CN120689447A

CN120689447A - A training method, device, equipment and storage medium for image correction model

Info

Publication number: CN120689447A
Application number: CN202510765610.9A
Authority: CN
Inventors: 林卉; 谢琦; 曾栋; 毛泽睿; 马建华; 孟德宇
Original assignee: Xian Jiaotong University
Current assignee: Xian Jiaotong University
Priority date: 2025-06-10
Filing date: 2025-06-10
Publication date: 2025-09-23

Abstract

The application discloses a training method, device, equipment and storage medium of an image correction model, which comprises the steps of obtaining a first scanning image, obtaining a plurality of scanning sinusoidal images in one-to-one correspondence through generation of the first scanning image without corresponding scanning sinusoidal images, carrying out back projection and summation on the scanning sinusoidal images to obtain a sparse angle scanning image, comparing the sparse angle scanning image with a full angle scanning image to obtain a back projection tensor error and a discrete sampling error, constructing an initial network model comprising a common convolution module and a traditional linear difference module, obtaining an intermediate network model according to the error correction modules, inputting the sinusoidal scanning images into the intermediate model to obtain a corrected scanning image, calculating a structural similarity index loss value of the corrected sinusoidal image and the first scanning image, and correcting the intermediate model according to the loss value to obtain a target network model. The method can solve the problems of projection tensor error and discrete sampling in the sparse view scanning image, improve the accuracy and detail holding capacity of the reconstructed image, and enhance the adaptability of the model to sparse sampling.

Description

Training method, device, equipment and storage medium for image correction model

Technical Field

The present application relates to the field of image recognition technologies, and in particular, to a training method, apparatus, device, and storage medium for an image correction model.

Background

With the development of medical imaging technology, sparse view scanning imaging is widely applied due to the advantages of low radiation dose and the like. However, the existing traditional filtering back projection method is difficult to effectively inhibit errors under a sparse view angle, and the partial depth learning method lacks of targeted modeling of rotation inconsistency and detector distance dimension continuity, so that the quality of a reconstructed image is unstable. Especially in an extremely sparse sampling scene, the angle alignment capability of the prior art on the back projection tensor is insufficient, the interpolation precision of discrete sampling points is low, and the accuracy of image reconstruction is further reduced.

Therefore, how to efficiently process the back projection tensor error and the discrete sampling error and improve the accuracy and the robustness of the scanned image reconstruction under the sparse view angle is a problem to be solved.

Disclosure of Invention

In view of this, the training method, device, equipment and storage medium for an image correction model provided by the embodiments of the present application can efficiently process the back projection tensor error and the discrete sampling error, and improve the accuracy and robustness of the scanned image reconstruction under a sparse view angle. The training method, device and equipment for the image correction model and the storage medium provided by the embodiment of the application are realized as follows:

the training method of the image correction model provided by the embodiment of the application comprises the following steps:

Acquiring a plurality of first scanning images, judging whether each first scanning image has a corresponding scanning sinusoidal image, and generating a scanning sinusoidal image corresponding to the current first scanning image when the current first scanning image does not have the corresponding scanning sinusoidal image, so as to obtain a plurality of scanning sinusoidal images corresponding to the plurality of first scanning images one by one;

acquiring all scanning sinusoidal images, and respectively carrying out back projection and summation on each scanning sinusoidal image to obtain a sparse angle scanning image;

Acquiring a full-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, wherein the initial network model comprises a common convolution module and a traditional linear difference module;

Correcting the common convolution module according to the back projection tensor error to obtain a rotatable convolution module, correcting the traditional linear difference module according to the discrete sampling error to obtain a two-dimensional local continuous representation difference module, and integrating the rotatable convolution module and the two-dimensional local continuous representation difference module to obtain an intermediate network model;

and inputting the sinusoidal scanning image into an intermediate network 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 network model according to the structural similarity index loss value to obtain a target network model.

In some embodiments, the acquiring all the scanned sinusoidal images, performing back projection and summation on each scanned sinusoidal image to obtain a sparse angle scanned image, and further includes:

and filtering each scanning sinusoidal image to obtain a filtered scanning sinusoidal image.

In some embodiments, the acquiring all the scanned sinusoidal images, performing back projection and summation on each scanned sinusoidal image to obtain a sparse angle scanned image, includes:

Acquiring all scanning sinusoidal images, and carrying out back projection and summation on each scanning sinusoidal image according to a formula (1) to obtain a sparse angle scanning image;

Wherein, the For a filtered scanned sinusoidal image, q is the filter kernel,For convolution operation, p _θ is a scanned sinusoidal image, θ is the back projection angle, H _θ [ i, j ] is the back projection image at the θ viewing angle, [ i, j ] is the coordinates of the back projection image,As a back projection function, x _i、y_j is the spatial coordinate corresponding to the sparse angle scan image, where i, j are the pixel coordinate indices in the image,For the sparse angle scan image, M is the number of sparse angle scan images, θ _m is the angle of the mth sparse angle scan image.

In some embodiments, the correcting the common convolution module according to the backprojection tensor error to obtain a rotatable convolution module, and correcting the conventional linear difference module according to the discrete sampling error to obtain a two-dimensional locally continuous representation difference module, and combining the rotatable convolution module and the two-dimensional locally continuous representation difference module to obtain an intermediate network model, including:

Analyzing and processing the back projection tensor error to obtain the distribution characteristics of rotation inconsistency of the sparse angle scanning image;

Correcting the learnable combination coefficient and the basis function parameter in the common convolution module according to the distribution characteristics to obtain a corrected 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;

And carrying out integrated processing on the corrected rotatable convolution module and the corrected two-dimensional local continuous representation difference module to obtain an intermediate network model.

In some embodiments, the correcting the conventional linear difference module according to the discrete sampling error to obtain a corrected two-dimensional locally continuous representation difference module includes:

analyzing and processing the discrete sampling errors to obtain an undersampled region;

And correcting a basis function combination mode in the traditional linear difference module according to the undersampled region to obtain a corrected two-dimensional local continuous representation difference module.

In some embodiments, the acquiring a plurality of first scan images, determining whether each first scan image has a corresponding scan sinusoidal image, generating a scan sinusoidal image corresponding to a current first scan image when the current first scan image does not have a corresponding scan sinusoidal image, and after obtaining a plurality of scan sinusoidal images corresponding to the plurality of first scan images one-to-one, further includes:

and carrying out data quality optimization processing on the plurality of scanning sinusoidal images to obtain the processed scanning sinusoidal images.

In some embodiments, the number of sampling angles of the full angle scan image is 1152 and the number of sampling angles of the sparse angle scan image is 72.

The training device for the image correction model provided by the embodiment of the application comprises the following components:

The acquisition module is used for acquiring a plurality of first scanning images, judging whether each first scanning image has a corresponding scanning sinusoidal image, and generating a scanning sinusoidal image corresponding to the current first scanning image when the current first scanning image does not have the corresponding scanning sinusoidal image, so as to obtain a plurality of scanning sinusoidal images corresponding to the plurality of first scanning images one by one;

The processing module is used for acquiring all scanning sinusoidal images, and respectively carrying out back projection and summation on each scanning sinusoidal image to obtain a sparse angle scanning image;

The processing module is further used for acquiring a full-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;

The construction module is used for constructing an initial network model, and the initial network model comprises a common convolution module and a traditional linear difference module;

The processing module is further configured to perform correction processing on the normal convolution module according to the backprojection tensor error to obtain a rotatable convolution module, 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 perform integrated processing on the rotatable convolution module and the two-dimensional local continuous representation difference module to obtain an intermediate network model;

The computer device provided by the embodiment of the application comprises a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and the processor realizes the method of the embodiment of the application when executing the program.

The computer readable storage medium provided by the embodiment of the present application stores a computer program thereon, which when executed by a processor implements the method provided by the embodiment of the present application.

The training method, the training device, the computer equipment and the computer readable storage medium for the image correction model are provided by the embodiment of the application, through acquiring a first scanning image, generating a plurality of scanning sinusoidal images corresponding to the first scanning image without the corresponding scanning sinusoidal image one by one, carrying out back projection and summation on the scanning sinusoidal images to obtain a sparse angle scanning image, comparing the sparse angle scanning image with a full angle scanning image to obtain a back projection tensor error and a discrete sampling error, constructing an initial network model comprising a common convolution module and a traditional linear difference module, obtaining an intermediate network model according to the two error correction modules, inputting the sinusoidal scanning images into the intermediate model to obtain a corrected scanning image, calculating a structural similarity index loss value of the corrected sinusoidal scanning image and the first scanning image, and correcting the intermediate model according to the loss value to obtain a target network model. The method can solve the problems of projection tensor error and discrete sampling in the sparse view scanning image, improve the accuracy and detail holding capacity of the reconstructed image, enhance the adaptability of the model to sparse sampling and solve the technical problems in the background technology.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following description will briefly explain the embodiments of the present application or the drawings used in the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic implementation flow chart of a training method of an image correction model according to an embodiment of the present application;

FIG. 2 is a schematic diagram of a model correction implementation flow in a training method of an image correction model according to an embodiment of the present application;

Fig. 3 is a training device for an image correction model according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. It will be apparent that the described embodiments are some, but not all, embodiments of the application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

Some of the techniques involved in the embodiments of the present application are described below to aid understanding, and they should be considered as merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the application. Also, for the sake of clarity and conciseness, descriptions of well-known functions and constructions are omitted in the following description.

Fig. 1 is a schematic implementation flow chart of a training method of an image correction model according to an embodiment of the present application, including steps 101 to 106. Wherein fig. 1 is only one execution order shown in the embodiment of the present application, and does not represent the only execution order of a training method of an image correction model, and the steps shown in fig. 1 may be executed in parallel or upside down in case that the final result can be achieved.

Step 101, acquiring a plurality of first scanning images, judging whether each first scanning image has a corresponding scanning sinusoidal image, and generating a scanning sinusoidal image corresponding to the current first scanning image when the current first scanning image does not have the corresponding scanning sinusoidal image, so as to obtain a plurality of scanning sinusoidal images corresponding to the plurality of first scanning images one by one.

In an embodiment of the application, a plurality of medical CT (Computed Tomography ) first scan images are acquired, and a determination is made for each image as to whether a corresponding scan sinogram (i.e., X-ray projection data) already exists. If not, generating a scanning sine image corresponding to the image through Radon transformation. Specifically, using the physical principle of X-ray attenuation, HU values of a CT image (henry's unit, a standardized unit for quantifying the attenuation degree of a tissue to a ray in a medical CT image) are converted into linear attenuation value distribution by a formula (1), projection data are generated by line integration calculation, and finally a scanning sinusoidal image corresponding to each first scanning image one by one is obtained.

V_μ＝V_CT*0.0192/1000+0.0192(1)

Wherein V _μ is the attenuation value, and V _CT is the HU value corresponding to the first scan image.

Step 102, acquiring all scanning sinusoidal images, and respectively carrying out back projection and summation on each scanning sinusoidal image to obtain a sparse angle scanning image.

In the embodiment of the application, after all generated scanning sinusoidal images are acquired, filtering back projection processing is performed on each sinusoidal image. Specifically, a filter such as Ram-Lak is used for filtering a sinusoidal image, and then the filtered projection data is mapped to an image space through a back projection algorithm to generate a back projection image with a single view angle. And then, carrying out summation processing on the back projection images of all the view angles to obtain CT reconstructed images (namely sparse angle scanning images) under sparse angles.

And 103, acquiring a full-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.

In the embodiment of the present application, a full-angle scan image (e.g., a reconstructed CT image with a sampling geometry of 1152 angles) of the same scan object is obtained, and compared with the sparse angle scan image obtained in step 102 pixel by pixel. Specifically, the difference of pixel values of the two is calculated to obtain a back projection tensor error of the back projection tensor in the angle dimension and a discrete sampling error of the detector and the distance dimension caused by sparse sampling.

Step 104, constructing an initial network model, wherein the initial network model comprises a common convolution module and a traditional linear difference module.

In the embodiment of the application, an initial network model comprising two large core modules is constructed, wherein the initial network model comprises a common convolution module and a traditional linear difference module, and the initial network model comprises the following components:

The common convolution module consists of a plurality of convolution layers, an activation function (such as ReLU) and a normalization layer (such as BatchNorm), and is responsible for extracting global features of an input image (such as an optimized scanning sinusoidal image), covering information such as contours, textures and the like, and providing basic feature support for subsequent processing.

And the traditional linear interpolation module adopts a linear interpolation algorithm to fill data in sparse sampling areas (such as detector edges and signal sparse areas of low-dose scanning) in the image. And the missing pixels are estimated by calculating the adjacent pixel values, so that the image continuity is enhanced, and the smoothness and the integrity of local details are improved.

And 105, carrying out correction processing on the common convolution module according to the back projection tensor error to obtain a rotatable convolution module, carrying out 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 carrying out integrated processing on the rotatable convolution module and the two-dimensional local continuous representation difference module to obtain an intermediate network model.

In the embodiment of the application, a rotatable convolution module introduces a weighted convolution structure based on distance guidance aiming at rotation inconsistency of back projection tensors of different view angles, and the mathematical expression is as follows: where D _θ is the distance tensor from each pixel point to the virtual detector, ρ ₁ and ρ ₂ are convolution kernels of different smoothness, In order to filter the backprojection tensor estimation,In order to convolve the operation symbols,And for the original back projection tensor under the angle theta, the output of the two convolution kernels is adaptively combined through the distance tensor, so that the feature extraction capability of different distance areas is enhanced.

The two-dimensional local continuous representation difference module is used for modeling continuity of the detector and the distance dimension, a continuous projection model is constructed through a bicubic interpolation basis function, a basis function expression is a piecewise function, different polynomial forms are adopted in different intervals, a convolution response range is limited by combining a radial mask function, and interpolation precision among discrete sampling points is improved.

The rotatable convolution module corrects the distribution characteristics according to the backprojection tensor error, and the combination coefficient and the basis function parameter can be learned in the adjustment module. Specifically, by analyzing the rotation error distribution of the back projection tensor under different angles, the combination coefficient of the rotation convolution kernel is optimized (namely, the convolution kernel can rotate and align the tensor under different angles through training), meanwhile, the shape parameters of the bicubic basis functions are finely adjusted, and the robustness of the module to angle change is enhanced.

The two-dimensional local continuous representation difference module corrects the undersampled region of the identification detector based on the discrete sampling error and the distance dimension, and an adaptive weight mechanism is introduced by adjusting the combination mode of the basis functions, so that the module can dynamically adjust the interpolation weight according to the sampling density and inhibit artifacts caused by undersampling.

And 106, inputting the sinusoidal scanning image into the intermediate network model to obtain a corrected scanning image, calculating the structural similarity index loss value of the corrected scanning image and the first scanning image, and correcting the intermediate network model according to the structural similarity index loss value to obtain the target network model.

In the embodiment of the application, the scanning sinusoidal image is input into the corrected intermediate network model to obtain a corrected scanning image. A Structural Similarity Index (SSIM) loss value of the modified image and the first scanned image is calculated, and a loss function is constructed by evaluating differences in brightness, contrast, and structural information of the images. And (3) performing iterative optimization (such as adjustment of convolution kernel weight, basis function combination coefficient and the like) on parameters of the intermediate network model according to the loss value by using a back propagation algorithm, and finally obtaining the target network model capable of effectively inhibiting sparse visual angle artifacts.

According to the embodiment of the application, the integrity of training data is ensured by generating the missing scanning sinusoidal image, the error is obtained by utilizing the contrast of the back projection and the full-angle image, the adaptability of the geometric transformation and the local characteristic representation capability of the model are corrected in a targeted manner by combining the rotatable module and the two-dimensional local continuous representation difference module, and finally the image structure is optimized through the structural similarity index, so that the training precision and the robustness of the image correction model are effectively improved.

In some embodiments, all the scanning sinusoidal images are obtained, back projection and summation processing are respectively carried out on each scanning sinusoidal image to obtain sparse angle scanning images, and filtering processing is carried out on each scanning sinusoidal image to obtain filtered scanning sinusoidal images.

Specifically, before the back projection processing is performed on the scanned sinusoidal images, the filtering processing is performed on each scanned sinusoidal image. Specifically, a frequency domain filtering method is adopted, and a Ram-Lak filter (or other known CT reconstruction filters) is selected to process the sinusoidal image. The filtering process is to carry out convolution operation on the filter and the scanned sinusoidal image in the frequency domain or the time domain, so that high-frequency components in projection data are enhanced, information loss caused by sparse sampling is compensated, and the accuracy of subsequent back projection reconstruction is improved.

The embodiment of the application further optimizes the reconstruction quality of the sparse angle scanning image by adding the filtering processing step of the scanning sinusoidal image. The method can effectively compensate low-frequency distortion caused by sparse sampling, and provides more accurate projection data for back projection reconstruction, so that the retention capacity of the model to image details and the noise resistance are improved.

In some embodiments, acquiring all the scanning sinusoidal images, and performing back projection and summation on each scanning sinusoidal image to obtain a sparse angle scanning image.

Specifically, after all the scanned sinusoidal images are acquired, each image is subjected to a filtering process to enhance the high frequency information. Specifically, a convolution operation is performed by using a preset filter kernel and a scanning sinusoidal image. The filtering process can be expressed as performing convolution operation on the filtering kernel and the scanned sinusoidal image in the time domain or the frequency domain, wherein the frequency domain filtering requires converting the image into the frequency domain through Fourier transformation, multiplying the image with the filter frequency domain expression, and then inversely converting the image back into the time domain.

And performing a back projection operation on the filtered scanned sinusoidal image, and mapping projection data to an image space. The method comprises the specific steps of carrying out distributed accumulation on filtered projection data along a projection line of each back projection angle theta by utilizing a back projection function. The mathematical expression of the back projection function needs to satisfy the physical process of X-ray attenuation, and each point on the projection line is mapped to the corresponding coordinate of the image space according to the geometric position. For example, for a coordinate point (x _i,y_j) in image space, whose projection line position at an angle θ is x _icosθ+y_j sin θ, the backprojection function assigns a value to the coordinate point (x _i,y_j) according to the projection value of the position, generating a backprojection image H _θ [ i, j ] at a single viewing angle.

And carrying out summation treatment on the back projection images of all the sparse views to obtain a final sparse angle scanning 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 angle of the full-angle scanned image is 1152, and the angle of each view angle isAfter the back projection images of each view angle are accumulated, the normalized coefficient is multipliedTo compensate for the viewing angle sampling density. The normalization operation ensures that the intensity of the reconstructed image is consistent with that of the full-angle scan, and avoids brightness deviation caused by sparse view angles. The mathematical expression of the summation process is that pixel-level accumulation is carried out on the back projection images of all view angles, and finally the obtained sparse angle scanning image can synthesize projection information of a plurality of sparse view angles, so that artifacts of single view angle reconstruction are reduced.

The filter processing formula: Corresponding convolution operations in which For the filtered scanned sinusoidal image, q is the filter kernel, p _θ is the original scanned sinusoidal image,Representing a convolution operation;

Back projection formula: corresponding projection line mapping process, wherein Is a back projection function, (x _i,y_j) is an image space coordinate, θ is a projection angle, and H _θ [ i, j ] is a back projection image at a θ viewing angle;

the sum formula: corresponding to a normalized accumulation operation, where M is the sparse view number, For a sparse angle scan image, θ _m is the angle of the mth view, and the final sparse angle scan image is generated by summing and normalizing.

According to the embodiment of the application, the filtering, back projection and summation processes of the scanned sinusoidal image are mathematically processed, the filtering process utilizes the known CT reconstruction filter to enhance the high-frequency component of projection data, the back projection process maps the projection data to the image space based on the Radon transformation inverse operation, the summation step compensates the sparse view angle sampling deviation through normalization accumulation, the feasibility and accuracy of the technical scheme are ensured, and the accuracy and reliability of the reconstruction of the sparse view angle CT image are effectively improved.

On the basis of fig. 1, the present application further provides a schematic implementation flow diagram of model correction in the training method of the image correction model, as shown in fig. 2, including steps 201 to 204:

step 201, analyzing and processing the back projection tensor error to obtain the distribution characteristics of rotation inconsistency of the sparse angle scanning image.

In the embodiment of the application, the back projection tensor error is quantitatively analyzed, the sparse angle scanning image and the full angle scanning image are compared pixel by pixel, the pixel value difference of the sparse angle scanning image and the full angle scanning image is calculated, and the region with obvious error is positioned. And obtaining the spatial characteristics of the rotation inconsistency by counting the error distribution of the back projection tensor. For example, the law of error variation with distance from the virtual detector is analyzed to determine which of the back projection tensors of the location intervals have significant rotational bias.

And 202, correcting the learnable combination coefficient and the basis function parameter in the common convolution module according to the distribution characteristics to obtain a corrected rotatable convolution module.

In an embodiment of the present application, the convolution kernel of the rotatable convolution module is formed by a set of convolution bases and a learnable combination coefficient as a formulaShown, where the combining coefficients w _n control the weights of the various convolution substrates. According to the distribution of rotation inconsistency, the value of w _n is optimized through training, so that the convolution kernel can perform more accurate rotation alignment on the back projection tensor of the error sensitive area, ρ (z, θ) is the convolution kernel, σ _n is a basis function, z is a space coordinate for describing the position of a pixel in an image, θ is a scanning angle for describing the rotation direction of a detector, N is the number of convolution substrates, N is the sequence number of the basis function, T _θ is an angle rotation matrix for performing rotation transformation on the space coordinate, wherein,

And in a position interval with larger error, the combination coefficient of the corresponding convolution substrate is increased, and the adjustment capability of the convolution check in the area is enhanced. For example, in the image edge region where the errors are concentrated and distributed, by learning a specific convolution base combination coefficient, the model can better adapt the convolution kernel to the local error feature and maintain the consistency of adjustment under different angles by learning a specific base combination coefficient, thereby realizing rotatable error mitigation effect.

And 203, correcting the traditional linear difference module according to the discrete sampling error to obtain a corrected two-dimensional local continuous representation difference module.

In the embodiment of the application, for discrete sampling errors, a sampling missing region (such as a projection data discontinuous region caused by sparse sampling) in the dimension of the detector and the dimension of the distance is determined by analyzing the distribution of the discrete sampling errors.

The module utilizes bicubic interpolation basis functions and radial mask functions to construct a continuous projection model, and adjusts the combination weight of the basis functions according to the distribution of undersampled areas. For example, in the area with low sampling density, the weight of the basic function adjacent to the sampling point is increased, the continuity is enhanced by local interpolation, and in the area with high sampling density, the default combination mode of the basic function is maintained, so that the overfitting is avoided.

And introducing a self-adaptive weight mechanism based on sampling density, and dynamically adjusting the contribution degree of each sampling point to the reconstruction result. For example, for sparse sampling detector positions, higher interpolation weights are given to adjacent sampling points, artifacts caused by insufficient sampling are restrained, and for dense sampling areas, uniform weights are adopted, so that detail retention is ensured.

And 204, carrying out integrated processing on the corrected rotatable convolution module and the corrected two-dimensional local continuous representation difference module to obtain an intermediate network model.

In the embodiment of the application, the corrected rotatable convolution module is integrated with the two-dimensional local continuous representation difference module, the rotatable convolution module is used for carrying out angle alignment treatment on the back projection tensor, the rotation inconsistency among different visual angles is eliminated, the treated tensor is input into the two-dimensional local continuous representation difference module, the continuity modeling is carried out on the detector and the distance dimension, and the discrete sampling error is compensated.

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

In the integration process, the matching of the input and output dimensions of the two modules is ensured, for example, the output tensor dimension of the rotatable convolution module is consistent with the input requirement of the two-dimensional local continuous representation difference module, and finally an intermediate network model capable of simultaneously processing the rotation error and the discrete sampling error is formed.

According to the embodiment of the application, the distribution characteristics of rotation inconsistency are obtained through analysis of the back projection tensor error, the problem of geometrical transformation inconsistency caused by angle deficiency in sparse angle scanning can be positioned in a targeted manner, and furthermore, the adaptability of a rotatable convolution module to image transformation under different angles is enhanced through correcting the learnable combination coefficient and the basis function parameter of the module, the constant change characteristics of convolution operation in a rotation space are ensured, and reconstruction distortion caused by angle sparsity is effectively restrained. The two-dimensional local continuous representation difference module is corrected aiming at discrete sampling errors, so that the loss of image details caused by discontinuous sampling points can be compensated, and the continuous modeling capability of a model on discrete data is improved by optimizing a parameterization mode of local continuous representation, so that the spatial structure information of an image can be accurately captured under a sparse angle.

In some embodiments, the traditional linear difference module is modified according to the discrete sampling error to obtain a modified two-dimensional local continuous representation difference module, which comprises analyzing and processing the discrete sampling error to obtain an undersampled region.

Specifically, quantitative analysis is carried out on discrete sampling errors, a sparse angle scanning image is compared with a full angle scanning image, and sampling point error distribution in the dimension and the distance dimension of the detector is calculated. By counting the spatial distribution of the pixel value differences, a reconstruction distortion region (namely, an undersampled region) caused by the sparse sampling points is identified. For example, at edge regions of the detector array or at locations further from the imaging center, significant discrete sampling errors often occur due to low sampling density, forming streak artifacts or structural blurring.

Further, a basis function combination mode in the traditional linear difference module is modified according to the undersampled region, and a modified two-dimensional local continuous representation difference module is obtained.

Specifically, according to the positioned undersampled region, the basis function combination mode of the two-dimensional local continuous representation difference module is subjected to targeted adjustment, weighted combination optimization of bicubic basis functions is adopted, and a continuous projection model is constructed through linear combination of the basis functions. In the undersampled region, the weight of the basis function corresponding to the adjacent sampling point is increased, and the local interpolation precision is enhanced. For example, for regions of larger sampling intervals in the detector dimension, the basis function weights of adjacent sampling points are increased by 20% -30%, and the lack of sampling is compensated for by enhancing the interpolation coverage.

The effective range of the mask is dynamically extended based on the spatial location of the undersampled region. 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, and the continuity modeling capability between discrete points is improved.

The combination coefficients of the basis functions are optimized through a training process, so that the module automatically activates more proper basis function combinations in the undersampled area. For example, by using a gradient descent algorithm, with the reconstruction error of the undersampled region as an optimization target, the linear combination coefficient of the basis functions is iteratively adjusted, so that the interpolation result of the combined basis functions in the region is closer to the true value of the full-angle scanning.

When the corrected module processes the undersampled region, weighted bicubic interpolation is adopted for sparse sampling points in the dimension of the detector, the weight is determined by the distance from the sampling point to the target point and the error distribution, and the sampling point with the closer distance or larger error has higher weight.

In the distance dimension, the effective acting range of the base function is expanded by combining the dynamically adjusted radial mask function, so that the area far away from the sampling point can still obtain a more accurate reconstruction value through adjacent multi-point interpolation.

The embodiment of the application identifies the undersampled region by analyzing the discrete sampling error, so that the correction operation has definite pertinence. Excessive intervention to a normal sampling area can be avoided, and the basis function combination is dynamically adjusted only in the error significant area, so that the local detail precision of the reconstructed image is maximally improved while the calculation efficiency is ensured.

In some embodiments, a plurality of first scanning images are obtained, whether each first scanning image has a corresponding scanning sinusoidal image is judged, when the current first scanning image does not have the corresponding scanning sinusoidal image, the scanning sinusoidal image corresponding to the current first scanning image is generated, and after the plurality of scanning sinusoidal images corresponding to the plurality of first scanning images one by one are obtained, data quality optimization processing is carried out on the plurality of scanning sinusoidal images, and the processed scanning sinusoidal images are obtained.

Specifically, after acquiring a plurality of scanning sinusoidal images corresponding to the first scanning image, performing data quality optimization processing on each scanning sinusoidal image, including traversing each pixel point in the scanning sinusoidal image, and calculating statistical differences (such as mean value and standard deviation) between the scanning sinusoidal image and other pixel points in the local neighborhood. For outliers that deviate from the statistical threshold (e.g., outliers due to detector noise), a median or weighted average within the local neighborhood is used for substitution. For example, if the value of a pixel exceeds 3 times the standard deviation of the average value of the neighboring pixels, the pixel is determined to be an abnormal value and corrected.

And comparing the scanned sinusoidal images at different angles, and detecting and eliminating repeated data caused by errors of scanning equipment or overlapping angles. The specific method is to calculate the similarity (such as structural similarity index SSIM) between the images, if the similarity of the two images exceeds a preset threshold (such as 0.95), one image with higher quality is reserved, and duplicate images are deleted.

And (3) for the missing region (such as data loss caused by detector faults) existing in the scanned sinusoidal image, performing complementation by adopting a bicubic interpolation method, and performing interpolation calculation by combining the values of surrounding effective pixel points to generate an estimated value of the missing region.

The embodiment of the application lays a high-precision reconstruction foundation on the physical signal level through multidimensional optimization of data quality, and realizes three-level data enhancement of noise suppression, redundancy elimination and information complementation. Compared with the traditional non-optimized scanning sinusoidal image, the signal to noise ratio can be improved.

Although the application provides method operational steps as described in the examples or flowcharts, more or fewer operational steps may be included based on conventional or non-inventive labor. The order of steps recited in the present embodiment is only one way of performing the steps in a plurality of steps, and does not represent a unique order of execution. When implemented by an actual device or client product, the method of the present embodiment or the accompanying drawings may be performed sequentially or in parallel (e.g., in a parallel processor or a multithreaded environment).

As shown in fig. 3, the embodiment of the application further provides a training device 300 for an image correction model. The device comprises:

The acquiring module 301 is configured to acquire a plurality of first scan images, determine whether each first scan image has a corresponding scan sinusoidal image, and generate a scan sinusoidal image corresponding to the current first scan image when the current first scan image does not have a corresponding scan sinusoidal image, so as to obtain a plurality of scan sinusoidal images corresponding to the plurality of first scan images one to one;

the processing module 302 is configured to obtain all the scanned sinusoidal images, and perform back projection and summation processing on each scanned sinusoidal image to obtain a sparse angle scanned image;

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

The construction module 303 is configured to construct an initial network model, where the initial network model includes a general convolution module and a traditional linear difference module;

The processing module 302 is further configured to perform correction processing on the normal convolution module according to the backprojection tensor error to obtain a rotatable convolution module, 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 perform integrated processing on the rotatable convolution module and the two-dimensional local continuous representation difference module to obtain an intermediate network model;

And inputting the sinusoidal scanning image into the intermediate network model to obtain a corrected scanning image, calculating the structural similarity index loss value of the corrected scanning image and the first scanning image, and correcting the intermediate network model according to the structural similarity index loss value to obtain the target network model.

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

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

In some embodiments, the processing module 302 is further configured to analyze the backprojection tensor error to obtain a distribution characteristic of rotation inconsistency of the sparse angle scan image;

the processing module 302 is further configured to correct the learnable combination coefficient and the basis function parameter in the common convolution module according to the distribution characteristic, so as to obtain a corrected rotatable convolution module;

the processing module 302 is further configured to modify the conventional linear difference module according to the discrete sampling error, so as to obtain a modified two-dimensional local continuous representation difference module;

the processing module 302 is further configured to perform integrated processing on the modified rotatable convolution module and the modified two-dimensional local continuous representation difference module, so as to obtain an intermediate network model.

In some embodiments, the processing module 302 is further configured to analyze and process the discrete sampling error to obtain an undersampled region;

the processing module 302 is further configured to modify a basis function combination manner in the conventional linear difference module according to the undersampled region, so as to obtain a modified two-dimensional locally continuous representation difference module.

The processing module 302 is further configured to perform data quality optimization processing on the plurality of scanned sinusoidal images, so as to obtain a processed scanned sinusoidal image.

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

The apparatus or module set forth in the embodiments of the application may be implemented in particular by a computer chip or entity, or by a product having a certain function. For convenience of description, the above devices are described as being functionally divided into various modules, respectively. The functions of each module may be implemented in the same piece or pieces of software and/or hardware when implementing the embodiments of the present application. Of course, a module that implements a certain function may be implemented by a plurality of sub-modules or a combination of sub-units.

The methods, apparatus or modules described herein may be implemented in computer readable program code means and in any suitable manner, e.g., the controller may take the form of, for example, a microprocessor or processor and a computer readable medium storing computer readable program code (e.g., software or firmware) executable by the (micro) processor, logic gates, switches, application-specific integrated circuits (english: application SPECIFIC INTEGRATED Circuit; ASIC), programmable logic controllers and embedded microcontrollers, examples of which include, but are not limited to, ARC 625D, atmel AT91SAM, microchip PIC18F26K20 and Silicone Labs C8051F320, and the memory controller may also be implemented as part of the control logic of the memory. Those skilled in the art will also appreciate that, in addition to implementing the controller in a pure computer readable program code, it is well possible to implement the same functionality by logically programming the method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. Such a controller can be regarded as a hardware component, and means for implementing various functions included therein can also be regarded as a structure within the hardware component. Or even means for achieving the various functions may be regarded as either software modules implementing the methods or structures within hardware components.

The embodiment of the application also provides equipment, which comprises a processor, a memory for storing executable instructions of the processor, and a method for realizing the embodiment of the application when the processor executes the executable instructions.

Embodiments of the present application also provide a non-transitory computer readable storage medium having stored thereon a computer program or instructions which, when executed, cause a method as described in embodiments of the present application to be implemented.

In addition, each functional module in the embodiments of the present invention may be integrated into one processing module, each module may exist alone, or two or more modules may be integrated into one module.

The storage medium includes, but is not limited to, a random access Memory (hereinafter referred to as RAM), a Read-Only Memory (hereinafter referred to as ROM), a Cache Memory (hereinafter referred to as Cache), a hard disk (hereinafter referred to as HARD DISKDRIVE; hereinafter referred to as HDD), or a Memory card (hereinafter referred to as Memory card). The memory may be used to store computer program instructions.

From the above description of embodiments, it will be apparent to those skilled in the art that the present application may be implemented in software plus necessary hardware. Based on such understanding, the technical solution of the present application may be embodied essentially or in a part contributing to the prior art in the form of a software product or may be embodied in the implementation of data migration. The computer software product may be stored on a storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., comprising instructions for causing a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to perform the methods described in the various embodiments or portions of the embodiments of the application.

In this specification, each embodiment is described in a progressive manner, and the same or similar parts of each embodiment are referred to each other, and each embodiment is mainly described as a difference from other embodiments. All or portions of the present application are operational with numerous general purpose or special purpose computer system environments or configurations. Such as a personal computer, a server computer, a hand-held or portable device, a tablet device, a mobile communication terminal, a multiprocessor system, a microprocessor-based system, a programmable electronic device, a network PC, a minicomputer, a mainframe computer, a distributed computing environment that includes any of the above systems or devices, and the like.

The foregoing embodiments are only for illustrating the technical solution of the present application, but not for limiting the same, and although the present application has been described in detail with reference to the foregoing embodiments, it will be understood by those skilled in the art that modifications may be made to the technical solution described in the foregoing embodiments, or equivalents may be substituted for some or all of the technical features thereof, without departing from the spirit of the corresponding technical solution from the scope of the technical solution of the present application.

Claims

1. A method of training an image correction model, comprising:

2. The method of claim 1, wherein the acquiring all the scanned sinusoidal images, respectively performing back projection and summation on each scanned sinusoidal image to obtain a sparse angle scanned image, further comprises:

3. The method according to claim 2, wherein the acquiring all the scanned sinusoidal images, respectively performing back projection and summation on each scanned sinusoidal image to obtain a sparse angle scanned image, includes:

4. The method according to claim 1, wherein the correcting the normal convolution module according to the backprojection tensor error to obtain a rotatable convolution module, and correcting the conventional linear difference module according to the discrete sampling error to obtain a two-dimensional locally continuous representation difference module, and combining the rotatable convolution module and the two-dimensional locally continuous representation difference module to obtain an intermediate network model includes:

5. The method of claim 4, wherein said modifying said conventional linear difference module based on said discrete sampling error results in a modified two-dimensional locally continuous representation difference module comprising:

6. The method according to claim 1, wherein the steps of obtaining a plurality of first scan images, determining whether each first scan image has a corresponding scan sinusoidal image, generating a scan sinusoidal image corresponding to a current first scan image when the current first scan image does not have a corresponding scan sinusoidal image, and obtaining a plurality of scan sinusoidal images corresponding to the plurality of first scan images one-to-one, further comprise:

7. The method of claim 1, wherein the number of sampling angles of the full angle scan image is 1152 and the number of sampling angles of the sparse angle scan image is 72.

8. An image correction model training apparatus, comprising:

9. A computer device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that the processor implements the steps of the method of any of claims 1 to 7 when the program is executed.

10. A computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the method according to any one of claims 1 to 7.