An x-ray image denoising method based on physical driving data synthesis and local constraint

Abstract

The application discloses an X-ray image denoising method based on physical driving data synthesis and local constraint, and particularly relates to the field of X-ray imaging denoising technology, and comprises the following steps: collecting an X-ray image through an imaging system; introducing a clean signal Isg, background texture Itx, particle cluster noise Icl and a calibration offset Offset, and superimposing them to the X-ray image to generate a synthetic noise image; performing domain adaptive enhancement and double-path denoising, and mapping the synthetic noise image to obtain an enhanced noise image; through a synthetic noise modeling process based on a physical imaging mechanism, the application uniformly models and superimposes a hot spot radiation form, background irradiation fluctuation and particle cluster pulse noise to generate a synthetic noise image, so that the training data is highly consistent with real X-ray experimental data in terms of statistical distribution and physical characteristics, and the adaptability and stability of the double-path denoising model in a complex scene are improved.

Description

Technical Field

[0001] This invention relates to the field of X-ray imaging noise reduction technology, specifically to an X-ray image noise reduction method based on physical-driven data synthesis and local constraints. Background Technology

[0002] In applications such as inertial confinement fusion (ICF), high-energy physics experiments, industrial non-destructive testing, and high-energy radiation imaging, X-ray imaging technology is widely used to acquire information on target structure, energy distribution, and physical state. However, due to the influence of detector noise, scattered radiation, statistical fluctuations in counting, and complex experimental environments during the imaging process, the actual acquired X-ray images usually contain significant noise interference, resulting in low image signal-to-noise ratios and difficulty in accurately identifying weak signal features and fine structures, thus adversely affecting subsequent quantitative analysis and physical diagnosis. To improve the quality of X-ray images, various image denoising methods have been proposed in existing technologies. Early methods were mainly based on traditional signal processing theories, including mean filtering, median filtering, bilateral filtering, wavelet transform denoising, and denoising algorithms based on variational models.

[0003] The existing technologies have the following main shortcomings: First, they are not good at covering real and complex experimental scenarios. Second, the existing denoising optimization targets lack explicit constraints on the strength of the physical background, which can easily cause background collapse and light distortion during the denoising process.

[0004] To address the aforementioned technical problems, a solution is proposed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an X-ray image denoising method based on physical-driven data synthesis and local constraints, which solves the problems of background collapse and photometric distortion caused during the denoising process.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an X-ray image denoising method based on physics-driven data synthesis and local constraints, comprising: S1. Raw image acquisition: X-ray images are acquired using an X-ray imaging system; S2. Based on the X-ray image, a clean signal Isg, background texture Itx, particle cluster noise Icl, and calibration offset Offset are introduced and superimposed to obtain a synthetic noise image; S3. Domain adaptive enhancement and dual-path denoising: The synthetic noise image is enhanced and mapped to obtain an enhanced noise image; an enhanced synthetic image and a real noisy image are constructed based on the enhanced noise image and the X-ray image, respectively; a dual-path denoising model is constructed to denoise the enhanced synthetic image and the real noisy image, and a clean X-ray image is output. S4. Hybrid Loss Constraint Optimization: Construct a total loss function in the training of the dual-path denoising model and introduce a background compensation constraint term; through multiple iterations and optimizations, the dual-path denoising model is obtained.

[0007] Furthermore, generating synthetic noisy images based on X-ray images includes: X-ray images were acquired in an experimental environment and used as unpaired target domain data storage. The radial boundary contour of the hot spot region in the X-ray image is extracted in polar coordinates. The angular change of the polar coordinate expression is expanded and fitted by a high-order orthogonal polynomial to obtain a set of multi-order deformation parameters. The set of deformation parameters is used to drive the generation of a two-dimensional ideal radiation distribution field, and pixel rasterization mapping is performed to obtain a clean signal Isg; A virtual flux Flx is introduced as a background irradiance term superimposed with hotspot radiation. The clean signal Isg and the virtual flux Flx are superimposed to obtain a joint irradiance field. The background noise intensity scale is calculated based on the pixel intensity distribution of the joint irradiation field, and a corresponding noise sampling field is generated. A convolution imaging operation is performed on the noise sampling field and the point spread function of the imaging system to obtain the background texture Itx. The point spread function is obtained by calibration of the pinhole array geometric parameters and the imaging link response parameters.

[0008] Furthermore, generating synthetic noisy images based on X-ray images also includes: Based on X-ray images, an additive particle cluster accumulation model is constructed. Multiple particle cluster center positions are randomly generated according to a preset particle flux distribution, and the particle track coverage area is simulated based on the particle incident angle and energy distribution. Pixel-level particle deposition components are calculated in the particle track coverage area to form a particle deposition signal; By introducing a photoelectric conversion efficiency constraint for the detector, energy mapping processing is performed on the particle deposition signal, and saturation truncation processing is performed on the pixel charge value that exceeds the full well capacity threshold of the detector to obtain the particle cluster noise Icl; The pixel intensity distribution of the synthesized image is subjected to low quantile statistical analysis, and the gray values ​​corresponding to the preset low quantiles are extracted as the physical background baseline. The physical baseline is aligned with the preset target baseline, and a calibration offset is generated based on the difference between the physical baseline and the preset target baseline.

[0009] Furthermore, the clean signal Isg, background texture Itx, particle cluster noise Icl, system gain Gsy, electronic readout noise Nrd, and calibration offset Offset are superimposed to generate a synthetic noise image lsyn. ; in, For system gain, For electronic readout noise, This represents the quantization process of a simulated ADC.

[0010] Furthermore, the synthesized noisy image lsyn is input into a domain adaptive enhancement network, which performs layer-by-layer enhancement mapping on the noise-related features to reconstruct an enhanced noisy image; The domain adaptive enhancement network is composed of multiple cascaded residual dense blocks, which enhances the low-level noise features of the synthesized noisy image. An enhanced synthetic image is constructed by superimposing the enhanced noise image and the synthetic reference signal at the pixel level, wherein the synthetic reference signal is a clean signal Isg generated based on the physical modeling process; The real noise of the X-ray image and the reconstructed signal are superimposed at the pixel level to construct a real noisy image, wherein the reconstructed signal is obtained by the residual of the real noise after denoising by a denoising sub-network.

[0011] Furthermore, the dual-path denoising module includes a denoising sub-network, a first denoising path, and a second denoising path. The first denoising path is used to denoise the enhanced synthetic image, and the second denoising path is used to denoise the real noisy image. The denoising sub-network adopts a U-Net-like network structure, including an encoder, a latent representation layer, and a decoder. The encoder maps the input tensor from the initial spatial resolution to a low-resolution high-dimensional feature space to extract deep semantic features. The latent representation layer performs centralized representation and cross-scale fusion of the deep semantic features. The decoder fuses the feature information from the encoding stage through skip connections to restore the image spatial resolution and output a clean ray image.

[0012] Furthermore, the background compensation constraint term includes: Based on the pixel intensity histogram characteristics of real noisy images, a background determination threshold is determined to distinguish between background and signal regions. Then, a binary mask for the background region is generated pixel-by-pixel in the real noisy image according to the background determination threshold. ; ; Where xr is the pixel intensity value of the real noisy image, π is the background determination threshold, and otherwise represents all pixels that do not belong to the background.

[0013] Furthermore, in the background region binary mask Within the covered pixel range, a consistency constraint calculation is performed on the pixel intensity difference, and the pixel intensity difference distance between the clean ray image and the real noisy image is calculated as the background preservation loss term Lbg. ; Where xr is the pixel intensity value of the real noisy image; yr is the pixel intensity value of the clean ray image; |yr xr∣ represents the pixel-level intensity difference between the two; This refers to the number of background pixels. As a selective constraint, pixels that are background are included in the loss calculation, while pixels that are signals are not included in the loss calculation. When the background region binary mask When the number of effective pixels is lower than a preset threshold, the calculation of the background preservation loss term is stopped; The background preservation loss term Lbg is introduced into the total loss function; the total loss function is used as the objective function, and the iterations of sample extraction, forward inference, and hybrid loss calculation are repeatedly executed. When the total loss function satisfies the preset convergence condition, the final dual-path denoising model is obtained.

[0014] This invention provides an X-ray image denoising method based on physical-driven data synthesis and local constraints, which has the following advantages compared with the prior art: This invention uses a synthetic noise modeling process based on physical imaging mechanisms to uniformly model and superimpose hotspot radiation patterns, background irradiance fluctuations, and particle cluster pulse noise to generate synthetic noise images. This makes the training data highly consistent with real X-ray experimental data in terms of statistical distribution and physical characteristics, and improves the adaptability and stability of the dual-path denoising model in complex scenarios. This invention introduces a background compensation constraint mechanism during the training of a dual-path denoising model. By adaptively identifying the background region in the real noisy image and constraining the pixel intensity consistency, it suppresses the excessive weakening of the low signal-to-noise ratio background region during the denoising process, thus avoiding background collapse and photometric distortion in the denoising result. This invention employs a dual-path denoising structure combining domain adaptive enhancement and parameter sharing, enabling the denoising network to learn universal feature representations that are insensitive to domain differences. This significantly improves the consistency and robustness of denoising effects on real X-ray images while maintaining detail recovery capabilities. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a backlit X-ray imaging system; Figure 2 A schematic diagram illustrating different types of images in the dataset; Figure 3 A schematic diagram illustrating the comparison between a synthetic image (top) and a real noisy image (bottom); Figure 4 This is a schematic diagram of the structure of the present invention; Figure 5 This is a flowchart of the method proposed in this invention; Figure 6 This is a comparison image showing the denoising effect of the present invention on real high-noise images; Figure 7 This is a comparison image showing the noise reduction effect of the proposed method on real images. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This application provides an X-ray image denoising method based on physics-driven data synthesis and local constraints, including: S1. Raw image acquisition: X-ray images are acquired through an X-ray imaging system.

[0018] like Figure 1 As shown, in an inertial confinement fusion experiment or equivalent experimental environment, real X-ray images are acquired using an X-ray imaging system. These X-ray images are stored as unpaired target domain data for use as real data constraints in subsequent domain adaptive training. The X-ray imaging system includes a pinhole array, filters, an imaging tube, a scintillator, a charge-coupled device (CCD), and a gas chamber. The pinhole array is configured with an 8×7 grid, with a row spacing of 0.5 mm and a column spacing of 0.45 mm, located in front of a pointer 112 mm from the black cavity target, generating 56 pinhole images per experiment. Each pinhole has a diameter of approximately 15 μm. Integrated into the imaging path is a filter assembly consisting of a 200 μm thick beryllium layer and a 50 μm thick aluminum layer, placed inside the imaging tube. Following the tube, a scintillator converts the X-rays into visible light, which is then connected to a charge-coupled device (CCD). The CCD is located 1186 mm from the pinhole array, capturing the generated visible photons.

[0019] S2. Based on the X-ray image, a clean signal Isg, background texture Itx, particle cluster noise Icl, and calibration offset Offset are introduced and superimposed to obtain a synthetic noise image.

[0020] As a preferred embodiment of S2, generating a synthetic noise image based on an X-ray image includes:

[0021] X-ray images were acquired in an experimental environment and used as unpaired target domain data storage.

[0022] The radial boundary contour of the hot spot region in the X-ray image is extracted in polar coordinates. The angular change of the polar coordinate expression is expanded and fitted by a high-order orthogonal polynomial to obtain a set of multi-order deformation parameters.

[0023] The set of deformation parameters is used to drive the generation of a two-dimensional ideal radiation distribution field, and pixel rasterization mapping is performed to obtain a clean signal Isg;

[0024] More specifically, boundary extraction processing is performed on the hotspot regions in the X-ray image. By converting the boundary contours of the hotspot regions into polar coordinates, a polar coordinate expression of the hotspot radius as a function of angle is obtained. Based on this polar coordinate expression, a high-order orthogonal polynomial is used for diagonal expansion and fitting to obtain a set of multi-order deformation parameters characterizing the hotspot asymmetry, center offset, and boundary perturbation features. Driven by this set of multi-order deformation parameters, a two-dimensional ideal radiation distribution field is constructed, and a corresponding synthetic clean signal Isg is generated through pixel rasterization mapping. Figure 2 As shown, the first column, composed of 4th Legendre images, models the radial boundaries of hotspots to obtain a synthetic clean signal, Isg, which characterizes the ideal hotspot radiation distribution free from noise. The first column uses a 4th Legendre function to represent the hotspot asymmetry, and the fourth column uses a 10th Legendre function. The signal reconstruction method is as described above. To prevent overfitting, we added natural texture images, as shown in the second and third columns. The second row shows the result of adding noise to the first row of images.

[0025] A virtual flux Flx is introduced as a background irradiance term superimposed with hotspot radiation. The clean signal Isg and the virtual flux Flx are superimposed to obtain a joint irradiance field.

[0026] Furthermore, a virtual flux Flx is introduced as a background irradiance term superimposed with hotspot radiation. The clean signal Isg and the virtual flux Flx are superimposed according to pixel intensity to obtain a joint irradiance field. Based on the statistical distribution of pixel intensity of the joint irradiance field, the background noise intensity scale is calculated, and a corresponding random noise sampling field is generated accordingly. Then, the noise sampling field is convolved with the point spread function of the imaging system to obtain a background texture component Itx with spatial correlation characteristics. The point spread function is obtained by calibration of the geometric parameters of the pinhole array and the system response parameters of the imaging link.

[0027] The background noise intensity scale is calculated based on the pixel intensity distribution of the joint irradiation field, and a corresponding noise sampling field is generated. A convolution imaging operation is performed on the noise sampling field and the point spread function of the imaging system to obtain the background texture Itx. The point spread function is obtained by calibration of the pinhole array geometric parameters and the imaging link response parameters.

[0028] As another preferred embodiment of S2, generating a synthetic noise image based on an X-ray image further includes:

[0029] Based on X-ray images, an additive particle cluster accumulation model is constructed. Multiple particle cluster center positions are randomly generated according to a preset particle flux distribution, and the particle track coverage area is simulated based on the particle incident angle and energy distribution.

[0030] Pixel-level particle deposition components are calculated in the particle track coverage area to form a particle deposition signal;

[0031] By introducing a photoelectric conversion efficiency constraint for the detector, energy mapping processing is performed on the particle deposition signal, and saturation truncation processing is performed on the pixel charge value that exceeds the full well capacity threshold of the detector to obtain the particle cluster noise Icl;

[0032] Specifically, to address the pulse-type noise generated by the interaction between high-energy particles and the detector, an additive particle cluster accumulation model is constructed based on the X-ray image. Multiple particle cluster center positions are randomly generated according to a preset particle flux distribution, and the coverage area of ​​particle tracks on the imaging plane is simulated by combining the particle incident angle distribution and energy distribution. Pixel-level particle deposition components are calculated within the particle track coverage area to form a particle deposition signal. At the same time, the photoelectric conversion efficiency constraint of the detector is introduced, energy mapping processing is performed on the particle deposition signal, and pixel charge values ​​exceeding the detector's full-well capacity threshold are saturated and truncated to obtain the particle cluster noise Icl.

[0033] The pixel intensity distribution of the synthesized image is subjected to low quantile statistical analysis, and the gray values ​​corresponding to the preset low quantiles are extracted as the physical background baseline.

[0034] The physical baseline is aligned with the preset target baseline, and a calibration offset is generated based on the difference between the physical baseline and the preset target baseline.

[0035] Specifically, a low quantile statistical analysis is performed on the overall pixel intensity distribution of the synthesized image, and the gray value corresponding to the preset low quantile is extracted as the physical baseline of the current synthesized image. The physical baseline is aligned with the preset target baseline, and a calibration offset is generated based on the difference between the two to compensate for the baseline drift under low signal-to-noise ratio conditions.

[0036] This invention utilizes a synthetic noise modeling process based on physical imaging mechanisms to uniformly model and superimpose hotspot radiation patterns, background irradiance fluctuations, and particle cluster pulse noise to generate synthetic noise images. This ensures that the training data closely matches real X-ray experimental data in terms of statistical distribution and physical characteristics, thereby improving the adaptability and stability of the dual-path denoising model in complex scenarios.

[0037] S3. Domain adaptive enhancement and dual-path denoising: The synthesized noisy image is enhanced and mapped to obtain an enhanced noisy image; an enhanced synthetic image and a real noisy image are constructed based on the enhanced noisy image and the X-ray image, respectively; a dual-path denoising model is constructed to denoise and fuse the enhanced synthetic image and the real noisy image to output a clean X-ray image, as shown in the appendix. Figure 4 As shown.

[0038] As a preferred embodiment of S3, the clean signal Isg, background texture Itx, particle cluster noise Icl, system gain Gsy, electronic readout noise Nrd, and calibration offset Offset are superimposed to generate a synthetic noise image lsyn: ; in, For system gain, For electronic readout noise, The term "Quantize" refers to the quantization process of an analog-to-digital converter (ADC), specifically a rounding operation (either up or down). In the physical process, the number of photoelectrons collected by the detector must be discrete integers, and the output image grayscale value (ADU) after system gain amplification must also be an integer. The system gain of the camera is used to map the physical charge (number of electrons) collected by the detector to the digital grayscale value (ADU) of the image. In a specific embodiment of the invention, to match the imaging device characteristics of a real inertial confinement fusion (ICF) experiment, The value is set to 2.1ADU / e-; This is used to simulate the substrate thermal noise and circuit fluctuations generated by the detector during electronic readout. It is determined by constructing a model with a mean of 0 and a standard deviation of... The Gaussian random distribution. To simulate the natural differences in background variability across different batches of experiments, the standard deviation... During the synthesis process, it is set to be dynamically randomized, and its value ranges from 600 to 700.

[0039] Specifically, the clean signal Isg, background texture Itx, particle cluster noise Icl, system gain Gsy, electronic readout noise Nrd, and calibration offset Offset are superimposed according to physical consistency, and the analog-to-digital conversion process is simulated through quantization to generate the final synthetic noise image lsyn.

[0040] Specifically, such as Figure 4As shown, the synthesized noise image lsyn is input into the domain adaptive enhancement network EnhancementModule. The domain adaptive enhancement network performs layer-by-layer enhancement mapping on the noise-related features and reconstructs the enhanced noise image Enhancednoise(Xnt).

[0041] The domain adaptive enhancement network EnhancementModule is composed of multiple residual dense blocks fused by residual skip cascade fusion, aiming to align low-level features. Each dense block contains a convolutional layer Conv, an instance normalization layer (InstanceNorm), and an activation function to enhance the low-level noise features of the synthetic noise image. The residual connections are designed to preserve high-level structural information while transforming low-level noise features.

[0042] The enhanced noise image (Xnt) and the synthetic reference signal (Syntheticsignal) are superimposed at the pixel level to construct an enhanced synthetic image. The synthetic reference signal is a clean signal Isg generated based on the physical modeling process.

[0043] The real noise (Xnr) of the X-ray image is superimposed with the reconstructed signal at the pixel level to construct a real noisy image, wherein the reconstructed signal is obtained by residualing the real noise after denoising by a denoising sub-network.

[0044] Specifically, the reconstructed signal is obtained by interpolating the signal-containing noisy image with the residual obtained by denoising using the Unet model. We performed rigorous quality checks on the noise generated by the residual (original denoising). If the residual patch showed a high correlation with the signal structure (indicating leakage), it was replaced with a background noise patch, which was cropped from the signal-free background region of the experimental image.

[0045] like Figure 4 As shown, the dual-path denoising module consists of two weight-shared U-Net-like networks, including a denoising sub-network, a first denoising path, and a second denoising path. The first denoising path is used to denoise the enhanced synthetic image to obtain the output Output1(yst), and the second denoising path is used to denoise the real noisy image to obtain the output Output2(yt).

[0046] The denoising sub-network adopts a U-Net-like network structure, including an encoder, a latent representation layer, and a decoder. The encoder maps the input tensor from the initial spatial resolution to a low-resolution high-dimensional feature space to extract deep semantic features. The latent representation layer performs centralized representation and cross-scale fusion of the deep semantic features. The decoder fuses the feature information from the encoding stage through skip connections to restore the image spatial resolution and output a clean ray image.

[0047] S4. Hybrid Loss Constraint Optimization: Construct a total loss function in the training of the dual-path denoising model and introduce a background compensation constraint term; through multiple iterations and optimizations, the dual-path denoising model is obtained.

[0048] As another preferred embodiment of S4, the background compensation constraint term includes:

[0049] Based on the pixel intensity histogram characteristics of real noisy images, a background determination threshold is determined to distinguish between background and signal regions. Then, a binary mask for the background region is generated pixel-by-pixel in the real noisy image according to the background determination threshold. ; ; in, For a real noisy image, π is the background threshold; otherwise, all pixels that do not belong to the background are included.

[0050] Specifically, in the background region binary mask Within the covered pixel range, a consistency constraint calculation is performed on the pixel intensity difference, and the pixel intensity difference distance between the clean ray image and the real noisy image is calculated as the background preservation loss term Lbg. ; in, yr represents the pixel intensity value of the real noisy image; yr represents the pixel intensity value of the clean ray image; |yr xr∣ represents the pixel-level intensity difference between the two; This represents the total number of pixels within the mask. As a selective constraint, pixels that are background are included in the loss calculation, while pixels that are signals are not included in the loss calculation.

[0051] When the background region binary mask When the number of effective pixels is lower than a preset threshold, the calculation of the background preservation loss term is stopped;

[0052] In this embodiment, the task constraints include: calculating the absolute error between the denoised output and the synthesized clean signal pixel by pixel in the pixel space to form a pixel-level reconstruction loss; extracting multi-layer perceptual features from the denoised output and the synthesized clean signal respectively through a pre-trained feature extraction network in the structure space, and measuring the differences between the corresponding layer features to form a perceptual consistency loss; and weighting and fusing the pixel-level reconstruction loss and the perceptual consistency loss according to preset weights to form the task constraint Ltask for denoising and reconstruction accuracy.

[0053] To reduce the noise distribution difference between the source and target domains, a discriminative enhancement network is introduced to enhance the synthesized image and the real noisy image, forming a noise domain alignment adversarial loss, LadNoise. A feature-level domain adversarial discriminator is introduced into the latent feature space of the denoising network to discriminate the domain source of latent feature representations from the synthesized and real domain paths. Through adversarial training, the denoising network gradually weakens the domain-related features in the latent space, forming a latent feature adversarial loss, Ladlatent. An output-level domain adversarial discriminator is introduced at the output of the denoising network. The denoised outputs from the synthesized and real domains are input into the discriminator. The domain discrimination loss is calculated based on the differences in global brightness distribution, texture statistical features, and local structural patterns of the output results. Through an adversarial game mechanism, the denoising network is guided to reduce the domain discriminability of its output results, forming an output result adversarial loss, Ladout.

[0054] Specifically, the total loss function Defined as: ; Where λt, λc, λl, λo, and λb are weight coefficients; Ltask is the task constraint term; LadNoise is the adversarial loss for noise domain alignment; Ladlatent is the adversarial loss for latent features; Ladout is the adversarial loss for output results; and Lbg is the baseline preservation loss term.

[0055] This invention introduces a background compensation constraint mechanism during the training process of the dual-path denoising model. By adaptively identifying the background region in the real noisy image and constraining the pixel intensity consistency, it suppresses the excessive weakening of the low signal-to-noise ratio background region during the denoising process, thus avoiding background collapse and photometric distortion in the denoising result.

[0056] The background preservation loss term Lbg is introduced into the total loss function; the total loss function is used as the objective function, and the iterations of sample extraction, forward inference, and hybrid loss calculation are repeatedly executed. When the total loss function satisfies the preset convergence condition, the final dual-path denoising model is obtained.

[0057] In this embodiment, before training, all images were normalized to a single-channel 16-bit format, adjusted to a resolution of 512×512 pixels, and normalized to the range [0,1]. Specifically, this normalization preserved the relative intensity distribution of the real data, providing a unified metric space for subsequent threshold-based background recognition. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The model was implemented using the PyTorch framework and trained on an NVIDIA GeForce RTX 3090 GPU with a batch size of 4.

[0058] like Figure 3 As shown, Figure 3 The four images above are augmented composite images. Figure 3 The four images below are real noisy images. During the model training phase, for each batch of input samples, the enhanced synthetic image and the real noisy image are input into a dual-path denoising model with shared encoder parameters to obtain the corresponding denoising output. Based on the denoising output, a total loss function is constructed, which includes task constraints, domain adversarial constraints, and background compensation constraints.

[0059] Specifically, the differences between the denoised output of the synthetic domain path and the corresponding synthesized clean signal in pixel space and structural space are jointly measured by task constraints to constrain the basic denoising and reconstruction capabilities of the model. Specifically, the pixel-wise absolute error between the denoised output and the synthesized clean signal is calculated to force the model to learn the global intensity distribution of the signal. High-level features of the denoised output and the clean signal are extracted through a pre-trained VGG network, and the feature differences are calculated to guide the model to recover the structural details and semantic information of the image.

[0060] Specifically, domain discrimination mechanisms are introduced into the feature and output spaces of potential adaptive augmentation networks and weight-sharing U-Net networks through domain adversarial constraints. A discriminator is then used to train the feature representations and denoising results of the synthetic and real domains to reduce the distribution differences between different data domains. For example, Figure 4 As shown, a discriminator is introduced into the adaptive enhancement network to distinguish between the enhanced synthetic image and the real noisy image, reducing the difference in noise distribution between the source and target domains; in the feature space of the weight-sharing U-Net network, the differences in style attributes (such as illumination distribution) between the synthetic and real domain features are adjusted, while semantic content is preserved; the denoised synthetic image and the denoised real image are distinguished to ensure that the output result has visual authenticity.

[0061] More specifically, the pixel intensity consistency of the background region in the real domain path is constrained by a background compensation constraint term, and this is applied to the mask. Within the covered area, the output of the forced denoising network With input To maintain consistency, the L1 distance between the two is calculated as the loss to prevent excessive suppression of the physical background during the denoising process.

[0062] like Figure 6 and Figure 7 As shown, Figure 6 The top two rows show the denoising effect of Unet on two images. In these two rows, the first column on the left is the real noisy image, the second column is the image after denoising by Unet, and the last four columns are grayscale value comparison images from different angles. The bottom two rows show the denoising effect of the same two images. Figure 7 The change in the signal-to-noise ratio (SNR) after denoising a real noisy image.

[0063] In contrast, this invention, by introducing an improved physical synthesis mechanism and background compensation loss, retains as many physical photometric features as possible at extremely low signal-to-noise ratios.

[0064] This invention employs a dual-path denoising structure combining domain-adaptive enhancement and parameter sharing. This enables the denoising network to learn universal feature representations insensitive to domain differences, significantly improving the consistency and robustness of denoising results for real X-ray images while maintaining detail recovery capabilities. The network structure achieves triple alignment at the image, feature, and result levels through enhancement, denoising, and joint adversarial learning performed by three discriminators. In particular, the introduction of a background preservation loss, as a regularization term, effectively prevents the model from misclassifying the physical background of the real image as pure noise for removal. The final denoising output correctly preserves the background grayscale level consistent with the experimental environment while removing impulse and shot noise, successfully bridging the photometric domain gap between synthetic and real data.

[0065] In each training iteration, backpropagation calculation is performed on the dual-path denoising model and the corresponding adversarial discrimination module based on the total loss function to update the network parameters; and sample extraction, forward inference, hybrid loss calculation and parameter update operations are alternately performed in multiple iterations. When the change of the total loss function meets the preset convergence condition or reaches the maximum number of iterations, the training process ends and the final dual-path denoising model is obtained.

[0066] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0067] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0068] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for denoising X-ray images based on physics-driven data synthesis and local constraints, characterized in that, include: Raw image acquisition: X-ray images are acquired using an X-ray imaging system; Based on the X-ray image, a clean signal Isg, background texture Itx, particle cluster noise Icl, and calibration offset Offset are introduced and superimposed to obtain a synthetic noise image; Domain adaptive enhancement and dual-path denoising are used to enhance and map the synthesized noisy image to obtain the enhanced noisy image. An enhanced synthetic image and a real noisy image are constructed based on the enhanced noisy image and the X-ray image, respectively; a dual-path denoising model is constructed to denoise the enhanced synthetic image and the real noisy image, and a clean X-ray image is output. Hybrid loss constraint optimization involves constructing a total loss function during the training of the dual-path denoising model and introducing a background compensation constraint term. Through multiple iterations and optimizations, the dual-path denoising model is obtained.

2. The X-ray image denoising method based on physical-driven data synthesis and local constraints according to claim 1, characterized in that, Generate synthetic noise images based on X-ray images, including: X-ray images were acquired in an experimental environment and used as unpaired target domain data storage. The radial boundary contour of the hot spot region in the X-ray image is extracted in polar coordinates. The angular change of the polar coordinate expression is expanded and fitted by a high-order orthogonal polynomial to obtain a set of multi-order deformation parameters. The set of deformation parameters is used to drive the generation of a two-dimensional ideal radiation distribution field, and pixel rasterization mapping is performed to obtain a clean signal Isg; A virtual flux Flx is introduced as a background irradiance term superimposed with hotspot radiation. The clean signal Isg and the virtual flux Flx are superimposed to obtain a joint irradiance field. The background noise intensity scale is calculated based on the pixel intensity distribution of the joint irradiation field, and a corresponding noise sampling field is generated. A convolution imaging operation is performed on the noise sampling field and the point spread function of the imaging system to obtain the background texture Itx. The point spread function is obtained by calibration of the pinhole array geometric parameters and the imaging link response parameters.

3. The X-ray image denoising method based on physical-driven data synthesis and local constraints according to claim 2, characterized in that, Generating synthetic noise images based on X-ray images also includes: Based on X-ray images, an additive particle cluster accumulation model is constructed. Multiple particle cluster center positions are randomly generated according to a preset particle flux distribution, and the particle track coverage area is simulated based on the particle incident angle and energy distribution. Pixel-level particle deposition components are calculated in the particle track coverage area to form a particle deposition signal; By introducing a photoelectric conversion efficiency constraint for the detector, energy mapping processing is performed on the particle deposition signal, and saturation truncation processing is performed on the pixel charge value that exceeds the full well capacity threshold of the detector to obtain the particle cluster noise Icl; The pixel intensity distribution of the synthesized image is subjected to low quantile statistical analysis, and the gray values ​​corresponding to the preset low quantiles are extracted as the physical background baseline. The physical baseline is aligned with the preset target baseline, and a calibration offset is generated based on the difference between the physical baseline and the preset target baseline.

4. The X-ray image denoising method based on physical-driven data synthesis and local constraints according to claim 3, characterized in that, The clean signal Isg, background texture Itx, particle cluster noise Icl, system gain Gsy, electronic readout noise Nrd, and calibration offset are superimposed to generate a synthetic noise image lsyn: ； in, For system gain, For electronic readout noise, This represents the quantization process of a simulated ADC.

5. The X-ray image denoising method based on physical-driven data synthesis and local constraints according to claim 4, characterized in that, The synthesized noisy image lsyn is input into a domain adaptive enhancement network, which performs layer-by-layer enhancement mapping on the noise-related features to reconstruct an enhanced noisy image. The domain adaptive enhancement network is composed of multiple cascaded residual dense blocks, which enhances the low-level noise features of the synthesized noisy image. An enhanced synthetic image is constructed by superimposing the enhanced noise image and the synthetic reference signal at the pixel level, wherein the synthetic reference signal is a clean signal Isg generated based on the physical modeling process; The real noise of the X-ray image and the reconstructed signal are superimposed at the pixel level to construct a real noisy image, wherein the reconstructed signal is obtained by taking the residual of the real noise after denoising the result of the denoising sub-network.

6. The X-ray image denoising method based on physical-driven data synthesis and local constraints according to claim 1, characterized in that, The dual-path denoising module includes a denoising sub-network, a first denoising path, and a second denoising path. The first denoising path is used to denoise the enhanced synthetic image, and the second denoising path is used to denoise the real noisy image. The denoising sub-network adopts a U-Net-like network structure, including an encoder, a latent representation layer, and a decoder. The encoder maps the input tensor from the initial spatial resolution to a low-resolution high-dimensional feature space to extract deep semantic features. The latent representation layer performs centralized representation and cross-scale fusion of the deep semantic features. The decoder fuses the feature information from the encoding stage through skip connections to restore the image spatial resolution and output a clean ray image.

7. The X-ray image denoising method based on physical-driven data synthesis and local constraints according to claim 1, characterized in that, The background compensation constraint includes: Based on the pixel intensity histogram characteristics of real noisy images, a background determination threshold is determined to distinguish between background and signal regions. Then, a binary mask for the background region is generated pixel-by-pixel in the real noisy image according to the background determination threshold. ; ； Where xr is the pixel intensity value of the real noisy image, π is the background determination threshold, and otherwise represents all pixels that do not belong to the background.

8. The X-ray image denoising method based on physical-driven data synthesis and local constraints according to claim 1, characterized in that, Binary mask in the background region Within the covered pixel range, a consistency constraint calculation is performed on the pixel intensity difference, and the pixel intensity difference distance between the clean ray image and the real noisy image is calculated as the background preservation loss term Lbg. ； Where xr is the pixel intensity value of the real noisy image; yr is the pixel intensity value of the clean ray image; |yr xr∣ represents the pixel-level intensity difference between the two; This refers to the number of background pixels. As a selective constraint, pixels that are background are included in the loss calculation, while pixels that are signals are not included in the loss calculation. When the background region binary mask When the number of effective pixels is lower than a preset threshold, the calculation of the background preservation loss term is stopped; The background preservation loss term Lbg is introduced into the total loss function; the total loss function is used as the objective function, and the iterations of sample extraction, forward inference, and hybrid loss calculation are repeatedly executed. When the total loss function satisfies the preset convergence condition, the final dual-path denoising model is obtained.

Images