A method, system, and device for joint denoising of dual-source data

By employing a dual-source data joint denoising method for DS-CBCT equipment, and utilizing a self-supervised denoising network model, combined with the correlation and redundancy information of the dual-source data, the noise problem in DS-CBCT images was solved, achieving high-quality image reconstruction and improving image clarity and diagnostic accuracy.

CN121258822BActive Publication Date: 2026-11-10JIANGSU FIRST-IMAGING MEDICAL EQUIPMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511334209.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-09-18
Publication Date
2026-11-10
Estimated Expiration
2045-09-18

AI Technical Summary

Technical Problem

Significant quantum noise, electronic noise, and noise introduced by reconstruction algorithms exist in DS-CBCT images, affecting image clarity and clinical diagnostic accuracy. Existing methods struggle to effectively denoise images without clean image supervision.

Method used

By utilizing the dual-source structure design of the DS-CBCT equipment, a self-supervised noise correction is performed by combining the correlation and redundancy information of the dual-source data with a self-supervised denoising network model. This includes data stitching, noise optimization, and training of the self-supervised denoising network to construct a noise mapping relationship to remove image noise.

Benefits of technology

Without external noise-free image supervision, it effectively suppresses image noise, improves image signal-to-noise ratio and low-contrast resolution, simplifies data preprocessing, is suitable for imaging scenarios with various noise levels, and improves image quality and diagnostic reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121258822B_ABST
    Figure CN121258822B_ABST
Patent Text Reader

Abstract

This application discloses a method, system, and device for joint denoising of dual-source data. The method includes: acquiring dual-source projection data based on a dual-source structure for preliminary reconstruction; stitching and reconstructing the dual-source image data obtained from the preliminary reconstruction, extracting inherent noise, and performing noise optimization processing on the dual-source image data based on the inherent noise; establishing a self-supervised denoising network model, using the noise-optimized dual-source image data as input to the self-supervised denoising network model; constraining the noise distribution characteristics based on the dual-source image data to construct a noise mapping relationship; and performing noise correction on the dual-source image data based on the trained self-supervised denoising network model to obtain the target reconstructed image. This application utilizes the correlation and redundancy information between the two sets of data in the dual-source structure of the DS-CBCT equipment to design a self-supervised noise correction algorithm, effectively suppressing various types of noise in the reconstructed image, thereby obtaining high-quality image reconstruction results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and more specifically, to a method, system, and device for joint denoising of dual-source data. Background Technology

[0002] Dual-Source Cone Beam CT (DS-CBCT) is a cone-beam CT imaging technique based on dual X-ray sources. Due to its dual-source structure, DS-CBCT equipment has inherent complementary advantages during data acquisition, enabling large-field-of-view imaging. However, it still inevitably faces noise issues. During actual scanning, DS-CBCT images are often significantly noisy due to multiple factors, including quantum noise, electronic noise, and the accuracy of reconstruction algorithms. This noise severely affects image clarity and detail, thus reducing the accuracy of clinical diagnosis. Current mainstream noise correction methods largely rely on high-quality, noise-free images as supervisory information, but obtaining ideal noise-free images is difficult in practical applications. Summary of the Invention

[0003] To address the shortcomings of DS-CBCT technology in image noise processing, this application discloses a dual-source data joint denoising method, system, and device. Utilizing the correlation and redundancy information between the two sets of data in the dual-source structure of the DS-CBCT device, a self-supervised noise correction algorithm is designed to effectively suppress various types of noise in DS-CBCT images, thereby obtaining high-quality image reconstruction results. Specifically, the technical solution of this application is as follows:

[0004] In a first aspect, this application discloses a method for joint denoising of dual-source data, comprising the following steps:

[0005] Based on the dual-source structure, dual-source projection data is acquired, and a reconstruction algorithm is used to perform preliminary reconstruction of the dual-source projection data respectively;

[0006] The dual-source image data obtained from the initial reconstruction, including the upper source image and the lower source image, is stitched together and reconstructed a second time to obtain a secondary reconstructed image; inherent noise is extracted from the secondary reconstructed image, and noise optimization processing is performed on the dual-source image data based on the inherent noise.

[0007] A self-supervised denoising network model is established, and the optimized dual-source image data is used as the input of the self-supervised denoising network model; the noise distribution characteristics are constrained based on the dual-source image data, and a noise mapping relationship is constructed.

[0008] Based on the trained self-supervised denoising network model, noise correction is performed on the optimized dual-source image data to obtain the target reconstructed image.

[0009] In some implementations, the initial reconstructed dual-source image data includes an upper source image and a lower source image, which are then stitched together and reconstructed a second time to obtain a second-reconstructed image; specifically, this includes:

[0010] The dual-source image data is stitched together to obtain stitched image data: the dual-source image data is aligned to the same coordinate system using a geometric registration method; and image pixels are fused in the common overlapping areas of the dual-source image data based on a data interpolation algorithm.

[0011] The stitched image data is segmented using a threshold segmentation algorithm: the skeletal region and the soft tissue region are segmented by setting an appropriate threshold; and the pixel values ​​of the segmented soft tissue region are replaced with preset constant values.

[0012] The thresholded soft tissue region data is merged with the bone region data to obtain merged image data;

[0013] The merged image data is back-projected, and then the reconstruction algorithm is used to reconstruct the back-projected data a second time to obtain the reconstructed image.

[0014] In some implementations, the step of extracting inherent noise from the secondary reconstructed image and performing noise optimization processing on the dual-source image data based on the inherent noise specifically includes:

[0015] For common regions that overlap spatially in the reconstructed images, the data difference between the merged image data and the reconstructed image is calculated to obtain the inherent noise introduced by the reconstruction algorithm;

[0016] Based on the dual-source image data, the inherent noise is removed to obtain the optimized dual-source image data.

[0017] In some implementations, the self-supervised denoising network model includes:

[0018] The feature extraction module is used to extract the features of noise signals and real signals in the input image using a multi-layer convolutional neural network, obtain noise distribution features, and fuse information at multiple scales;

[0019] Noise mapping learning module: used to construct a noise mapping matrix based on the overlap information of the input image in the common region, so as to learn the correspondence between noise distribution and real signal in a self-supervised manner;

[0020] Denoising and reconstruction module: used to remove noise components from the input image based on the obtained noise mapping prediction results, and output the denoised target reconstructed image.

[0021] In some implementations, during the model training process, the self-supervised denoising network model utilizes the optimized dual-source image data, including the optimized upper source image and the optimized lower source image, as mutual supervision signals.

[0022] When the optimized source image is used as network input, the corresponding optimized source image is used as the label; conversely, when the optimized source image is used as network input, the corresponding optimized source image is used as the label.

[0023] In some implementations, the reconstruction algorithm is the FDK reconstruction algorithm.

[0024] Secondly, this application also discloses a dual-source data joint denoising system, comprising:

[0025] Data acquisition module; used to acquire dual-source projection data based on a dual-source structure, including upper source projection data and lower source projection data;

[0026] The noise optimization module is used to perform preliminary reconstruction of the dual-source projection data using a reconstruction algorithm; it is also used to stitch together and perform secondary reconstruction of the dual-source image data obtained from the preliminary reconstruction, including the upper source image and the lower source image, to obtain a secondary reconstructed image; it extracts inherent noise from the secondary reconstructed image and performs noise optimization processing on the dual-source image data based on the inherent noise;

[0027] A self-supervised network denoising module is used to establish a self-supervised denoising network model, taking optimized dual-source image data as input to the self-supervised denoising network model; constraining the noise distribution characteristics based on the dual-source image data, and constructing a noise mapping relationship;

[0028] The self-supervised network denoising module is also used to perform noise correction on the optimized dual-source image data based on the trained self-supervised denoising network model to obtain the target reconstructed image.

[0029] In some embodiments, the noise optimization module includes:

[0030] The stitching submodule is used to stitch the dual-source image data to obtain stitched image data: aligning the dual-source image data to the same coordinate system using a geometric registration method; and performing image pixel fusion on the common overlapping areas of the dual-source image data based on a data interpolation algorithm.

[0031] The segmentation submodule is used to segment the stitched image data using a threshold segmentation algorithm: by setting an appropriate threshold, the skeletal region and the soft tissue region are segmented; and the pixel values ​​of the segmented soft tissue region are replaced with preset constant values.

[0032] The merging submodule is used to merge the thresholded soft tissue region data with the bone region data to obtain merged image data.

[0033] The back-projection submodule is used to perform back-projection processing on the merged image data, and then use the reconstruction algorithm to perform secondary reconstruction on the back-projected data to obtain the secondary reconstructed image.

[0034] In some embodiments, the noise optimization module further includes:

[0035] The noise calculation submodule is used to calculate the data difference between the merged image data and the secondary reconstructed image for common regions that overlap in space in the secondary reconstructed image, to obtain the inherent noise introduced by the reconstruction algorithm; it is also used to remove the inherent noise based on the dual-source image data to obtain the optimized dual-source image data.

[0036] Thirdly, this application also discloses a DS-CBCT device, which includes a dual-source data joint denoising system as described in any of the above embodiments.

[0037] Compared with the prior art, this application has at least one of the following beneficial effects:

[0038] 1. This application utilizes the correlation and redundancy information between two sets of data in the dual-source structure of the DS-CBCT equipment to design a self-supervised noise correction algorithm. By utilizing the inherent redundancy information between the dual-source data, self-supervised noise modeling is achieved under the condition of no clean image supervision, which fundamentally suppresses quantum noise, electronic noise and noise caused by algorithm insufficiency.

[0039] 2. This application does not rely on external noise-free images as supervision labels, and is applicable to DS-CBCT imaging scenarios with various noise levels. It has good adaptability and promotional application value. This application simplifies the data preprocessing process, reduces noise complexity through a secondary reconstruction process, provides stable input for the self-supervised network, and improves the overall denoising effect.

[0040] 3. This application provides an adaptive noise correction mechanism for statistical analysis of noise within the common region. By separating the inherent noise components introduced by the reconstruction algorithm, targeted noise suppression is achieved. Compared to global filtering denoising, it better preserves the edge and detail features of anatomical structures. The entire process fully utilizes the data redundancy characteristics of the dual-source system, significantly improving the signal-to-noise ratio and low-contrast resolution of the image while reducing noise complexity, resulting in high-quality image output. Attached Figure Description

[0041] The preferred embodiments will now be described in a clear and easy-to-understand manner, in conjunction with the accompanying drawings, to further explain the above-mentioned characteristics, technical features, advantages, and implementation methods of this application.

[0042] Figure 1 This is a flowchart illustrating the steps of an embodiment of a dual-source data joint denoising method according to this application;

[0043] Figure 2 This is a schematic diagram illustrating the process of stitching and reconstructing dual-source image data in an embodiment of this application.

[0044] Figure 3 This is a schematic diagram of the noise optimization process for dual-source image data in an embodiment of this application;

[0045] Figure 4 This is a schematic diagram illustrating the application process of the self-supervised denoising network model in the embodiments of this application. Detailed Implementation

[0046] 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 can 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.

[0047] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or sets.

[0048] To keep the drawings concise, each figure only schematically shows the parts relevant to the invention, and these do not represent the actual structure of the product. Furthermore, to facilitate understanding, in some figures, only one of components with the same structure or function is schematically depicted, or only one is labeled. In this document, "one" not only means "only one," but can also mean "more than one."

[0049] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0050] Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the specific implementation methods of this application will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without creative effort.

[0052] Dual-source cone-beam computed tomography (DS-CBCT) is now widely used in medical imaging, especially in scenarios requiring high-resolution imaging, such as oral, head and neck, and orthopedic procedures. Compared to traditional single-tube CBCT, DS-CBCT uses two independent X-ray sources and detector systems, typically arranged orthogonally or at a fixed angle, enabling simultaneous acquisition of projection data from different angles. This dual-source design significantly improves imaging efficiency and quality. The advantages of this dual-source structure include expanded imaging field of view, reduced scan time, and improved temporal resolution, making it particularly suitable for dynamic or large-area scans.

[0053] While the dual-source design of DS-CBCT improves data acquisition efficiency, it also introduces some drawbacks. The increased complexity of the dual-source system, coupled with the potential for increased scattering noise and radiation dose during actual scanning due to the simultaneous operation of both X-ray tubes, leads to issues. DS-CBCT image quality is affected by various noise factors, including quantum noise, electronic noise, scattering noise, and approximation errors in the reconstruction algorithm under limited angle or low-dose conditions. These noises can reduce image contrast and resolution, and obscure details of minute lesions, thus impacting the reliability of clinical diagnosis.

[0054] Currently, denoising methods for DS-CBCT can be mainly divided into the following categories: The first category is hardware-optimized methods, including using anti-scattering grids to reduce scattering noise or using high-sensitivity detectors to improve the signal-to-noise ratio. These methods are costly and cannot completely eliminate noise. The second category is based on traditional image processing methods, such as non-local means, wavelet transform, or anisotropic diffusion filtering, which directly smooth noise on the reconstructed image. However, these methods are prone to over-smoothing, resulting in the loss of minute structures. The third category is based on deep learning methods: using convolutional neural networks (CNNs) or generative adversarial networks (GANs) to recover high-quality images from low-quality images. Mainstream approaches include supervised learning: constructing end-to-end neural network models using pairs of "noisy-low-quality images" and "clean-high-quality images" as training data to learn the mapping relationship from noisy images to ideal images. However, in clinical practice, it is difficult to obtain noise-free reference images as ideal noise-free labels. Therefore, all of the above methods have certain limitations.

[0055] In view of the above, this application makes full use of the correlation and redundancy information between the two sets of data in the dual-source structure of the DS-CBCT device. In the absence of external noise-free image labels, a self-supervised noise correction algorithm is designed to effectively suppress quantum noise, electronic noise and various noises introduced by the reconstruction algorithm in DS-CBCT images, thereby obtaining high-quality image reconstruction results and improving imaging effect and diagnostic accuracy.

[0056] Reference manual attached Figure 1 As shown, Figure 1 This is a flowchart illustrating the steps of an embodiment of a dual-source data joint denoising method according to this application. An embodiment of the dual-source data joint denoising method specifically includes the following steps: S100, acquiring dual-source projection data based on a dual-source structure, including upper source projection data and lower source projection data. Preliminary reconstruction is performed on the dual-source projection data using a reconstruction algorithm. S200, stitching and secondary reconstruction are performed on the preliminarily reconstructed dual-source image data, including the upper source image and the lower source image, to obtain a secondary reconstructed image. Inherent noise is extracted from the secondary reconstructed image, and noise optimization processing is performed on the dual-source image data based on the inherent noise. S300, establishing a self-supervised denoising network model, using the optimized dual-source image data as input to the self-supervised denoising network model. The noise distribution characteristics are constrained based on the dual-source image data to construct a noise mapping relationship. S400, based on the trained self-supervised denoising network model, noise correction is performed on the optimized dual-source image data to obtain the target reconstructed image.

[0057] In this embodiment, the dual-source data joint denoising method is applicable to DS-CBCT equipment. The dual-source structure refers to the imaging system configured with two independent X-ray sources and corresponding detectors, symmetrically arranged at a certain angle, enabling simultaneous data acquisition from different angles. During DS-CBCT image reconstruction, the noise structure in the reconstructed image is quite complex due to the influence of multiple noise sources. This application first performs secondary reconstruction processing on the original data acquired through dual-source acquisition. By statistically analyzing the two images separately and extracting their respective noise features, the noise complexity is reduced within the common area of ​​the images, providing more stable and less complex input data for subsequent self-supervised denoising processing.

[0058] Compared to existing technologies, the improvements in this application achieve the following: Highly efficient noise correction: Utilizing the inherent redundancy between dual-source data, self-supervised noise modeling is achieved even without clean image supervision, fundamentally suppressing quantum noise, electronic noise, and noise caused by algorithmic deficiencies. Simplified data preprocessing: A secondary reconstruction process reduces noise complexity, providing stable input to the self-supervised network and improving overall denoising performance. Enhanced image quality: After denoising, DS-CBCT images show significant improvements in tissue resolution and detail, providing more reliable imaging evidence for clinical diagnosis. Strong adaptability: This application does not rely on external noise-free images as supervisory labels, making it suitable for DS-CBCT imaging scenarios with various noise levels, demonstrating good adaptability and application value.

[0059] This application provides another embodiment of a dual-source data joint denoising method. Based on any embodiment of the above method, step S200 includes step S210: stitching together and reconstructing the initially reconstructed dual-source image data, including the upper source image and the lower source image, to obtain a secondary reconstructed image. Specifically, it includes the following steps: refer to the appendix to the specification. Figure 2 As shown, Figure 2 This is a schematic diagram illustrating the process of stitching and reconstructing dual-source image data in an embodiment of this application.

[0060] S211, The dual-source image data is stitched together to obtain stitched image data. Specific steps include: aligning the dual-source image data to the same coordinate system using a geometric registration method; and fusing image pixels in the overlapping common areas of the dual-source image data based on a data interpolation algorithm. S212, The stitched image data is segmented using a threshold segmentation algorithm. Specific steps include: segmenting the skeletal region and soft tissue region by setting an appropriate threshold; and replacing the pixel values ​​of the segmented soft tissue region with preset constant values. S213, The thresholded soft tissue region data and skeletal region data are merged to obtain merged image data. S214, The merged image data is back-projected, and then the reconstruction algorithm is used to perform secondary reconstruction on the back-projected data to obtain the secondary reconstructed image.

[0061] In this embodiment, during CT imaging, the upper and lower source images acquired simultaneously by the dual-source structure first need to be stitched together. This step is achieved through precise geometric registration and projection data interpolation algorithms to ensure that the projection data acquired by the two X-ray tubes are accurately aligned in spatial position, thus obtaining stitched image data.

[0062] After stitching, a thresholding algorithm is used to process the stitched data. For example, a thresholding algorithm based on grayscale histogram analysis accurately segments the soft tissue region by setting an appropriate threshold and replaces the pixel values ​​of these regions with pre-determined constants. This process effectively reduces the interference of soft tissue on subsequent reconstruction. Then, the thresholded soft tissue region data and the segmented skeletal region data are merged pixel by pixel. The merging process involves weighted superposition or logical operations on pixel values, generating merged image data.

[0063] After data merging, a second reconstruction is performed on the merged image data using a back-projection reconstruction algorithm. This second reconstruction process reduces noise complexity, provides stable input for the self-supervised network, and improves the overall denoising effect. In some implementations, the reconstruction algorithm is the FDK (Feldkamp-Davis-Kress) reconstruction algorithm.

[0064] This application provides another embodiment of a dual-source data joint denoising method. Based on any embodiment of the above method, step S200 further includes step S220: extracting inherent noise from the secondary reconstructed image, and performing noise optimization processing on the dual-source image data based on the inherent noise. (See attached specification) Figure 3 As shown, Figure 3 This is a schematic diagram of the noise optimization process for dual-source image data in an embodiment of this application.

[0065] In some implementations, step S220 specifically includes: S221, calculating the data difference between the merged image data and the secondary reconstructed image for common regions that overlap spatially in the secondary reconstructed images, to obtain the inherent noise introduced by the reconstruction algorithm; S222, removing the inherent noise based on the dual-source image data to obtain the optimized dual-source image data.

[0066] In this embodiment, the upper and lower source images in the secondary reconstructed image have a partially overlapping common region in space. This common region can be used to quantify the systematic noise introduced by the reconstruction algorithm. The secondary reconstructed image and the merged image data are the image data before and after the secondary reconstruction, respectively. The inherent noise introduced by the reconstruction method can be obtained by subtracting the data before and after the secondary reconstruction.

[0067] In practice, corresponding pixels within the common region of the two images are compared point-by-point to calculate the grayscale difference between the merged image data and the secondary reconstructed image. This difference matrix represents the structured inherent noise introduced by the iterative reconstruction, filtering, and backprojection algorithms. Subtracting this inherent noise from the initially reconstructed dual-source projection data reduces the noise introduced by the reconstruction method. This process not only reduces the overall noise level of the image but, more importantly, reduces the spatial correlation of noise, thereby significantly improving the noise power spectrum characteristics of the image. This results in a final output image that retains its original image information while exhibiting lower noise complexity, providing a more reliable image foundation for subsequent model training.

[0068] This application achieves targeted noise suppression by statistically analyzing noise within common regions and separating the inherent noise components introduced by the reconstruction algorithm. Compared to global filtering, this method better preserves the edge and detail features of anatomical structures. While reducing noise complexity, it significantly improves the signal-to-noise ratio and low-contrast resolution of the image.

[0069] Based on the above embodiments, this application discloses another embodiment of a dual-source data joint denoising method. In this embodiment, without clean images as supervisory information, the application fully utilizes the inherent correlation between dual-source data in a DS-CBCT device to design a self-supervised denoising network model. Its core structure includes:

[0070] The feature extraction module is used to extract features of noise signals and real signals in the input image using a multi-layer convolutional neural network, obtain noise distribution features, and fuse information at multiple scales.

[0071] Noise mapping learning module: used to construct a noise mapping matrix based on the overlap information of the input image in the common region, so as to learn the correspondence between noise distribution and real signal in a self-supervised manner.

[0072] Denoising and reconstruction module: used to remove noise components from the input image based on the obtained noise mapping prediction results, and output the denoised target reconstructed image.

[0073] In this embodiment, the self-supervised denoising network model is based on the TDSR-Net network model. During the feature extraction stage, a multi-layer convolutional neural network architecture with skip connections is employed, extracting local and global features of the input image layer by layer through convolutional kernels and activation functions at different levels. In some implementations, shallow networks capture high-frequency detail features of noise, while deep networks learn low-frequency structural features of the real signal and utilize pyramid pooling to fuse multi-resolution feature information at multiple scales, ultimately outputting a feature vector containing noise distribution characteristics. In other implementations, the TDSR-Net network mainly consists of three parts: an initial convolutional layer, a deep feature extraction module, and a final reconstruction layer. The deep feature extraction module is composed of multiple residual-in-residual dense blocks (RRDB modules) stacked together. Each RRDB module contains three dense blocks and a 1×1 convolutional layer. The DenseBlocks employ the idea of ​​dense connections; in each DenseBlock, the input to each convolutional layer is obtained by concatenating the outputs of all preceding layers. The 1×1 convolutional layer is used to fuse the output of the DenseBlock and compress the number of channels back to the original size.

[0074] The noise mapping learning module is based on a self-supervised learning mechanism. First, it establishes a noise-signal correspondence in the common region of the input image through an image patch matching algorithm. Then, it uses a two-branch Transformer network with an attention mechanism to analyze the spatial correlation between noise features and the real signal and construct a noise mapping matrix. Each pixel value in this matrix represents the probability that the corresponding position belongs to noise.

[0075] The denoising and reconstruction module integrates the results of feature extraction and noise mapping, and uses pixel-level weighted fusion to subtract the predicted noise distribution from the original input image, outputting a high-quality target reconstruction image.

[0076] In some embodiments, the self-supervised denoising network model described in this application, based on the above embodiments, further includes: a data input module for receiving optimized dual-source image data as the input image of the network model. (See attached specification) Figure 4 As shown. Figure 4 This is a schematic diagram illustrating the application process of the self-supervised denoising network model in the embodiments of this application.

[0077] In other embodiments, during the model training process, the self-supervised denoising network model utilizes the optimized dual-source image data, including the optimized upper source image and the optimized lower source image, as mutual supervision signals: when the optimized upper source image is used as network input, the corresponding optimized lower source image is used as the label; conversely, when the optimized lower source image is used as network input, the corresponding optimized upper source image is used as the label.

[0078] In this embodiment, the Noise2Noise self-supervised learning framework is used to achieve medical image denoising. The innovation lies in constructing training sample pairs using optimized upper and lower source images acquired by a dual-source imaging system, without relying on clean real images as supervision signals. Specifically, when the upper source image is input, the lower source image is used as the label, and vice versa.

[0079] Based on the same concept, this application also discloses a dual-source data joint denoising system. The system is used to implement the steps described in any of the above method embodiments. Specifically, one embodiment of the dual-source data joint denoising system of this application includes:

[0080] The data acquisition module is used to acquire dual-source projection data based on the dual-source structure, including upper source projection data and lower source projection data.

[0081] The noise optimization module is used to perform preliminary reconstruction of the dual-source projection data using a reconstruction algorithm. It is also used to stitch together the preliminarily reconstructed dual-source image data (including the upper and lower source images) and perform a secondary reconstruction to obtain a secondary reconstructed image. Inherent noise is extracted from the secondary reconstructed image, and noise optimization processing is performed on the dual-source image data based on this inherent noise.

[0082] A self-supervised network denoising module is used to establish a self-supervised denoising network model, taking optimized dual-source image data as input to the model. Based on the dual-source image data, the noise distribution characteristics are constrained to construct a noise mapping relationship.

[0083] The self-supervised network denoising module is also used to perform noise correction on the optimized dual-source image data based on the trained self-supervised denoising network model to obtain the target reconstructed image.

[0084] Based on the above embodiments, this application discloses another embodiment of a dual-source data joint denoising system, wherein the noise optimization module includes the following sub-modules:

[0085] The stitching submodule is used to stitch the dual-source image data to obtain stitched image data: the dual-source image data is aligned to the same coordinate system using a geometric registration method; and image pixels are fused in the common overlapping areas of the dual-source image data based on a data interpolation algorithm.

[0086] The segmentation submodule is used to segment the stitched image data using a threshold segmentation algorithm: by setting an appropriate threshold, the skeletal region and the soft tissue region are segmented; and the pixel values ​​of the segmented soft tissue region are replaced with preset constant values.

[0087] The merging submodule is used to merge thresholded soft tissue region data with bone region data to obtain merged image data.

[0088] The back-projection submodule is used to perform back-projection processing on the merged image data, and then use the reconstruction algorithm to perform secondary reconstruction on the back-projected data to obtain the secondary reconstructed image.

[0089] Based on the above embodiments, this application discloses another embodiment of a dual-source data joint denoising system. The noise optimization module further includes: a noise calculation submodule, used to calculate the data difference between the merged image data and the secondary reconstructed image for common regions that overlap spatially in the secondary reconstructed image, to obtain the inherent noise introduced by the reconstruction algorithm; and also used to remove the inherent noise based on the dual-source image data to obtain the optimized dual-source image data.

[0090] Based on the same concept, this application also discloses a DS-CBCT device, which includes a dual-source data joint denoising system as described in any of the above embodiments.

[0091] The dual-source data joint denoising method, system, and device of this application have the same technical concept, and the technical details of the embodiments of the two are applicable to each other. In order to reduce repetition, they will not be repeated here.

[0092] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of program modules is merely an example. In practical applications, the above functions can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program units or modules to complete all or part of the functions described above. The program 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 processing unit. The integrated unit can be implemented in hardware or as a software program unit. Furthermore, the specific names of the program modules are only for easy differentiation and are not intended to limit the scope of protection of this application.

[0093] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for joint denoising of dual-source data, characterized in that, Includes the following steps: Based on the dual-source structure, dual-source projection data is acquired, and a reconstruction algorithm is used to perform preliminary reconstruction of the dual-source projection data respectively; The dual-source image data obtained from the initial reconstruction, including the upper source image and the lower source image, is stitched together and reconstructed a second time to obtain a secondary reconstructed image; inherent noise is extracted from the secondary reconstructed image, and noise optimization processing is performed on the dual-source image data based on the inherent noise. A self-supervised denoising network model is established, and the optimized dual-source image data is used as the input of the self-supervised denoising network model; the noise distribution characteristics are constrained based on the dual-source image data, and a noise mapping relationship is constructed. Based on the trained self-supervised denoising network model, noise correction is performed on the optimized dual-source image data to obtain the target reconstructed image; The process of stitching together and reconstructing the dual-source image data obtained from the initial reconstruction (including the upper and lower source images) to obtain a secondary reconstructed image specifically includes: The dual-source image data is stitched together to obtain stitched image data: the dual-source image data is aligned to the same coordinate system using a geometric registration method; and image pixels are fused in the common overlapping areas of the dual-source image data based on a data interpolation algorithm. The stitched image data is segmented using a threshold segmentation algorithm: the skeletal region and the soft tissue region are segmented by setting an appropriate threshold; and the pixel values ​​of the segmented soft tissue region are replaced with preset constant values. The thresholded soft tissue region data is merged with the bone region data to obtain merged image data; The merged image data is back-projected, and then the reconstruction algorithm is used to reconstruct the back-projected data a second time to obtain the reconstructed image. During the training process, the self-supervised denoising network model utilizes the optimized dual-source image data, including the optimized upper source image and the optimized lower source image, as mutual supervision signals. When the optimized source image is used as network input, the corresponding optimized source image is used as the label; conversely, when the optimized source image is used as network input, the corresponding optimized source image is used as the label.

2. The dual-source data joint denoising method as described in claim 1, characterized in that, The process involves extracting inherent noise from the reconstructed image and performing noise optimization processing on the dual-source image data based on the inherent noise. Specifically, it includes: For common regions that overlap spatially in the reconstructed images, the data difference between the merged image data and the reconstructed image is calculated to obtain the inherent noise introduced by the reconstruction algorithm; Based on the dual-source image data, the inherent noise is removed to obtain the optimized dual-source image data.

3. The dual-source data joint denoising method as described in claim 1, characterized in that, The self-supervised denoising network model includes: The feature extraction module is used to extract the features of noise signals and real signals in the input image using a multi-layer convolutional neural network, obtain noise distribution features, and fuse information at multiple scales; Noise mapping learning module: used to construct a noise mapping matrix based on the overlap information of the input image in the common region, so as to learn the correspondence between noise distribution and real signal in a self-supervised manner; Denoising and reconstruction module: used to remove noise components from the input image based on the obtained noise mapping prediction results, and output the denoised target reconstructed image.

4. A dual-source data joint denoising method as described in any one of claims 1-3, characterized in that: The reconstruction algorithm is the FDK reconstruction algorithm.

5. A dual-source data joint denoising system, characterized in that, The system is used to implement the dual-source data joint denoising method according to any one of claims 1-4, comprising: Data acquisition module; used to acquire dual-source projection data based on a dual-source structure, including upper source projection data and lower source projection data; The noise optimization module is used to perform preliminary reconstruction of the dual-source projection data using a reconstruction algorithm; it is also used to stitch together and perform secondary reconstruction of the dual-source image data obtained from the preliminary reconstruction, including the upper source image and the lower source image, to obtain a secondary reconstructed image; it extracts inherent noise from the secondary reconstructed image and performs noise optimization processing on the dual-source image data based on the inherent noise; A self-supervised network denoising module is used to establish a self-supervised denoising network model, taking optimized dual-source image data as input to the self-supervised denoising network model; constraining the noise distribution characteristics based on the dual-source image data, and constructing a noise mapping relationship; The self-supervised network denoising module is also used to perform noise correction on the optimized dual-source image data based on the trained self-supervised denoising network model to obtain the target reconstructed image.

6. The dual-source data joint denoising system as described in claim 5, characterized in that, The noise optimization module includes: The stitching submodule is used to stitch the dual-source image data to obtain stitched image data: aligning the dual-source image data to the same coordinate system using a geometric registration method; and performing image pixel fusion on the common overlapping areas of the dual-source image data based on a data interpolation algorithm. The segmentation submodule is used to segment the stitched image data using a threshold segmentation algorithm: by setting an appropriate threshold, the skeletal region and the soft tissue region are segmented; and the pixel values ​​of the segmented soft tissue region are replaced with preset constant values. The merging submodule is used to merge the thresholded soft tissue region data with the bone region data to obtain merged image data. The back-projection submodule is used to perform back-projection processing on the merged image data, and then use the reconstruction algorithm to perform secondary reconstruction on the back-projected data to obtain the secondary reconstructed image.

7. A dual-source data joint denoising system as described in claim 6, characterized in that, The noise optimization module further includes: The noise calculation submodule is used to calculate the data difference between the merged image data and the secondary reconstructed image for common regions that overlap in space in the secondary reconstructed image, to obtain the inherent noise introduced by the reconstruction algorithm; it is also used to remove the inherent noise based on the dual-source image data to obtain the optimized dual-source image data.

8. A DS-CBCT device, characterized in that, The device includes a dual-source data joint denoising system as described in any one of claims 5-7.