Image correction model training method and device, equipment and storage medium

Abstract

The invention discloses an image correction model training method and device, equipment and a storage medium, and the method comprises the steps: obtaining a first scanning image, and carrying out the generation of the first scanning image without a corresponding scanning sine image, and obtaining a plurality of scanning sine images in one-to-one correspondence; performing back projection on the scanning sine image, performing summation to obtain a sparse angle scanning image, and comparing the sparse angle scanning image with the full-angle scanning image to obtain a back projection tensor error and a discrete sampling error; constructing an initial network model containing a common convolution module and a traditional linear difference module, and correcting the two modules according to an error to obtain an intermediate network model; and inputting the sine scanning image into the intermediate model to obtain a corrected scanning image, calculating a structural similarity index loss value of the corrected scanning image and the first scanning image, and correcting the intermediate model according to the loss value to obtain a target network model. Projection tensor error and discrete sampling problems in a sparse view angle scanning image can be solved, the precision and detail retention capability of a reconstructed image are improved, and the adaptability of a model to sparse sampling is enhanced.

Description

Technical Field

[0001] The present application relates to the field of image recognition technology, and in particular to a training method, apparatus, device and storage medium for an image correction model. Background Art

[0002] With the advancement of medical imaging technology, sparse-view scanning imaging has gained widespread application due to its advantages, such as low radiation dose. However, existing traditional filtered back-projection methods struggle to effectively suppress errors in sparse-view scenarios, and some deep learning methods lack targeted modeling of rotational inconsistencies and continuity in the detector distance dimension, resulting in unstable reconstructed image quality. In particular, in extremely sparse sampling scenarios, existing technologies lack the ability to align the angles of the back-projected tensor, resulting in low interpolation accuracy at discrete sampling points, further reducing image reconstruction accuracy.

[0003] Therefore, how to efficiently deal with back-projection tensor errors and discrete sampling errors to improve the accuracy and robustness of scanned image reconstruction under sparse perspective is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the training method, apparatus, device, and storage medium of an image correction model provided in the embodiments of the present application can efficiently process back-projection tensor errors and discrete sampling errors, thereby improving the accuracy and robustness of scanned image reconstruction under sparse viewing angles. The training method, apparatus, device, and storage medium of an image correction model provided in the embodiments of the present application are implemented as follows:

[0005] The present invention provides a method for training an image correction model, including:

[0006] Acquire multiple first scan images, determine whether each first scan image has a corresponding scanning sinusoidal image, and if no corresponding scanning sinusoidal image exists for the current first scan image, generate a scanning sinusoidal image corresponding to the current first scan image, thereby obtaining multiple scanning sinusoidal images that correspond one-to-one to the multiple first scan images;

[0007] Obtain all scanning sinusoidal images, perform back-projection and summation processing on each scanning sinusoidal image to obtain a sparse angle scanning image;

[0008] Acquire a full-angle scan image, compare the sparse angle scan image with the full-angle scan image, and obtain a back-projection tensor error and a discrete sampling error;

[0009] Constructing an initial network model, wherein the initial network model includes a common convolution module and a traditional linear difference module;

[0010] The ordinary convolution module is corrected according to the back-projection tensor error to obtain a rotatable convolution module, and the traditional linear difference module is corrected according to the discrete sampling error to obtain a two-dimensional local continuous representation difference module, and the rotatable convolution module and the two-dimensional local continuous representation difference module are integrated to obtain an intermediate network model;

[0011] The sinusoidal scan image is input into the intermediate network model to obtain a corrected scan image, a structural similarity index loss value between the corrected scan image and the first scan image is calculated, and the intermediate network model is corrected according to the structural similarity index loss value to obtain a target network model.

[0012] In some embodiments, the step of acquiring all scanning sinusoidal images and performing back-projection and summation processing on each scanning sinusoidal image to obtain a sparse angle scanning image further includes:

[0013] Each of the scanned sinusoidal images is filtered to obtain a filtered scanned sinusoidal image.

[0014] In some embodiments, acquiring all scanning sinusoidal images, performing back-projection and summation processing on each scanning sinusoidal image, and obtaining a sparse angle scanning image, includes:

[0015] Acquire all scanning sinusoidal images, perform back-projection and summation processing on each scanning sinusoidal image according to formula (1) to obtain a sparse angle scanning image;

[0016]

[0017] in, is the filtered scanned sinusoidal image, q is the filter kernel, is the convolution operation, p θ is the scanned sinusoidal image, θ is the back-projection angle, H θ [i,j] is the back-projected image at the viewing angle θ, [i,j] is the coordinate of the back-projected image, is the back projection function, x i 、y j is the spatial coordinate corresponding to the sparse angle scanned image, where i and j are the pixel coordinate indices in the image. is the sparse angle scan image, M is the number of sparse angle scan images, θ m The angle of the image scanned for the mth sparse angle.

[0018] In some embodiments, the ordinary convolution module is corrected according to the back-projection tensor error to obtain a rotatable convolution module, and the traditional linear difference module is corrected according to the discrete sampling error to obtain a two-dimensional local continuous representation difference module. The rotatable convolution module and the two-dimensional local continuous representation difference module are combined to obtain an intermediate network model, including:

[0019] Analyzing and processing the back-projection tensor error to obtain a distribution feature of rotation inconsistency of the sparse angular scan image;

[0020] Modifying the learnable combination coefficients and basis function parameters in the common convolution module according to the distribution characteristics to obtain a modified rotatable convolution module;

[0021] Correcting the traditional linear difference module according to the discrete sampling error to obtain a corrected two-dimensional local continuous representation difference module;

[0022] The modified rotatable convolution module and the modified two-dimensional local continuous representation difference module are integrated to obtain an intermediate network model.

[0023] In some embodiments, the method of correcting the traditional linear difference module according to the discrete sampling error to obtain a corrected two-dimensional local continuous representation difference module includes:

[0024] Analyzing and processing the discrete sampling error to obtain an under-sampling area;

[0025] The basis function combination mode in the traditional linear difference module is corrected according to the insufficiently sampled area to obtain a corrected two-dimensional local continuous representation difference module.

[0026] In some embodiments, the method of acquiring multiple first scanned images, determining whether each first scanned image has a corresponding scanning sinusoidal image, and generating a scanning sinusoidal image corresponding to the current first scanned image when no corresponding scanning sinusoidal image exists for the current first scanned image, and obtaining multiple scanning sinusoidal images corresponding one-to-one to the multiple first scanned images, further includes:

[0027] Data quality optimization processing is performed on the plurality of scanning sinusoidal images to obtain processed scanning sinusoidal images.

[0028] In some embodiments, the number of sampling angles of the full-angle scanning image is 1152, and the number of sampling angles of the sparse-angle scanning image is 72.

[0029] An embodiment of the present application provides a training device for an image correction model, comprising:

[0030] an acquisition module, configured to acquire a plurality of first scanned images, determine whether a corresponding scanning sinusoidal image exists for each first scanned image, and generate a scanning sinusoidal image corresponding to the current first scanned image if no corresponding scanning sinusoidal image exists for the current first scanned image, thereby obtaining a plurality of scanning sinusoidal images corresponding one-to-one to the plurality of first scanned images;

[0031] A processing module is used to obtain all scanning sinusoidal images, and perform back-projection and summation processing on each scanning sinusoidal image to obtain a sparse angle scanning image;

[0032] The processing module is further configured to obtain a full-angle scan image, compare the sparse angle scan image with the full-angle scan image, and obtain a back-projection tensor error and a discrete sampling error;

[0033] A construction module, used to construct an initial network model, wherein the initial network model includes a common convolution module and a traditional linear difference module;

[0034] The processing module is further configured to perform correction processing on the ordinary convolution module according to the back-projection tensor error to obtain a rotatable convolution module, and perform correction processing on the traditional linear difference module according to the discrete sampling error to obtain a two-dimensional local continuous representation difference module, and integrate the rotatable convolution module and the two-dimensional local continuous representation difference module to obtain an intermediate network model;

[0035] The sinusoidal scan image is input into the intermediate network model to obtain a corrected scan image, a structural similarity index loss value between the corrected scan image and the first scan image is calculated, and the intermediate network model is corrected according to the structural similarity index loss value to obtain a target network model.

[0036] The computer device provided in an embodiment of the present application includes a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the program, the method described in the embodiment of the present application is implemented.

[0037] The computer-readable storage medium provided in the embodiment of the present application stores a computer program thereon, and when the computer program is executed by a processor, the method provided in the embodiment of the present application is implemented.

[0038] The embodiments of the present application provide a training method, apparatus, computer device, and computer-readable storage medium for an image correction model. The method obtains a first scanned image, generates multiple corresponding scanned sinusoidal images for the first scanned image without a corresponding scanned sinusoidal image, back-projects and sums the scanned sinusoidal images to obtain a sparse angle scanned image, and compares it with the full-angle scanned image to obtain the back-projection tensor error and discrete sampling error. An initial network model containing a common convolution module and a traditional linear difference module is constructed, and an intermediate network model is obtained based on the error correction modules. The sinusoidal scanned image is input into the intermediate model to obtain a corrected scanned image, and the structural similarity index loss value between the image and the first scanned image is calculated. The intermediate model is corrected based on the loss value to obtain a target network model. The method can solve the problems of projection tensor error and discrete sampling in sparse angle scanned images, improve the accuracy of reconstructed images and the ability to retain details, enhance the adaptability of the model to sparse sampling, and solve the technical problems raised in the background technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments of the present application or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 A schematic diagram of an implementation flow of a training method for an image correction model provided in an embodiment of the present application;

[0041] Figure 2 A schematic diagram of the implementation flow of model correction in a training method for an image correction model provided in an embodiment of the present application;

[0042] Figure 3 A training device for an image correction model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] The following description of some of the technologies involved in the embodiments of this application is provided to facilitate understanding and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for the sake of clarity and conciseness, some descriptions of well-known functions and structures are omitted from the following description.

[0045] Figure 1 This is a schematic diagram of the implementation flow of a training method for an image correction model provided in an embodiment of the present application, including steps 101 to 106. Figure 1 This is only an execution order shown in the embodiment of the present application, and does not represent the only execution order of a training method for an image correction model. If the final result can be achieved, Figure 1 The steps shown may be performed in parallel or reversed.

[0046] Step 101: Acquire multiple first scan images, determine whether each first scan image has a corresponding scan sinusoidal image, and when no corresponding scan sinusoidal image exists for the current first scan image, generate a scan sinusoidal image corresponding to the current first scan image, thereby obtaining multiple scan sinusoidal images that correspond one-to-one to the multiple first scan images.

[0047] In an embodiment of the present application, a plurality of medical CT (Computed Tomography) first scan images are obtained, and for each image, it is determined whether a corresponding scanning sinusoidal image (i.e., X-ray projection data) already exists. If not, a scanning sinusoidal image corresponding to the image is generated by Radon transform. Specifically, using the physical principle of X-ray attenuation, the HU value (Henry unit, a standardized unit used in medical CT images to quantify the degree of tissue attenuation of radiation) of the CT image is converted into a linear attenuation value distribution through formula (1), and then projection data is generated by line integral calculation, and finally a scanning sinusoidal image corresponding to each first scan image is obtained.

[0048] V μ =V CT *0.0192 / 1000+0.0192(1)

[0049] Among them, V μ is the attenuation value, V CT is the HU value corresponding to the first scan image.

[0050] Step 102 : Acquire all scanning sinusoidal images, perform back-projection and summation processing on each scanning sinusoidal image, and obtain a sparse angle scanning image.

[0051] In an embodiment of the present application, after obtaining all generated scanned sinusoidal images, a filtered back-projection process is performed on each sinusoidal image. Specifically, the sinusoidal image is first filtered using a filter such as Ram-Lak, and then the filtered projection data is mapped to the image space using a back-projection algorithm to generate a single-view back-projection image. Subsequently, the back-projection images of all viewpoints are summed to obtain a CT reconstructed image at a sparse angle (i.e., a sparse angle scan image).

[0052] Step 103 : Acquire a full-angle scan image, compare the sparse angle scan image with the full-angle scan image, and obtain a back-projection tensor error and a discrete sampling error.

[0053] In this embodiment of the present application, a full-angle scan image of the same scanned object (e.g., a reconstructed CT image with a sampling geometry of 1152 angles) is obtained and compared pixel by pixel with the sparse angle scan image obtained in step 102. Specifically, by calculating the difference in pixel values ​​between the two, the back-projection tensor error in the angular dimension and the discrete sampling error in the detector and distance dimensions caused by sparse sampling are obtained.

[0054] Step 104: construct an initial network model, which includes a common convolution module and a traditional linear difference module.

[0055] In the embodiment of the present application, an initial network model including two core modules is constructed. The initial network model includes a common convolution module and a traditional linear difference module, wherein:

[0056] Ordinary convolution module: It consists of multiple convolution layers, activation functions (such as ReLU) and normalization layers (such as BatchNorm). It is responsible for extracting the global features of the input image (such as the optimized scanned sinusoidal image), covering information such as contour and texture, and providing basic feature support for subsequent processing.

[0057] Traditional Linear Interpolation Module: This module uses a linear interpolation algorithm to fill in sparsely sampled areas in the image (such as detector edges and signal-sparse areas in low-dose scans). It estimates missing pixels by calculating neighboring pixel values, enhancing image continuity and improving the smoothness and integrity of local details.

[0058] Step 105: Correct the ordinary convolution module according to the back-projection tensor error to obtain a rotatable convolution module, and correct the traditional linear difference module according to the discrete sampling error to obtain a two-dimensional local continuous representation difference module. The rotatable convolution module and the two-dimensional local continuous representation difference module are integrated to obtain an intermediate network model.

[0059] In the embodiment of the present application, the rotatable convolution module addresses the rotation inconsistency of the back-projected tensors at different viewpoints. The module introduces a distance-guided weighted convolution structure, the mathematical expression of which is: Among them, D θ is the distance tensor from each pixel to the virtual detector, ρ1 and ρ2 are convolution kernels with different smoothness levels, is the estimated value of the filtered back-projection tensor, is the convolution operation symbol, is the original back-projection tensor under angle θ. The outputs of the two convolution kernels are adaptively combined through the distance tensor to enhance the feature extraction capability of different distance regions.

[0060] The two-dimensional local continuous representation difference module is used to model the continuity of the detector and the distance dimension. A continuous projection model is constructed through the bicubic interpolation basis function. The basis function expression is a piecewise function, using different polynomial forms in different intervals. It is combined with a radial mask function to limit the convolution response range and improve the interpolation accuracy between discrete sampling points.

[0061] The rotational convolution module is modified based on the distribution of back-projection tensor errors, adjusting the learnable combination coefficients and basis function parameters in the module. Specifically, by analyzing the distribution of rotational errors in the back-projection tensor at different angles, the combination coefficients of the rotational convolution kernel are optimized (i.e., the convolution kernel is trained to rotate and align tensors at different angles). The shape parameters of the bicubic basis functions are also fine-tuned to enhance the module's robustness to angle variations.

[0062] The two-dimensional local continuous representation difference module corrects the under-sampling areas of the detector and distance dimension based on discrete sampling errors. By adjusting the combination of basis functions and introducing an adaptive weight mechanism, the module can dynamically adjust the interpolation weight according to the sampling density, thereby suppressing artifacts caused by insufficient sampling.

[0063] Step 106: Input the sinusoidal scan image into the intermediate network model to obtain a corrected scan image, calculate the structural similarity index loss value between the corrected scan image and the first scan image, and correct the intermediate network model according to the structural similarity index loss value to obtain the target network model.

[0064] In an embodiment of the present application, a scanned sinusoidal image is input into a modified intermediate network model to obtain a modified scanned image. The structural similarity index (SSIM) loss value between the modified image and the first scanned image is calculated, and a loss function is constructed by evaluating the brightness, contrast, and structural information differences of the images. Using a backpropagation algorithm, the parameters of the intermediate network model are iteratively optimized according to the loss value (such as adjusting the convolution kernel weights, basis function combination coefficients, etc.), and finally a target network model that can effectively suppress sparse perspective artifacts is obtained.

[0065] The embodiment of the present application ensures the integrity of the training data by generating missing scanned sinusoidal images, obtains errors by comparing back projection with full-angle images, combines a rotatable module and a two-dimensional local continuous representation difference module, and specifically corrects the model's geometric transformation adaptability and local feature representation capabilities. Finally, the image structure is optimized through the structural similarity index, effectively improving the training accuracy and robustness of the image correction model.

[0066] In some embodiments, all scanning sinusoidal images are acquired, and each scanning sinusoidal image is back-projected and summed to obtain a sparse angle scanning image. The process also includes: filtering each scanning sinusoidal image to obtain a filtered scanning sinusoidal image.

[0067] Specifically, before back-projecting the scanned sinusoidal images, each scanned sinusoidal image is filtered. Specifically, frequency-domain filtering methods, such as the Ram-Lak filter (or other well-known CT reconstruction filters), are used to process the sinusoidal images. This filtering process convolves the filter with the scanned sinusoidal image in the frequency or time domain, enhancing the high-frequency components in the projection data and compensating for information loss caused by sparse sampling, thereby improving the accuracy of subsequent back-projection reconstruction.

[0068] This embodiment of the present application further optimizes the reconstruction quality of sparse angular scanned images by adding a filtering step to the scanned sinusoidal images. This effectively compensates for low-frequency distortion caused by sparse sampling, providing more accurate projection data for back-projection reconstruction, thereby improving the model's ability to retain image detail and its noise immunity.

[0069] In some embodiments, all scanning sinusoidal images are acquired, and each scanning sinusoidal image is back-projected and summed to obtain a sparse angle scanning image, including: acquiring all scanning sinusoidal images, and each scanning sinusoidal image is back-projected and summed to obtain a sparse angle scanning image.

[0070] Specifically, after acquiring all scanned sinusoidal images, each image is filtered to enhance high-frequency information. Specifically, a pre-set filter kernel is convolved with the scanned sinusoidal image. The filtering process can be represented as convolving the filter kernel with the scanned sinusoidal image in either the time or frequency domain. Frequency-domain filtering requires first converting the image to the frequency domain via a Fourier transform, then multiplying it with the filter's frequency-domain representation, and then inversely transforming it back to the time domain.

[0071] The back projection operation is performed on the filtered scanned sinusoidal image to map the projection data to the image space. The specific steps are as follows: for each back projection angle θ, the back projection function is used to distribute and accumulate the filtered projection data along the projection line of the angle. The mathematical expression of the back projection function must satisfy the physical process of X-ray attenuation, and each point on the projection line is mapped to the corresponding coordinate in the image space according to the geometric position. For example, for the coordinate point (x i ,y j ), the projection line position at angle θ is x i cosθ+y j sinθ, the back projection function calculates the coordinate point (x i ,y j ) is assigned to generate the back-projection image H under a single perspective θ [i,j].

[0072] The back-projected images of all sparse view angles are summed to obtain the final sparse angle scan image. Specifically, assuming that the sampling angles of the sparse view angle are M, in the embodiment of the present application, the value of M is 72 sampling angles, the sampling angles of the full angle scan image are 1152, and the angle of each view angle is After accumulating the back-projected images of each viewpoint, multiply them by the normalization coefficient This normalization operation ensures that the intensity of the reconstructed image is consistent with the full-angle scan, avoiding brightness deviations caused by sparse viewing angles. The mathematical expression of the summation process is the pixel-by-pixel accumulation of the back-projected images from all viewing angles. The resulting sparse viewing angle scan image integrates the projection information of multiple sparse viewing angles, reducing artifacts associated with single-view reconstruction.

[0073] Filter processing formula: Corresponding to the convolution operation, where is the filtered scanned sinusoidal image, q is the filter kernel, p θ is the original scanned sinusoidal image, Represents the convolution operation;

[0074] Back projection formula: Corresponding to the projection line mapping process, where is the back projection function, (x i ,y j ) is the image space coordinate, θ is the projection angle, H θ [i,j] is the back-projected image at the viewing angle θ;

[0075] Summation formula: Corresponding to the normalized accumulation operation, where M is the number of sparse viewpoints, is the sparse angle scan image, θ mis the angle of the m-th view, and the final sparse angle scan image is generated by summing and normalizing.

[0076] The embodiment of the present application mathematizes the filtering, back-projection and summation processes of the scanned sinusoidal image. The filtering process uses the well-known CT reconstruction filter to enhance the high-frequency components of the projection data. The back-projection process maps the projection data to the image space based on the inverse operation of the Radon transform. The summation step compensates for the sparse view sampling deviation through normalized accumulation, ensuring the feasibility and accuracy of the technical solution and effectively improving the accuracy and reliability of sparse view CT image reconstruction.

[0077] In the above Figure 1 On the basis of , this application also provides a schematic diagram of the implementation process of model correction in the training method of an image correction model, such as Figure 2 As shown, it includes steps 201 to 204:

[0078] Step 201 : Analyze and process the back-projection tensor error to obtain the distribution characteristics of the rotation inconsistency of the sparse angular scan image.

[0079] In the present embodiment, the back-projection tensor error is quantitatively analyzed by comparing the sparse angle scan image with the full angle scan image pixel by pixel, calculating the difference in pixel values ​​between the two, and locating areas with significant error. By statistically analyzing the error distribution of the back-projection tensor, the spatial characteristics of rotational inconsistency are obtained. For example, by analyzing how the error changes with distance from the virtual detector, it is possible to determine which position intervals in the back-projection tensor exhibit significant rotational deviation.

[0080] Step 202: Modify the learnable combination coefficients and basis function parameters in the common convolution module according to the distribution characteristics to obtain a modified rotatable convolution module.

[0081] In the embodiment of the present application, the convolution kernel of the rotatable convolution module is composed of a set of convolution bases and learnable combination coefficients as shown in the formula As shown, the combination coefficient w n Control the weight of each convolution basis. According to the distribution of rotation inconsistency, w is optimized through training n The value of , enables the convolution kernel to more accurately rotate and align the back-projection tensor of the error-sensitive area, ρ(z,θ) is the convolution kernel, σ n is the basis function, z is the spatial coordinate, which is used to describe the position of the pixel in the image, θ is the scanning angle, which is used to describe the rotation orientation of the detector, N is the number of convolution bases, n is the serial number of the basis function, T θ is the angle rotation matrix, which is used to rotate the spatial coordinates.

[0082] In locations with large errors, the corresponding convolution basis combination coefficients are increased to enhance the convolution kernel's ability to adjust to that area. For example, at the edge of an image where errors are concentrated, by learning specific convolution basis combination coefficients, the model can better adapt the convolution kernel to local error characteristics and maintain consistency in adjustment at different angles, thereby achieving a rotatable error mitigation effect.

[0083] Step 203: Correct the traditional linear difference module according to the discrete sampling error to obtain a corrected two-dimensional local continuous representation difference module.

[0084] In an embodiment of the present application, for discrete sampling errors, by analyzing the distribution of discrete sampling errors, sampling missing areas in the detector dimension and the distance dimension (such as discontinuous areas of projection data caused by sparse sampling) are determined.

[0085] The module uses bicubic interpolation basis functions and radial mask functions to construct a continuous projection model. The module adjusts the weights of the basis functions based on the distribution of undersampled areas. For example, in areas with low sampling density, the weights of the basis functions of adjacent sampling points are increased to enhance continuity through local interpolation. In areas with high sampling density, the default combination of basis functions is maintained to avoid overfitting.

[0086] An adaptive weighting mechanism based on sampling density is introduced to dynamically adjust the contribution of each sampling point to the reconstruction result. For example, for sparsely sampled detector locations, adjacent sampling points are given higher interpolation weights to suppress artifacts caused by undersampling; for densely sampled areas, uniform weighting is used to ensure detail preservation.

[0087] Step 204 , integrating the modified rotatable convolution module and the modified two-dimensional local continuous representation difference module to obtain an intermediate network model.

[0088] In an embodiment of the present application, the corrected rotatable convolution module is integrated with the two-dimensional local continuous representation difference module. The back-projection tensor is first angle-aligned by the rotatable convolution module to eliminate the rotation inconsistency between different perspectives. The processed tensor is then input into the two-dimensional local continuous representation difference module to perform continuity modeling on the detector and distance dimensions and compensate for discrete sampling errors.

[0089] The two modules process the back-projection tensor in parallel. The rotatable convolution module outputs the angle-aligned tensor, and the two-dimensional local continuous representation difference module outputs the continuous tensor. The output of the intermediate network model is generated through weighted fusion (such as dynamically adjusting the fusion weight according to the error distribution).

[0090] During the integration process, ensure that the input and output dimensions of the two modules match. For example, the output tensor size of the rotatable convolution module is consistent with the input requirements of the two-dimensional local continuous representation difference module, and finally form an intermediate network model that can handle both rotation error and discrete sampling error.

[0091] The embodiment of the present application obtains the distribution characteristics of rotation inconsistency by analyzing the back-projection tensor error, which can specifically locate the geometric transformation inconsistency problem caused by missing angles in sparse angle scanning. Then, by correcting the learnable combination coefficients and basis function parameters of the rotatable convolution module, the adaptability of the module to image transformation at different angles is enhanced, ensuring that the convolution operation maintains equivariant characteristics in the rotation space, and effectively suppressing the reconstruction distortion introduced by angular sparsity. The two-dimensional local continuous representation difference module is corrected for discrete sampling errors to compensate for the loss of image details caused by discontinuous sampling points. By optimizing the parameterization method of the local continuous representation, the model's ability to continuously model discrete data is improved, so that the spatial structure information of the image can still be accurately captured at sparse angles.

[0092] In some embodiments, the conventional linear difference module is corrected according to the discrete sampling error to obtain a corrected two-dimensional local continuous representation difference module, including: analyzing and processing the discrete sampling error to obtain an under-sampling area.

[0093] Specifically, the discrete sampling error is quantitatively analyzed by comparing the sparsely scanned image with the full-angle scan image, and calculating the distribution of sampling point errors in the detector and distance dimensions. By statistically analyzing the spatial distribution of pixel value differences, areas of reconstruction distortion caused by sparse sampling points (i.e., undersampling areas) are identified. For example, at the edge of the detector array or at locations far from the imaging center, significant discrete sampling errors often occur due to low sampling density, resulting in streaky artifacts or blurred structures.

[0094] Furthermore, the basis function combination method in the traditional linear interpolation module is corrected according to the under-sampling area to obtain a corrected two-dimensional local continuous representation interpolation module.

[0095] Specifically, based on the identified undersampled areas, the basis function combination of the two-dimensional local continuous representation difference module is adjusted. This is optimized through a weighted combination of bicubic basis functions, and a continuous projection model is constructed through a linear combination of basis functions. In undersampled areas, the weights of the basis functions corresponding to adjacent sampling points are increased to enhance local interpolation accuracy. For example, in areas with large sampling intervals in the detector dimension, the basis function weights of adjacent sampling points are increased by 20%-30%, compensating for the lack of sampling by enhancing the interpolation coverage.

[0096] The effective range of the mask is dynamically expanded based on the spatial location of undersampled areas. For example, in undersampled areas farther from the imaging center, the convolution kernel size is temporarily increased by 1-2 pixels, so that the response range of the basis function covers more adjacent sampling points, improving the ability to model continuity between discrete points.

[0097] The training process optimizes the combination coefficients of the basis functions, allowing the module to automatically activate a more appropriate combination of basis functions in undersampled areas. For example, using a gradient descent algorithm, with the reconstruction error in undersampled areas as the optimization target, the linear combination coefficients of the basis functions are iteratively adjusted to make the interpolation result of the combined basis functions in this area closer to the actual value of the full-angle scan.

[0098] When dealing with under-sampled areas, the revised module uses weighted bicubic interpolation for sparse sampling points in the detector dimension. The weight is determined by the distance from the sampling point to the target point and the error distribution. The closer the distance or the larger the error, the higher the weight.

[0099] In the distance dimension, combined with the dynamically adjusted radial mask function, the effective range of the basis function is expanded, so that areas far away from the sampling points can still obtain more accurate reconstructed values ​​through neighboring multi-point interpolation.

[0100] This embodiment of the application analyzes discrete sampling errors to identify undersampled areas, making corrections more targeted. This avoids excessive intervention in properly sampled areas and dynamically adjusts the basis function combination only in areas with significant errors, thereby maximizing the accuracy of local details in the reconstructed image while maintaining computational efficiency.

[0101] In some embodiments, multiple first scanning images are acquired, and it is determined whether there is a corresponding scanning sinusoidal image for each first scanning image. When no corresponding scanning sinusoidal image exists for the current first scanning image, a scanning sinusoidal image corresponding to the current first scanning image is generated. After obtaining multiple scanning sinusoidal images corresponding one-to-one to the multiple first scanning images, it also includes: performing data quality optimization processing on the multiple scanning sinusoidal images to obtain the processed scanning sinusoidal images.

[0102] Specifically, after acquiring multiple scanning sinusoidal images corresponding to the first scanning image, data quality optimization processing is performed on each scanning sinusoidal image, specifically including traversing each pixel in the scanning sinusoidal image and calculating its statistical difference (such as mean and standard deviation) with other pixels in the local neighborhood. For outliers that deviate from the statistical threshold (such as outliers caused by detector noise), they are replaced with the median or weighted average in the local neighborhood. For example, for a certain pixel, if its value exceeds 3 times the standard deviation of the neighborhood pixel mean, it is determined to be an outlier and corrected.

[0103] Compare sinusoidal images scanned at different angles to detect and eliminate duplicate data caused by scanning device errors or angle overlap. This method calculates the similarity between the images (such as the Structural Similarity Index (SSIM)). If the similarity between two images exceeds a preset threshold (such as 0.95), the higher-quality image is retained and the duplicate image is deleted.

[0104] For missing areas in the scanned sinusoidal image (such as data loss caused by detector failure), the bicubic interpolation method is used to fill them in. The interpolation calculation is combined with the values ​​of the surrounding valid pixels to generate an estimated value of the missing area.

[0105] The embodiments of this application lay the foundation for high-precision reconstruction at the physical signal level through multi-dimensional optimization of data quality, achieving a three-level data enhancement of "noise suppression, redundancy elimination, and information completion." Compared to traditional, unoptimized scanned sinusoidal images, this can improve the signal-to-noise ratio.

[0106] Although the present application provides method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-creative work. The order of steps listed in this embodiment is only one way of executing the order of many steps and does not represent the only execution order. When the actual device or client product is executed, it can be executed in sequence or in parallel according to the method shown in this embodiment or the accompanying drawings (for example, in a parallel processor or multi-threaded processing environment).

[0107] like Figure 3 As shown, the embodiment of the present application also provides a training device 300 for an image correction model. The device includes:

[0108] An acquisition module 301 is configured to acquire a plurality of first scanned images, determine whether a corresponding scanning sinusoidal image exists for each first scanned image, and generate a scanning sinusoidal image corresponding to the current first scanned image if no corresponding scanning sinusoidal image exists for the current first scanned image, thereby obtaining a plurality of scanning sinusoidal images that correspond one-to-one to the plurality of first scanned images;

[0109] The processing module 302 is used to obtain all scanning sinusoidal images, perform back-projection and summation processing on each scanning sinusoidal image, and obtain a sparse angle scanning image;

[0110] The processing module 302 is further configured to obtain a full-angle scan image, compare the sparse angle scan image with the full-angle scan image, and obtain a back-projection tensor error and a discrete sampling error;

[0111] A construction module 303 is used to construct an initial network model, which includes a common convolution module and a traditional linear difference module;

[0112] The processing module 302 is further configured to correct the ordinary convolution module according to the back-projection tensor error to obtain a rotatable convolution module, correct the traditional linear difference module according to the discrete sampling error to obtain a two-dimensional local continuous representation difference module, and integrate the rotatable convolution module and the two-dimensional local continuous representation difference module to obtain an intermediate network model.

[0113] The sinusoidal scan image is input into the intermediate network model to obtain a corrected scan image, the structural similarity index loss value between the corrected scan image and the first scan image is calculated, and the intermediate network model is corrected according to the structural similarity index loss value to obtain the target network model.

[0114] In some embodiments, the processing module 302 is further configured to perform filtering on each scanned sinusoidal image to obtain a filtered scanned sinusoidal image.

[0115] In some embodiments, the processing module 302 is further configured to obtain all scanned sinusoidal images, perform backprojection and summation processing on each scanned sinusoidal image according to formula (1), and obtain a sparse angle scanned image;

[0116]

[0117] in, is the filtered scanned sinusoidal image, q is the filter kernel, is the convolution operation, p θ is the scanned sinusoidal image, θ is the back projection angle, H θ [i,j] is the back-projected image at the viewing angle θ, [i,j] is the coordinate of the back-projected image, is the back projection function, x i 、y j is the spatial coordinate corresponding to the sparse angle scanned image, where i and j are the pixel coordinate indices in the image. is the sparse angle scan image, M is the number of sparse angle scan images, θ m The angle of the image scanned for the mth sparse angle.

[0118] In some embodiments, the processing module 302 is further configured to analyze and process the back-projection tensor error to obtain a distribution feature of the rotation inconsistency of the sparse angle scan image;

[0119] The processing module 302 is further configured to modify the learnable combination coefficients and basis function parameters in the common convolution module according to the distribution characteristics to obtain a modified rotatable convolution module;

[0120] The processing module 302 is further configured to correct the traditional linear difference module according to the discrete sampling error to obtain a corrected two-dimensional local continuous representation difference module;

[0121] The processing module 302 is further used to integrate the modified rotatable convolution module and the modified two-dimensional local continuous representation difference module to obtain an intermediate network model.

[0122] In some embodiments, the processing module 302 is further configured to analyze and process the discrete sampling error to obtain an under-sampling region;

[0123] The processing module 302 is further configured to modify the basis function combination method in the traditional linear difference module according to the insufficiently sampled area, so as to obtain a modified two-dimensional local continuous representation difference module.

[0124] The processing module 302 is further configured to perform data quality optimization processing on the plurality of scanning sinusoidal images to obtain processed scanning sinusoidal images.

[0125] Some modules in the apparatus described herein may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform specific tasks or implement specific abstract data types. The present application may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communications network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0126] The devices or modules described in the above application embodiments can be implemented by computer chips or physical devices, or by products with certain functions. For ease of description, the above devices are described separately by function in various modules. When implementing the embodiments of this application, the functions of each module can be implemented in the same or multiple software and / or hardware. Of course, a module that implements a certain function can also be implemented by combining multiple sub-modules or sub-units.

[0127] The methods, devices, or modules described in this application can be implemented in the form of computer-readable program code. The controller can be implemented in any appropriate manner. For example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (such as software or firmware) that can be executed by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also know that in addition to implementing the controller in the form of pure computer-readable program code, it is entirely possible to implement the same function of the controller in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be considered as structures within the hardware component. Or even, the means for implementing various functions may be considered to be both a software module for implementing the method and a structure within a hardware component.

[0128] An embodiment of the present application further provides a device comprising: a processor; a memory for storing processor-executable instructions; and when the processor executes the executable instructions, the method described in the embodiment of the present application is implemented.

[0129] The embodiments of the present application also provide a non-volatile computer-readable storage medium having a computer program or instruction stored thereon. When the computer program or instruction is executed, the method described in the embodiments of the present application is implemented.

[0130] In addition, each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist independently, or two or more modules may be integrated into one module.

[0131] The above-mentioned storage medium includes, but is not limited to, random access memory (RAM), read-only memory (ROM), cache, hard disk drive (HDD), or memory card. The memory can be used to store computer program instructions.

[0132] It can be seen from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus necessary hardware. Based on this understanding, the technical solution of the present application can essentially or the part that contributes to the prior art can be embodied in the form of a software product, or it can be embodied through the implementation process of data migration. The computer software product can be stored in a storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a mobile terminal, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0133] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referenced to each other. Each embodiment focuses on the differences from other embodiments. All or part of this application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, mobile communication terminals, multi-processor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0134] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit the present application. Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present application.

Claims

1. A training method for an image correction model, characterized in that: include: Acquire multiple first scan images, determine whether each first scan image has a corresponding scanning sinusoidal image, and if no corresponding scanning sinusoidal image exists for the current first scan image, generate a scanning sinusoidal image corresponding to the current first scan image, thereby obtaining multiple scanning sinusoidal images that correspond one-to-one to the multiple first scan images; Obtain all scanning sinusoidal images, perform back-projection and summation processing on each scanning sinusoidal image to obtain a sparse angle scanning image; Acquire a full-angle scan image, compare the sparse angle scan image with the full-angle scan image, and obtain a back-projection tensor error and a discrete sampling error; Constructing an initial network model, wherein the initial network model includes a common convolution module and a traditional linear difference module; The ordinary convolution module is corrected according to the back-projection tensor error to obtain a rotatable convolution module, and the traditional linear difference module is corrected according to the discrete sampling error to obtain a two-dimensional local continuous representation difference module, and the rotatable convolution module and the two-dimensional local continuous representation difference module are integrated to obtain an intermediate network model; The sinusoidal scan image is input into the intermediate network model to obtain a corrected scan image, a structural similarity index loss value between the corrected scan image and the first scan image is calculated, and the intermediate network model is corrected according to the structural similarity index loss value to obtain a target network model.

2. The method according to claim 1, characterized in that The method of acquiring all scanning sinusoidal images, performing back-projection and summation processing on each scanning sinusoidal image to obtain a sparse angle scanning image further includes: Each of the scanned sinusoidal images is filtered to obtain a filtered scanned sinusoidal image.

3. The method according to claim 2, characterized in that The method of acquiring all scanning sinusoidal images and performing back-projection and summation processing on each scanning sinusoidal image to obtain a sparse angle scanning image includes: Acquire all scanning sinusoidal images, perform back-projection and summation processing on each scanning sinusoidal image according to formula (1) to obtain a sparse angle scanning image; in, is the filtered scanned sinusoidal image, q is the filter kernel, is the convolution operation, p θ is the scanned sinusoidal image, θ is the back projection angle, H θ [i,j] is the back-projected image at the viewing angle θ, [i,j] is the coordinate of the back-projected image, is the back projection function, x i 、y j is the spatial coordinate corresponding to the sparse angle scanned image, where i and j are the pixel coordinate indices in the image. is the sparse angle scan image, M is the number of sparse angle scan images, θ m The angle of the image scanned for the mth sparse angle.

4. The method according to claim 1, wherein The ordinary convolution module is corrected according to the back-projection tensor error to obtain a rotatable convolution module, and the traditional linear difference module is corrected according to the discrete sampling error to obtain a two-dimensional local continuous representation difference module. The rotatable convolution module and the two-dimensional local continuous representation difference module are combined to obtain an intermediate network model, including: Analyzing and processing the back-projection tensor error to obtain a distribution feature of rotation inconsistency of the sparse angular scan image; Modifying the learnable combination coefficients and basis function parameters in the common convolution module according to the distribution characteristics to obtain a modified rotatable convolution module; Correcting the traditional linear difference module according to the discrete sampling error to obtain a corrected two-dimensional local continuous representation difference module; The modified rotatable convolution module and the modified two-dimensional local continuous representation difference module are integrated to obtain an intermediate network model.

5. The method according to claim 4, characterized in that The method of correcting the traditional linear difference module according to the discrete sampling error to obtain a corrected two-dimensional local continuous representation difference module includes: Analyzing and processing the discrete sampling error to obtain an under-sampling area; The basis function combination mode in the traditional linear difference module is corrected according to the insufficiently sampled area to obtain a corrected two-dimensional local continuous representation difference module.

6. The method according to claim 1, wherein The method further comprises: acquiring a plurality of first scanned images, determining whether a corresponding scanning sinusoidal image exists for each first scanned image, and generating a scanning sinusoidal image corresponding to the current first scanned image when no corresponding scanning sinusoidal image exists for the current first scanned image. After obtaining a plurality of scanning sinusoidal images corresponding one-to-one to the plurality of first scanned images, the method further comprises: Data quality optimization processing is performed on the plurality of scanning sinusoidal images to obtain processed scanning sinusoidal images.

7. The method according to claim 1, characterized in that The number of sampling angles of the full-angle scanning image is 1152, and the number of sampling angles of the sparse-angle scanning image is 72.

8. A training device for an image correction model, characterized in that: include: an acquisition module, configured to acquire a plurality of first scanned images, determine whether a corresponding scanning sinusoidal image exists for each first scanned image, and generate a scanning sinusoidal image corresponding to the current first scanned image if no corresponding scanning sinusoidal image exists for the current first scanned image, thereby obtaining a plurality of scanning sinusoidal images corresponding one-to-one to the plurality of first scanned images; A processing module is used to obtain all scanning sinusoidal images, and perform back-projection and summation processing on each scanning sinusoidal image to obtain a sparse angle scanning image; The processing module is further configured to obtain a full-angle scan image, compare the sparse angle scan image with the full-angle scan image, and obtain a back-projection tensor error and a discrete sampling error; A construction module, used to construct an initial network model, wherein the initial network model includes a common convolution module and a traditional linear difference module; The processing module is further configured to perform correction processing on the ordinary convolution module according to the back-projection tensor error to obtain a rotatable convolution module, and perform correction processing on the traditional linear difference module according to the discrete sampling error to obtain a two-dimensional local continuous representation difference module, and integrate the rotatable convolution module and the two-dimensional local continuous representation difference module to obtain an intermediate network model; The sinusoidal scan image is input into the intermediate network model to obtain a corrected scan image, a structural similarity index loss value between the corrected scan image and the first scan image is calculated, and the intermediate network model is corrected according to the structural similarity index loss value to obtain a target network model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.

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

Images