved.

[0060] Based on step S1, such as Figure 5 As shown, a set of multiple phase step projection images obtained using phase step scanning are used as labels for the training dataset, and the first step position of the step scan is taken as a single projection image in the training dataset.

[0061] Further, based on steps S2 and S3, a diffusion model based on the physical sampling process is trained, and a single projection image is input into the trained backbone network to obtain multiple projection images. Subsequently, as... Figure 6 As shown, three contrast images can be resolved with high image quality (using the three contrast images resolved directly using multi-step phase step projection maps as a reference).

[0062] In summary, according to the present invention, compared with traditional X-ray multi-contrast single-exposure imaging methods, most of which require changes to hardware such as grating interferometers and detectors and additional debugging, the method provided by the present invention can be implemented on a common X-ray three-grating interferometer and the results conform to the real physical sampling process, improving the imaging quality of single-exposure resolution of three contrasts and promoting the practical application of low-dose X-ray grating multi-contrast imaging.

[0063] The above detailed embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, only the aforementioned backbone network, loss function implementation, and backpropagation optimization algorithm are used as examples. In practical applications, the above functions can be assigned to different network modules or algorithms as needed to complete all or part of the functions described above. The specific working process of the model described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0064] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned static X-ray multicontrast imaging method based on a physical sampling process diffusion model.

[0065] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned static X-ray multicontrast imaging method based on a diffusion model of a physical sampling process.

[0066] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. 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 static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process, characterized in that, The method includes: Step S1: Obtain the phase-stepped projection map of X-ray grating interferometry imaging and construct the training dataset; Step S2: Construct a diffusion model based on the physical sampling process. The diffusion model includes a backbone network for converting a single X-ray exposure projection image into multiple phase-stepped images. Step S3: Train the diffusion model using the training dataset. The training process includes multi-step diffusion that simulates the physical sampling process and loss function calculation that combines image information and physical information. Step S4: Input the single X-ray exposure projection image of the object under test into the trained diffusion model to generate multiple corresponding phase step projection images, and analyze the X-ray absorption, phase and dark field contrast images.

2. The static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process according to claim 1, characterized in that, In step S2, the backbone network adopts the Unet architecture, with its input being a single-channel single X-ray exposure projection image and its output being a k-channel image, where k is the total number of phase steps.

3. The static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process according to claim 2, characterized in that, The backbone network of the Unet architecture consists of multiple convolutional layers connected sequentially, and residual connections are set between at least two sets of non-adjacent convolutional layers.

4. The static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process according to claim 1, characterized in that, In step S3, the multi-step diffusion process includes: Single-step diffusion step: Input the single X-ray exposure projection image into the backbone network to obtain multiple initial phase-stepped images; perform a fast Fourier transform on the multiple phase-stepped images to obtain an initial phase-stepped exposure curve; perform a virtual phase-stepping operation on the phase-stepped exposure curve according to the physical sampling rules to generate a single X-ray exposure projection image representing the next sampling position; use the newly generated single X-ray exposure projection image as input and re-input it into the backbone network to generate a series of intermediate projection images in sequence; Repeat the single-step diffusion step t times to finally output the complete multi-step projection map predicted by the model.

5. The static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process according to claim 1, characterized in that, In step S3, the loss function is composed of the image information loss and physical information loss in each diffusion process. The image information loss calculates the difference in pixel values ​​between the output image of the diffusion model and the real image, while the physical information loss calculates the difference in physical parameters between the output image of the diffusion model and the real image after Fourier transform.

6. The static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process according to claim 5, characterized in that, The image information loss is the mean square error of the pixel values ​​at corresponding phase positions between the phase step projection map generated by the model and the real phase step projection map.

7. The static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process according to claim 5, characterized in that, The physical information loss is the mean square error between the 0th-order amplitude, 1st-order amplitude, and 1st-order phase obtained by fast Fourier transform of the phase step image generated by the model, and the corresponding physical parameters obtained by fast Fourier transform of the real phase step projection image.

8. The static X-ray multi-contrast imaging method based on a diffusion model of the physical sampling process according to claim 1, characterized in that, Step S3 further includes updating the parameters of the backbone network based on the loss function calculation results after the loss function calculation is completed, using a stochastic gradient descent optimization algorithm.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the static X-ray multicontrast imaging method based on the diffusion model of the physical sampling process as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the static X-ray multi-contrast imaging method based on the diffusion model of the physical sampling process as described in any one of claims 1-8.

Images