Multi-source straight-line ct reconstruction method and system based on projection fusion and depth unfolding network

Abstract

This invention relates to the field of X-ray computed tomography imaging technology, and particularly to a multi-source linear CT reconstruction method and system based on projection fusion and depth unfolding networks. First, using at least three X-ray sources in conjunction with a single area array detector, the object under examination is controlled to translate linearly, and multiple sets of projection data sequences are acquired time-divisionally. Second, a mapping model is constructed based on spatial geometric relationships, fusing and rearranging the projection data into a two-dimensional sine graph with missing angles. Next, a cascaded network of sine graph completion and image reconstruction is constructed. First, a completion submodule fills in the missing data in the sine graph while retaining the true measurement values. Then, an intelligent reconstruction submodule with a depth unfolding architecture containing multiple iteration stages, combined with a weight-sharing backprojection operator and a residual regularization module, updates the image. Finally, an end-to-end joint training is performed using a composite loss function of the sine graph domain and the image domain. This invention achieves high-fidelity, artifact-free, and rapid reconstruction of CT tomographic images, balancing imaging quality and efficiency.

Description

Technical Field

[0001] This invention relates to the field of X-ray computed tomography (CT) imaging technology, and in particular to a multi-source linear CT reconstruction method and system based on projection fusion and depth unfolding network, which can be widely used in scenarios such as industrial non-destructive testing and security inspection imaging where both scanning speed and image quality are highly demanding. Background Technology

[0002] Traditional circular trajectory CT systems require mechanical rotation, resulting in slow scanning speeds and difficulty meeting the demands of industrial online inspection. Linear trajectory CT, on the other hand, replaces rotation with object translation, significantly improving efficiency. However, due to limitations in equipment structure, single-source linear scanning can only acquire projections at a limited angle, leading to poor reconstruction quality.

[0003] To improve data coverage, existing technologies propose multi-X-ray source linear array schemes. However, even so, due to the fixed viewing angle of each X-ray source and the object only translating in one direction, the synthesized projection still exhibits asymmetric, discontinuous, and large-scale missing characteristics in the angular domain (e.g., missing central angles, sparse edges). Such data cannot be directly used for traditional filtered back projection (FBP) reconstruction.

[0004] Although existing research has attempted to incorporate deep learning, such as using U-Net to remove artifacts from reconstructed images or employing generative networks to complete sine waves, these methods typically separate completion from reconstruction, fail to consider the geometric correlation of projected data under straight-line trajectories, and lack image quality feedback during the completion process, which can easily introduce physically inconsistent false structures.

[0005] Furthermore, although some patents (such as "A Deep Learning Hilbert Inverse Transform Reconstruction Method for Source-Oriented Linear Scan CT") combine analytical methods with learning, they still assume that the projection data is relatively complete and do not solve the problem of end-to-end high-quality reconstruction under highly truncated sine waves. Summary of the Invention

[0006] Addressing the core challenge of severely missing and unevenly distributed projection angles in multi-source linear trajectory CT systems, this invention proposes a multi-source linear CT reconstruction method based on projection fusion and depth unfolding networks. By using a cascaded network of multi-source linear scanning projection fusion rearrangement and sine curve completion-image reconstruction, this method effectively solves the artifact and distortion problems caused by the limited angles in traditional linear scanning CT while ensuring reconstruction efficiency, thus achieving high-fidelity, artifact-free CT tomographic image reconstruction.

[0007] To achieve the above objectives, the technical solution adopted is:

[0008] This invention provides a multi-source linear CT reconstruction method based on projection fusion and depth unfolding networks, comprising the following steps:

[0009] Step 1, Multi-source linear trajectory CT scan: Arrange no fewer than three X-ray sources at a specific interval and tilt angle to form a radiation source group, and fix it on one side of the scanning area; configure a single area array detector on the opposite side; control the object under examination to translate along a linear trajectory parallel to the plane of the detector and the line connecting the radiation sources; during the translation process, each radiation source is exposed sequentially according to a preset pulse sequence, and the detector receives the projection signal penetrating the sample in a time-division manner to obtain multiple sets of projection data sequences reflecting different translation positions;

[0010] Step 2, Geometrically Guided Projection Fusion and Rearrangement: Based on the object translation axis and the spatial geometric relationship between the fixed X-ray source and the detector, an imaging spatial mapping model under the moving coordinate system is constructed. The projection data sequence is fused and rearranged into a two-dimensional sine graph according to spatial correlation. The sine graph has structural missing regions in the angular direction due to the geometric limitations of straight-line scanning.

[0011] Step 3: End-to-end cascaded network reconstruction: Construct a cascaded network for sine curve completion and image reconstruction, consisting of a projection generation completion submodule and an image intelligent reconstruction submodule. The missing sine curve and mask marked with missing locations obtained in Step 2 are input into the cascaded network. First, the generation completion submodule predicts and fills in the missing data while retaining the actual measured values. Then, the completed sine curve is input into the image intelligent reconstruction submodule, which adopts a deep unfolding architecture and includes multiple iterative stages. Each stage updates the reconstructed image through a weight-sharing backprojection operator and a residual regularization module. Finally, the cascaded network directly outputs the reconstructed CT tomographic image.

[0012] Step 4, End-to-end joint training: The sinusoidal graph completion-image reconstruction cascade network is jointly trained end-to-end using a composite loss function that includes both the sinusoidal graph domain and the image domain.

[0013] According to the multi-source linear CT reconstruction method based on projection fusion and depth unfolding network of the present invention, in step 2, the moving coordinate system takes the geometric center of the object as the origin and moves synchronously with the object. The construction process of the imaging space mapping model is as follows: Let the translation position of the object along the x-axis be x. obj The vertical distance from the X-ray source to the translation trajectory is D, and the fan beam angle of each X-ray source is β. Using the geometric transformation formula θ = arctan(x... obj / D)+β and t=-x obj cosθ+Dsinθ establishes the physical fan-beam space (x obj The mapping relationship from (β) to the virtual parallel beam space (θ,t), where θ is the equivalent projection angle and t is the detector channel coordinate;

[0014] Based on this mapping relationship, a rearranged lookup table is constructed. The construction logic is as follows: traverse the discrete grid points of the target parallel beam sine diagram, calculate the floating-point coordinate index of the target parallel beam projection data through inverse geometric transformation, and pre-calculate and store the neighboring point addresses and weights required for bilinear interpolation. If the grid point has corresponding original data, retrieve the measurement value of the neighboring point from the original data, perform bilinear interpolation according to the weights, and fill the interpolation result into the sine diagram.

[0015] According to the multi-source linear CT reconstruction method based on projection fusion and depth unfolding network of the present invention, further, the generation and completion submodule in step 3 adopts a generative adversarial network generator based on U-Net structure, and integrates a data consistency constraint layer after the output of the generator; let the predicted sine curve output by the generator be y. pred The actual measurement data is y gt If the binary mask is M, then the data consistency constraint layer is used to calculate the completed sine curve y of the final input image intelligent reconstruction submodule. completed as follows:

[0016] y completed =M⊙y gt +(1-M)⊙y pred .

[0017] According to the multi-source linear CT reconstruction method based on projection fusion and depth unfolding network of the present invention, the generation logic of the binary mask M is further as follows: in the data fusion rearrangement process in step 2, according to the geometric mapping relationship of the multi-source linear scanning trajectory, it is determined whether there is a corresponding effective physical ray projection for the coordinates of each pixel point in the sine grid. If there is, mark M(θ,t) = 1; if there is not, mark M(θ,t) = 0, corresponding to the region to be filled.

[0018] According to the multi-source linear CT reconstruction method based on projection fusion and depth unfolding network of the present invention, the weight-sharing backprojection operator in step 3 is a learnable data-driven module, comprising three computational layers connected in sequence:

[0019] (1) Frequency domain filtering layer: A learnable one-dimensional convolution kernel is used to convolve the sinusoidal residual data along the detector channel direction. The kernel size is an odd number greater than 15 to simulate the slope filtering effect.

[0020] (2) Trajectory Tracking Rearrangement Layer: The coordinate grid (x,y) of the image domain to be reconstructed is predefined. For each coordinate point (x,y) in the image domain... i ,y i and each projection angle θ j The sampling index t of the straight line trajectory in the filtered sine curve is calculated based on the geometric relationship of the straight line trajectory. ij =x icosθ j +y i sinθ j Using the bilinear interpolation algorithm based on index t ij Extract the corresponding projection density value from the filtered sine wave and construct the feature vector corresponding to that pixel.

[0021] (3) Weight-sharing backprojection layer: Define a weight vector that is shared across the entire image domain, perform a dot product operation with all pixel feature vectors, and output the data fidelity update value. The weight vector does not change with the pixel position.

[0022] According to the multi-source linear CT reconstruction method based on projection fusion and depth unfolding network of the present invention, the residual regularization module in step 3 is a convolutional neural network based on global residual learning. The network includes an input feature extraction layer, a stacked nonlinear mapping layer, and an output reconstruction layer. The network is configured to take the current iterative image as input, output an estimated artifact and noise residual map, and generate a denoised updated image by subtracting the input image from the residual map.

[0023] According to the multi-source linear CT reconstruction method based on projection fusion and depth unfolding network of the present invention, the image update process of each iteration stage of the intelligent image reconstruction submodule is further defined by the following formula:

[0024]

[0025] Where, x (k) For the input reconstructed image in the k-th iteration stage, x (k+1) For the updated reconstructed image output in the k-th iteration stage, η (k) Let be the learnable step size parameter for the k-th iteration stage. Let A denote the weight-sharing back projection operator, and let y denote the forward projection operator. completed This represents the completed sine curve from the completion generation submodule. This represents the residual regularization module.

[0026] According to the multi-source linear CT reconstruction method based on projection fusion and depth unfolding network of the present invention, the composite loss function in step 4 further includes:

[0027] Sine domain loss L sino This is a weighted sum of the L1 norm distance and gradient differences between the predicted sine curve and the true complete sine curve;

[0028] Image domain loss L image The weighted sum of the mean squared error and structural similarity difference between the network-reconstructed image and the real labeled image;

[0029] Total loss L total =L sino +γ·L image , where γ is a hyperparameter that balances the losses of the two domains.

[0030] According to the multi-source linear CT reconstruction method based on projection fusion and depth unfolding network of the present invention, the sinusoidal domain loss L sino The calculation formula is:

[0031]

[0032] in, Let λ represent the Sobel gradient operator. grad y represents the gradient weight coefficients. true For a true and complete sine curve, y completed This is the completed sine curve from the completion generation submodule;

[0033] The image domain loss L image The calculation formula is:

[0034]

[0035] Where, x output To reconstruct images for the network, x label For real CT labeled images, λ ssim SSIM represents the structural similarity weights and the structural similarity index.

[0036] Furthermore, the present invention also provides a multi-source linear CT reconstruction system based on projection fusion and depth unfolding networks, comprising:

[0037] The scanning device includes a fixed X-ray source group, a single area array detector, and a transmission mechanism for controlling the linear translation of the object, used to acquire multi-view projection data;

[0038] The processor is configured to execute the methods described above;

[0039] The processor contains a trained sine curve completion-image reconstruction cascade network for end-to-end reconstruction of CT tomographic images from an input truncated sine curve.

[0040] The beneficial effects achieved by adopting the above technical solution are:

[0041] This invention significantly improves the imaging quality and reconstruction efficiency of linear scanning security CT scans. Its outstanding effect lies in addressing the industry problem of severe stripe artifacts and geometric distortions caused by the limited angle of traditional linear scanning. This invention innovatively adopts a cascaded architecture of "intelligent sine wave completion + depth unfolding reconstruction". First, it uses a generative adversarial network combined with data consistency constraints to accurately repair missing projection data in the sine wave domain, eliminating the physical causes of artifacts at the source. Then, it replaces the traditional time-consuming iterative reconstruction algorithm with a depth unfolding network that includes a prior physical model. While ensuring real-time reconstruction speed, it uses physical imaging operators to constrain the network output, effectively avoiding the risk of false textures that are easily generated by pure deep learning methods. This allows low-cost linear scanning systems using fixed X-ray sources to obtain high-fidelity, artifact-free images comparable to full-angle rotating CT scans. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0043] Figure 1 This is a schematic diagram of the structure of a multi-X-ray source linear trajectory CT scanning system according to an embodiment of the present invention. In the figure, 1-X-ray source group; 11-left X-ray source; 12-middle X-ray source; 13-right X-ray source; 2-conveyor belt transmission module; 3-detector module;

[0044] Figure 2 This is a schematic diagram of the sine curve completion-image reconstruction cascade network reconstruction process according to an embodiment of the present invention. Detailed Implementation

[0045] The exemplary solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art.

[0046] This invention discloses a multi-source linear CT reconstruction method based on projection fusion and depth unfolding networks, comprising the following steps:

[0047] Step S101, Multi-source linear trajectory CT scan: Arrange no fewer than three (preferably an odd number) X-ray sources at a specific interval and tilt angle to form a X-ray source group, and fix it on one side of the scanning area; configure a single high-sensitivity area array detector on the opposite side; control the object under examination to move at a constant speed along a straight trajectory parallel to the detector plane and the line connecting the X-ray sources; during the translation process, each X-ray source is exposed sequentially according to a preset pulse sequence, and the detector uses a time-division multiplexing method to receive the projection signal penetrating the sample, and acquire multiple sets of projection data sequences reflecting different translation positions.

[0048] The spacing between the X-ray sources is set according to the size of the scanning area and the required imaging accuracy. The tilt angle ranges from 0° to 45°, and the tilt directions of adjacent X-ray sources are alternately distributed to expand the field of view coverage. Figure 1 As shown.

[0049] Step S102, Geometric-Guided Projection Fusion and Rearrangement: Based on the object's translation axis and the spatial geometric relationship between the fixed X-ray source and the detector, an imaging spatial mapping model in a moving coordinate system is constructed. The projection data sequence is fused and rearranged into a two-dimensional sine curve according to spatial correlation. The sine curve contains structurally missing regions in the angular direction due to the geometric limitations of linear scanning. This projection fusion and rearrangement uniformly maps multi-source, multi-location cone-beam data into a geometrically consistent two-dimensional parallel-beam sine curve, preserving the spatial structure that can be used for learning.

[0050] The specific implementation of projection blending and rearrangement is as follows: Define a translation coordinate system with the geometric center of the object as the origin, which moves synchronously with the object. Let the translation position of the object along the x-axis be x. obj The vertical distance from the ray source to the translation trajectory is D, and the fan beam angle of each ray source is β. Using the principle of ray path equivalence, a path is established from the physical fan beam space (x obj The mapping relationship from (β) to the virtual parallel bundle space (θ,t):

[0051] θ = arctan(x) obj / D)+β

[0052] t = -x obj cosθ+Dsinθ

[0053] Where θ is the equivalent projection angle and t is the detector channel coordinate.

[0054] Based on this mapping relationship, a rearranged lookup table is constructed, and its construction logic is as follows:

[0055] The algorithm iterates through all discrete grid points in the target parallel-beam sine map, performs an inverse geometric transformation using the mapping formula described above to calculate their floating-point coordinate indices in the original fan-beam projection data, and pre-calculates and stores the neighboring point addresses and weights required for bilinear interpolation. If a grid point has corresponding original data, the values ​​of the neighboring points (which are the detector's measurements, i.e., the intensity of X-rays after attenuation by the object) are retrieved from the original data, and bilinear interpolation is performed according to the weights. The interpolation results are then filled into the sine map. If a grid point does not have corresponding original data (outside the scanning geometry), it remains a missing value (0). These missing points constitute structurally missing regions. This lookup table is used to quickly resample the physical scan data into a parallel-beam sine map.

[0056] Bilinear interpolation is used to fill the original projection data into a uniformly spaced grid (θ,t) to generate a two-dimensional sine graph. Due to the geometric constraints of linear scanning, this sine graph has structurally missing regions in the angular directions (i.e., a finite-angle problem).

[0057] Step S103, End-to-end cascaded network reconstruction: (e.g.) Figure 2 As shown, a cascaded network for sine curve completion and image reconstruction is constructed. This network consists of a projection generation completion submodule and an image intelligent reconstruction submodule. The missing sine curve and the mask marking the missing positions obtained in step S102 are input into the cascaded network. First, the generation completion submodule predicts and fills in the missing data while retaining the actual measurement values. Then, the completed sine curve is input into the image intelligent reconstruction submodule. This submodule adopts a deep unfolding architecture and contains multiple iterative stages. Each stage updates the reconstructed image through a weight-sharing backprojection operator and a residual regularization module. Finally, the cascaded network directly outputs the reconstructed CT tomographic image.

[0058] (1) Generate completion submodule

[0059] The completion submodule employs a Generative Adversarial Network (GAN) generator based on the U-Net architecture. The input is a sine wave with missing regions and its mask (where the real regions are 1 and the missing regions are 0). The network uses the geometric continuity of the sine wave to predict the missing regions and integrates a data consistency constraint layer after the generator's output.

[0060] The data consistency constraint layer is used to force the fusion of real measurement data and network prediction data; let the predicted sine curve output by the generator be y. pred The actual measurement data is y gt If the binary mask is M, then the completed sine curve y of the final input image intelligent reconstruction submodule is... completed as follows:

[0061] y completed =M⊙ygt +(1-M)⊙y pred

[0062] The generation logic of the binary mask M is as follows: In the data fusion and rearrangement process in step S102, based on the geometric mapping relationship of the multi-source straight line scanning trajectory, it is determined whether there is a corresponding effective physical ray projection for the coordinates of each pixel in the sine grid. If there is an effective projection mapping, then mark M(θ,t) = 1, corresponding to the known data area; if there is no effective projection mapping (i.e., within the angle missing range caused by the straight line scanning), then mark M(θ,t) = 0, corresponding to the area to be filled; ensure that the predicted value is used only in the missing area, and the real measurement data is strictly preserved.

[0063] (2) Image intelligent reconstruction submodule: adopts the deep unfolding architecture (LEARN) and contains N iterative stages; each stage consists of a weight-sharing backprojection operator and a residual regularization module.

[0064] The weight-sharing backprojection operator is a learnable, data-driven module containing three sequentially connected computational layers:

[0065] ① Frequency domain filtering layer: A learnable one-dimensional convolution kernel is used to convolve the sinusoidal residual data along the detector channel direction. The kernel size is an odd number greater than 15 to simulate the ramp filtering effect and extract high-frequency edges.

[0066] ② Trajectory Tracking Rearrangement Layer: Based on the geometric relationship of the straight line trajectory, the sampling index of each pixel in the image domain in the filtered sine wave is calculated, and the corresponding projection density value is extracted through bilinear interpolation to construct the pixel feature vector. The specific implementation is as follows:

[0067] A coordinate grid (x, y) for the image domain to be reconstructed is predefined. For each coordinate point (x, y) in the image domain... i ,y i and each projection angle θ j The sampling index t of the straight line trajectory in the filtered sine curve is calculated based on the geometric relationship of the straight line trajectory. ij =x i cosθ j +y i sinθ j Using the bilinear interpolation algorithm based on index t ij Extract the corresponding projection density value from the filtered sine wave and construct the feature vector corresponding to that pixel. Where N view This represents the total number of projection angles. The role of the bilinear interpolation algorithm here is: because the calculated t... ijIt could be a non-integer like 50.3. There is no "50.3rd pixel". Therefore, we need to use the values ​​of the 50th and the lower 51st pixels to interpolate and calculate a value at the "virtual position 50.3". The value obtained by this interpolation is the corresponding projection density.

[0068] For example: For the pixel (30, 40) in the image, iterating through all angles, when angle θ... j Calculate t when =0° ij =30cos0°+40sin0°=30, in the 0th row of the filtered sine wave (corresponding to θ) j =0°), take the value at t=30, and get the value v. i [0] = 0.85; when angle θ j When = 1°, calculate t ij =30cos1°+40sin1°=30.7, in the first row of the filtered sine wave (corresponding to θ) j =1°), take the value at t=30.7 (interpolation is required) to obtain the value v. i [1] = 0.83, and so on. Finally, a value containing 180(θ) is obtained. j A vector with values ​​ranging from 0° to 179°.

[0069] ③ Weight-sharing backprojection layer: Defines a weight vector shared across the entire image domain, performs a dot product operation with all pixel feature vectors, and outputs a data fidelity update value. The weight vector does not change with pixel position and is used to learn general physical projection rules. The weight-sharing backprojection layer is a learnable, data-driven physical projection simulator, significantly different from the fixed-kernel FBP.

[0070] The residual regularization module is a convolutional neural network based on global residual learning. This network includes an input feature extraction layer, stacked nonlinear mapping layers, and an output reconstruction layer. It contains a total of 3 to 5 3×3 convolutional layers, with ReLU activation functions embedded in the intermediate layers. The network is configured to take the current iteration image as input, output an estimated artifact and noise residual map, and generate a denoised updated image by subtracting the residual map from the input image.

[0071] In this embodiment, the regularization module in each iteration stage A lightweight, shallow, fully convolutional network structure is employed, comprising three layers: The first layer is a feature extraction layer, using convolution (Conv) + ReLU operations with a 3×3 kernel, 1 input channel, and 32 output channels; the middle layers are layers 1 to 3, which are non-linear mapping layers, all using convolution (Conv) + ReLU operations with a 3×3 kernel and 32 input and 32 output channels; the last layer is a reconstruction output layer, using only convolution (Conv) operations, with 32 input channels and 1 output channel, its function being to map features back to the image space, outputting a "residual map." This is achieved through the Global Skip Connection formula. To achieve residual learning, where x input For the input image, For residual regularization module, The noise residual map output by the residual regularization module, x clean This is the image after denoising.

[0072] Finally, the image update formula for each iteration stage of the intelligent image reconstruction submodule is given:

[0073]

[0074] Where, x (k) For the input reconstructed image in the k-th iteration stage, x (k+1) For the updated reconstructed image output in the k-th iteration stage, η (k) Let be the learnable step size parameter for the k-th iteration stage. Let A denote the weight-sharing back projection operator, and let y denote the forward projection operator. completed This represents the completed sine curve from the completion generation submodule. This represents the residual regularization module. Initial image x (0) A graph consisting entirely of zeros, i.e., x, can typically be used. (0) =0.

[0075] Step S104, End-to-End Joint Training: The sinusoidal incomplete-image reconstruction cascaded network is jointly trained end-to-end using a composite loss function encompassing both the sinusoidal and image domains. Adjustable parameters are used to balance the losses of the two parts. An image quality-oriented joint optimization mechanism is employed: the image domain reconstruction error gradient is backpropagated to the generation and incomplete submodule via a differentiable operator in the intelligent image reconstruction submodule, simultaneously optimizing the parameters of both submodules to achieve sinusoidal incomplete with a focus on final image quality. The global loss function is a dual-domain composite loss.

[0076] Sine domain loss L sinoTo predict the L1 norm distance and the weighted sum of gradient differences between the predicted sine curve and the true complete sine curve, the formula is as follows:

[0077]

[0078] in, Let λ represent the Sobel gradient operator. grad y represents the gradient weight coefficients. true For a true and complete sine curve, y completed The completed sine curve is derived from the completion submodule; λ grad This is used to force the edges of the sine curve generated by the network to remain sharp. The first term on the right-hand side of the equation ensures that the numerical value of the completed sine curve is close to the true complete sine curve, and the second term ensures that the edge structure of the completed sine curve is consistent with the true one.

[0079] Image domain loss L image The weighted sum of the mean squared error and structural similarity difference between the network-reconstructed image and the ground truth labeled image is calculated using the following formula:

[0080]

[0081] Where, x output To reconstruct images for the network, x label For real CT labeled images, λ ssim λ represents the structural similarity weight. ssim It is used to restore the visual texture structure of the image while restoring the numerical accuracy of CT; SSIM is the structural similarity index, which is used to measure the similarity between two images in terms of brightness, contrast and structure; the first term on the right side of the equation guarantees the pixel value accuracy of the reconstructed image, and the second term guarantees the visual quality and texture structure of the reconstructed image.

[0082] Total loss L total =L sino +γ·L image , where γ is a hyperparameter balancing the losses of the two domains, and the value of the balancing hyperparameter γ is 0.1. This strategy ensures that sine wave completion not only satisfies its own smoothness, but also aims to preserve the fidelity of the final image structure, avoiding the accumulation of errors in staged processing.

[0083] The training data was constructed by generating non-rigid deformation models containing bone, soft tissue, and metal implants using finite element simulation (FEM), combined with Monte Carlo ray tracing to simulate multi-source linear scanning, and batch generating paired datasets of "truncated sine curves – complete CT images".

[0084] This invention also discloses a multi-source linear CT reconstruction system based on projection fusion and depth unfolding networks, comprising:

[0085] The scanning device includes a fixed X-ray source group, a single area array detector, and a transmission mechanism for controlling the linear translation of the object, used to acquire multi-view projection data;

[0086] The processor is configured to execute the aforementioned multi-source linear CT reconstruction method based on projection fusion and depth unfolding networks;

[0087] The processor contains a trained sine curve completion-image reconstruction cascade network for end-to-end reconstruction of CT tomographic images from an input truncated sine curve.

[0088] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-source linear CT reconstruction method based on projection fusion and depth unfolding network, characterized in that, Includes the following steps: Step 1, Multi-source linear trajectory CT scan: Arrange no fewer than three X-ray sources into a radiation source group at a specific interval and tilt angle, and fix it on one side of the scanning area; configure a single area array detector on the opposite side; control the object under examination to translate along a linear trajectory parallel to the plane of the detector and the line connecting the radiation sources; During the translation process, each X-ray source is exposed sequentially according to a preset pulse sequence, and the detector receives the projection signal penetrating the sample in a time-division manner to obtain multiple sets of projection data sequences reflecting different translation positions; Step 2, Geometrically Guided Projection Fusion and Rearrangement: Based on the object translation axis and the spatial geometric relationship between the fixed X-ray source and the detector, an imaging spatial mapping model under the moving coordinate system is constructed. The projection data sequence is fused and rearranged into a two-dimensional sine graph according to spatial correlation. The sine graph has structural missing regions in the angular direction due to the geometric limitations of straight-line scanning. Step 3, End-to-end cascaded network reconstruction: Construct a cascaded network of sine curve completion and image reconstruction, which consists of a projection generation completion submodule and an image intelligent reconstruction submodule; Input the mask with missing sine curve and marked missing positions obtained in Step 2 into the cascaded network, first predict and fill the missing data through the generation completion submodule, while retaining the real measurement value; The completed sine wave is then input into the image intelligent reconstruction submodule. This submodule adopts a deep unfolding architecture and contains multiple iterative stages. Each stage updates the reconstructed image by jointly updating the reconstructed image through a weight-sharing backprojection operator and a residual regularization module. Finally, the reconstructed CT tomographic image is directly output by the cascaded network. Step 4, End-to-end joint training: The sinusoidal graph completion-image reconstruction cascade network is jointly trained end-to-end using a composite loss function that includes both the sinusoidal graph domain and the image domain.

2. The multi-source linear CT reconstruction method based on projection fusion and depth unfolding network according to claim 1, characterized in that, The moving coordinate system mentioned in step 2 has its origin at the geometric center of the object and moves synchronously with the object. The construction process of the imaging space mapping model is as follows: Let the object's translation position along the x-axis be x. obj The vertical distance from the X-ray source to the translation trajectory is D, and the fan beam angle of each X-ray source is β. Using the geometric transformation formula θ = arctan(x... obj / D)+β and t=-x obj cosθ+Dsinθ establishes the physical fan-beam space (x obj The mapping relationship from (β) to the virtual parallel beam space (θ,t), where θ is the equivalent projection angle and t is the detector channel coordinate; Based on this mapping relationship, a rearranged lookup table is constructed. The construction logic is as follows: traverse the discrete grid points of the target parallel beam sine diagram, calculate the floating-point coordinate index of the target parallel beam projection data through inverse geometric transformation, and pre-calculate and store the neighboring point addresses and weights required for bilinear interpolation. If the grid point has corresponding original data, retrieve the measurement value of the neighboring point from the original data, perform bilinear interpolation according to the weights, and fill the interpolation result into the sine diagram.

3. The multi-source linear CT reconstruction method based on projection fusion and depth unfolding network according to claim 1, characterized in that, The generation completion submodule described in step 3 employs a generative adversarial network generator based on a U-Net architecture, and integrates a data consistency constraint layer after the generator's output; let the predicted sine curve output by the generator be y. pred The actual measurement data is y gt If the binary mask is M, then the data consistency constraint layer is used to calculate the completed sine curve y of the final input image intelligent reconstruction submodule. completed as follows: y completed =M⊙y gt +(1-M)⊙y pred 。 4. The multi-source linear CT reconstruction method based on projection fusion and depth unfolding network according to claim 3, characterized in that, The generation logic of the binary mask M is as follows: In the data fusion and rearrangement process in step 2, based on the geometric mapping relationship of the multi-source straight line scanning trajectory, it is determined whether there is a corresponding effective physical ray projection for the coordinates of each pixel point in the sine grid. If it exists, mark M(θ,t) = 1; if it does not exist, mark M(θ,t) = 0, corresponding to the area to be filled.

5. The multi-source linear CT reconstruction method based on projection fusion and depth unfolding network according to claim 1, characterized in that, The weight-sharing backprojection operator mentioned in step 3 is a learnable data-driven module, which includes three computational layers connected in sequence: (1) Frequency domain filtering layer: A learnable one-dimensional convolution kernel is used to convolve the sinusoidal residual data along the detector channel direction. The kernel size is an odd number greater than 15 to simulate the slope filtering effect. (2) Trajectory Tracking Rearrangement Layer: The coordinate grid (x,y) of the image domain to be reconstructed is predefined. For each coordinate point (x,y) in the image domain... i ,y i and each projection angle θ j The sampling index t of the straight line trajectory in the filtered sine curve is calculated based on the geometric relationship of the straight line trajectory. ij =x i cosθ j +y i sinθ j Using the bilinear interpolation algorithm based on index t ij Extract the corresponding projection density value from the filtered sine wave and construct the feature vector corresponding to that pixel. (3) Weight-sharing backprojection layer: Define a weight vector that is shared across the entire image domain, perform a dot product operation with all pixel feature vectors, and output the data fidelity update value. The weight vector does not change with the pixel position.

6. The multi-source linear CT reconstruction method based on projection fusion and depth unfolding network according to claim 1, characterized in that, The residual regularization module mentioned in step 3 is a convolutional neural network based on global residual learning. This network includes an input feature extraction layer, stacked nonlinear mapping layers, and an output reconstruction layer. The network is configured to take the current iteration image as input, output an estimated artifact and noise residual map, and generate a denoised updated image by subtracting the residual map from the input image.

7. The multi-source linear CT reconstruction method based on projection fusion and depth unfolding network according to claim 1, characterized in that, The image update process for each iteration stage of the intelligent image reconstruction submodule is defined by the following formula: Where, x (k) For the input reconstructed image in the k-th iteration stage, x (k+1) For the updated reconstructed image output in the k-th iteration stage, η (k) Let be the learnable step size parameter for the k-th iteration stage. Let A denote the weight-sharing back projection operator, and let y denote the forward projection operator. completed This represents the completed sine curve from the completion generation submodule. This represents the residual regularization module.

8. The multi-source linear CT reconstruction method based on projection fusion and depth unfolding network according to claim 1, characterized in that, The composite loss function mentioned in step 4 specifically includes: Sine domain loss L sino This is a weighted sum of the L1 norm distance and gradient differences between the predicted sine curve and the true complete sine curve; Image domain loss L image The weighted sum of the mean squared error and structural similarity difference between the network-reconstructed image and the real labeled image; Total loss L total =L simo +γ·L image , where γ is a hyperparameter that balances the losses of the two domains.

9. The multi-source linear CT reconstruction method based on projection fusion and depth unfolding network according to claim 8, characterized in that, The sinusoidal domain loss L sino The calculation formula is: in, Let λ represent the Sobel gradient operator. grad y represents the gradient weight coefficients. true For a true and complete sine curve, y completed This is the completed sine curve from the completion generation submodule; The image domain loss L image The calculation formula is: Where, x output To reconstruct images for the network, x label For real CT labeled images, λ ssim SSIM represents the structural similarity weights and the structural similarity index.

10. A multi-source linear CT reconstruction system based on projection fusion and depth unfolding network, characterized in that, include: The scanning device includes a fixed X-ray source group, a single area array detector, and a transmission mechanism for controlling the linear translation of the object, used to acquire multi-view projection data; The processor is configured to perform the method as described in any one of claims 1 to 9; The processor contains a trained sine curve completion-image reconstruction cascade network for end-to-end reconstruction of CT tomographic images from an input truncated sine curve.

Images