A sparse-view ct sinogram continuity implicit neural super-resolution method

Abstract

The application provides a sparse-view CT sinogram continuity implicit neural super-resolution method, belonging to the technical field of data processing, comprising: acquiring sparse-view CT projection data to construct an input sparse sinogram, and extracting high-frequency enhancement features through an edge-aware encoder; predicting Fourier residual parameters and detector axis shift fields by a decoder double branch respectively; calculating forward and backward trajectory-aware twist coordinates according to the relative position of the target angle in the adjacent sampling interval, and generating bidirectional residuals in combination with the shift field; weighting and fusing the basic interpolation results of adjacent sampling angles and bidirectional residuals according to the relative position to obtain target angle recovery projection; and repeatedly processing multiple target angles to output a continuous dense sinogram. The scheme of the disclosure improves the structure fidelity, angle continuity and downstream CT reconstruction quality of sparse projection recovery through continuous angle representation and trajectory-aware shift modeling.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to a continuous implicit neural super-resolution method for sparse-view CT sine waves. Background Technology

[0002] Currently, CT is an important imaging tool in clinical diagnosis, but repeated scans can lead to high radiation doses. Sparse-view CT reduces radiation dose and scan time by decreasing the number of sampling angles, but insufficient projection angles can result in missing sine wave data, further causing fringe artifacts, structural blurring, and detail distortion after backprojection reconstruction. Since projection domain errors are propagated and amplified by the inverse Radon transform, directly recovering a high-fidelity dense sine wave in the projection domain is key to improving the reconstruction quality of sparse-view CT.

[0003] Most existing methods treat the sine wave as a regular two-dimensional image, using interpolation, image inpainting, or fixed-magnification super-resolution for restoration. While these methods can achieve some success under specific sampling settings, they typically suffer from the following shortcomings: First, they fail to explicitly describe the continuous acquisition characteristics of CT projection data along the angular direction; second, even when introducing a continuous representation model, they often ignore the physical fact that the same anatomical structure drifts along the detector coordinate direction as the scanning angle changes; third, existing continuous rendering methods tend to over-smooth, making it difficult to recover abrupt responses near the boundaries of high-contrast structures. These problems result in insufficient generalization ability of existing methods in cross-scale, unseen angular density, or complex structural scenarios.

[0004] It is evident that there is an urgent need for a continuous implicit neural super-resolution method for recovering sparse-view CT sinusoids with high accuracy and strong generalization ability. Summary of the Invention

[0005] In view of this, the present disclosure provides a continuous implicit neural super-resolution method for sparse-view CT sine waves, which at least partially solves the problems existing in the prior art.

[0006] This disclosure provides a continuous implicit neural super-resolution method for sparse-view CT sinusoidal plots, including: Step 1: Obtain sparse view CT projection data of the target object and construct the input sparse sine curve according to the order of the sampled angles; Step 2: Input the input sparse sinusoidal graph into the edge-aware encoder to extract multi-directional high-frequency mutation features and obtain high-frequency enhanced feature representation; Step 3: Input the high-frequency enhancement feature representation into the decoder, predict the Fourier residual parameters by the first branch of the decoder, and predict the detector axis offset field at each sampled angle position by the second branch of the decoder. Step 4: Based on the relative position of the target angle within the adjacent sampled angle intervals, and combined with the detector axis offset field, calculate the forward trajectory sensing distortion coordinates and the backward trajectory sensing distortion coordinates respectively. Step 5: Based on the forward trajectory sensing twisted coordinates and the backward trajectory sensing twisted coordinates, generate the forward residual and the backward residual respectively using the Fourier residual parameters; Step 6: Perform basic interpolation on the projection values ​​of adjacent sampled angles to obtain the basic interpolation results; Step 7: The basic interpolation results are weighted and fused with the forward and backward residuals according to their relative positions to generate the restored projection at the target angle; Step 8: Repeat steps 1 to 7 for multiple target angles to output a continuous dense sine curve.

[0007] According to a specific implementation of an embodiment of this disclosure, the edge-aware encoder includes an explicit edge extraction layer and a residual backbone network. The explicit edge extraction layer uses multiple anisotropic convolutional masks with different directions and aspect ratios to extract multi-directional abrupt responses. The residual backbone network fuses and enhances the multi-directional abrupt responses to obtain high-frequency enhanced feature representations.

[0008] According to one specific implementation of this disclosure, the anisotropic convolution mask includes multiple combinations of 7×3, 3×7, 7×7, 5×5, 5×3 and 3×5 directional masks.

[0009] According to a specific implementation of an embodiment of this disclosure, the first branch of the decoder adopts a 1×1 convolution structure, outputting several sets of cosine coefficients and sine coefficients for each local anchor point to form Fourier residual parameters. The second branch of the decoder adopts a 1×1 convolution structure and outputs the detector axis offset field.

[0010] According to a specific implementation of this disclosure, the forward trajectory sensing distortion coordinates and the backward trajectory sensing distortion coordinates are determined in the following manner: Set target angle Located at adjacent sampled angles and Between, relative position The detector coordinates are The detector axis offset field is and ,but: The forward trajectory perception distortion coordinates are: ; The backward trajectory perception distortion coordinates are: .

[0011] According to a specific implementation of this disclosure, the Fourier residual parameters are modeled on a local angle interval basis, and the residual term at any relative position t within each local angle interval is explicitly calculated. The expression for the residual term is:

[0012] in, This represents the magnitude of the cosine of the Fourier series. This represents the magnitude of the sine wave in the Fourier series. Indicates a predefined frequency reference; When calculating the forward residual, the detector coordinates in the residual term expression are used. Replace with forward trajectory-aware warped coordinates ,get: ; When calculating the backward residual, the detector coordinates in the residual term expression are... Replace with the backward trajectory-aware warp coordinates and relative position Replace with ,get: .

[0013] According to a specific implementation of this disclosure, the expression for the basic interpolation result is:

[0014] in, This represents the projected value of the previously measured angle. This represents the projected value of the next measured angle; The expression for the restored projection at the target angle is:

[0015] in, This represents the repaired residual from the projection of the previous measurement angle. This represents the repaired residual from the projection of the next measurement angle.

[0016] According to a specific implementation of this disclosure, after step 8, the method further includes: A continuous dense sine wave is input into a filtered back-projection reconstruction algorithm or an iterative reconstruction algorithm to generate a CT image of the target object.

[0017] The continuous implicit neural super-resolution scheme for sparse-view CT sinusoidal images in this embodiment includes: Step 1, acquiring sparse-view CT projection data of the target object and constructing an input sparse sinusoidal image according to the sampled angle order; Step 2, inputting the input sparse sinusoidal image into an edge-aware encoder to extract multi-directional high-frequency abrupt features and obtain a high-frequency enhanced feature representation; Step 3, inputting the high-frequency enhanced feature representation into a decoder, predicting Fourier residual parameters by the first branch of the decoder, and predicting the detector axis offset field at each sampled angle position by the second branch of the decoder; Step 4, based on the target angle in adjacent sampled angles... Step 5: Based on the relative positions within the interval and the detector axis offset field, calculate the forward trajectory sensing distortion coordinates and the backward trajectory sensing distortion coordinates respectively; Step 6: Based on the forward trajectory sensing distortion coordinates and the backward trajectory sensing distortion coordinates, generate the forward residual and the backward residual respectively using the Fourier residual parameters; Step 7: Perform basic interpolation on the projection values ​​of adjacent sampled angles to obtain the basic interpolation result; Step 8: Weight and fuse the basic interpolation result with the forward residual and the backward residual according to their relative positions to generate the restored projection under the target angle; Step 9: Repeat steps 1 to 7 for multiple target angles to output a continuous dense sine wave.

[0018] The beneficial effects of the embodiments disclosed herein are as follows: 1. Sparse view CT projection restoration is redefined as a continuous super-resolution problem along the angular dimension, rather than a discrete interpolation problem with a fixed multiple.

[0019] 2. An explicit edge extraction module is introduced in the encoding stage, and multiple anisotropic orientation masks are combined with learnable convolutions to enhance the high-frequency mutation structure.

[0020] 3. A lightweight dual-head structure is used in the decoding stage, with one head predicting the Fourier phase residual parameters and the other head predicting the detector axial velocity flow or offset field.

[0021] 4. Based on the relative position of the target angle, perform forward and backward trajectory sensing distortion on the detector coordinates corresponding to adjacent sampling angles to establish a continuous coordinate system that conforms to the CT acquisition rules.

[0022] 5. The final target projection is formed by using basic interpolation plus two-way residual fusion, thus taking into account local continuity, global consistency and high-frequency detail recovery. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart illustrating a continuous implicit neural super-resolution method for sparse-view CT sinusoidal plots provided in this embodiment of the disclosure; Figure 2 The overall network architecture diagram corresponding to the continuous implicit neural super-resolution method for sparse view CT sinusoidal graphs provided in the embodiments of this disclosure; Figure 3 A comparison diagram illustrating the principles of traditional super-resolution rendering and continuous trajectory-aware rendering of the present invention, provided for embodiments of this disclosure; Figure 4 A schematic diagram of the decoder header structure provided in an embodiment of this disclosure; Figure 5 A visualization of arbitrary-scale reconstruction under a 30→90 training setting provided in an embodiment of this disclosure; Figure 6 This is a visualization of the noise power spectrum under a training setting of 30→90, provided in an embodiment of this disclosure. Figure 7 The figure shows the cross-scale generalization experiment results provided in the embodiments of this disclosure under training settings of 45→90, 45→180 and 30→90; Figure 8 A schematic diagram of the learned sensor offset trajectory provided in an embodiment of this disclosure. Detailed Implementation

[0025] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0026] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0027] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0028] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The illustrations only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0029] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0030] This disclosure provides a continuous implicit neural super-resolution method for sparse-view CT sine waves, which can be applied to CT image reconstruction in medical settings.

[0031] See Figure 1 This is a flowchart illustrating a continuous implicit neural super-resolution method for sparse-view CT sinusoidal images provided in this embodiment of the disclosure. Figure 1 and Figure 2 As shown, the method mainly includes the following steps: Step 1: Obtain sparse view CT projection data of the target object and construct the input sparse sine curve according to the order of the sampled angles; In practice, sparse view CT projection data of the target object can be obtained and the input sparse sine curve can be constructed according to the order of the sampled angles.

[0032] Step 2: Input the input sparse sinusoidal graph into the edge-aware encoder to extract multi-directional high-frequency mutation features and obtain high-frequency enhanced feature representation; In specific implementation, such as Figure 4As shown, a sparse sinusoidal graph can be input into an edge-aware encoder, and abrupt responses can be extracted using multi-directional anisotropic masks. High-frequency enhanced features are then obtained through the residual backbone. Since the sinusoidal graph contains both smooth angular variation components and abrupt structures caused by high-contrast anatomical boundaries, directly using a general implicit renderer can easily over-smooth local abrupt changes. Therefore, this invention incorporates an explicit edge extraction module at the encoding end, preferably employing multiple convolutional masks with fixed directions and different aspect ratios, such as 7×3, 3×7, 7×7, 5×5, 5×3, and 3×5, to capture abrupt responses in the projection domain from different directions. These responses are then fused with a learnable convolutional layer and a residual backbone to obtain the enhanced high-frequency feature representation.

[0033] Its high-frequency characteristics can be represented as: .

[0034] Optionally, the encoder consists of an explicit edge extraction layer and an EDSR-style residual backbone. The decoder employs a lightweight 1×1 convolutional structure, with one branch outputting Fourier coefficients. and Another output velocity stream or offset field The residual rendering is preferably performed locally in the angular dimension, rather than sharing a global implicit function across the entire sine curve.

[0035] Step 3: Input the high-frequency enhancement feature representation into the decoder, predict the Fourier residual parameters by the first branch of the decoder, and predict the detector axis offset field at each sampled angle position by the second branch of the decoder. In practical implementation, the decoder header structure is as follows: Figure 5 As shown, the high-frequency enhanced features can be input into the decoder, which outputs the Fourier residual parameters and the detector axis offset field, respectively.

[0036] Step 4: Based on the relative position of the target angle within the adjacent sampled angle intervals, and combined with the detector axis offset field, calculate the forward trajectory sensing distortion coordinates and the backward trajectory sensing distortion coordinates respectively. In practice, the trajectory perception distortion coordinates in the forward and backward directions can be calculated based on the relative position of the target angle between adjacent sampling angles.

[0037] Specifically, considering that target angle recovery between adjacent sampling angles should not simply be taken at the same detector coordinates, because according to the Radon projection mechanism, the position mapped to the detector by the same image domain structure changes at different scanning angles. For a fixed point in the image domain, its corresponding detector position at different angles will exhibit a sinusoidal drift along the detector direction. To model this physical phenomenon, this invention learns a detector axis offset field for each adjacent sampling angle interval. For the relative position of the target angle between two adjacent sampling angles, this invention constructs forward and backward trajectory warping coordinates respectively, so that the rendered query point adaptively moves along the detector axis as the angle changes.

[0038] According to Radon's projection mechanism, projection data can be expressed as:

[0039] The corresponding structural trajectory expression is:

[0040] Its derivative with respect to angle is:

[0041] Let the relative position of the target angle between adjacent sampling angles be:

[0042] The coordinates of the forward and backward trajectory distortions are as follows:

[0043] .

[0044] Step 5: Based on the forward trajectory sensing twisted coordinates and the backward trajectory sensing twisted coordinates, generate the forward residual and the backward residual respectively using the Fourier residual parameters; In practice, forward and backward residuals can be generated using Fourier residual parameters under distorted coordinates.

[0045] After establishing a trajectory-aware coordinate system, this invention continuously models the changes in projection intensity within local angle intervals. The Fourier branch in the decoder outputs several sets of cosine and sine coefficients for each local anchor point, constituting local Fourier residual parameters, and explicitly calculates the residual terms at any relative angle position within the interval. To reduce artifacts caused by single-direction rendering, this invention employs a bidirectional residual rendering strategy: first, basic interpolation is performed on two adjacent sampling angles; then, the forward and backward branches predict the residuals respectively; finally, the trajectory distortion results are combined for weighted fusion to obtain the final target angle projection.

[0046] The Fourier residual term can be expressed as: .

[0047] in, This represents the magnitude of the cosine of the Fourier series. This represents the magnitude of the sine wave in the Fourier series. Indicates a predefined frequency reference; When calculating the forward residual, the detector coordinates in the residual term expression are used. Replace with forward trajectory-aware warped coordinates ,get: ; When calculating the backward residual, the detector coordinates in the residual term expression are... Replace with the backward trajectory-aware warp coordinates and relative position Replace with ,get: .

[0048] Optionally, to enhance the symmetry and stability of continuous rendering, the residual calculation employs simultaneous forward and backward branches, weighted by the relative position of the target angle within the interval. This design better maintains the structural continuity between adjacent angles.

[0049] Step 6: Perform basic interpolation on the projection values ​​of adjacent sampled angles to obtain the basic interpolation results; In practice, basic interpolation can be performed on the projected values ​​of adjacent sampled angles to obtain the basic interpolation result: .

[0050] Step 7: The basic interpolation results are weighted and fused with the forward and backward residuals according to their relative positions to generate the restored projection at the target angle; In practice, after obtaining the basic interpolation results, the basic interpolation results and the two-way residuals can be fused according to their relative positions to generate the restored projection at the target angle: .

[0051] Step 8: Repeat steps 1 to 7 for multiple target angles to output a continuous dense sine curve.

[0052] Furthermore, after step 8, the method further includes: A continuous dense sine wave is input into a filtered back-projection reconstruction algorithm or an iterative reconstruction algorithm to generate a CT image of the target object.

[0053] In practice, the above steps are repeated for multiple target angles to output a continuous dense sine wave, which can then be further processed by filtered back projection or other reconstruction algorithms to generate CT images. The method disclosed herein treats sparse view projection restoration as a continuous rendering problem in the angular dimension. Let the input sparse sine wave be X. The goal is not simply to restore a fixed number of missing viewpoints, but to establish a continuous function representation that can be queried and generated at any target angle position, thereby outputting a dense sine wave with arbitrary angular density.

[0054] This embodiment provides a continuous implicit neural super-resolution method for sparse-view CT sinusoidal images. By taking the sparse sinusoidal image as input, it first extracts multi-directional, high-frequency abrupt features through an explicit edge-aware encoder, and then uses a dual-branch decoder to simultaneously predict Fourier residual parameters and detector axis offset fields. Subsequently, based on the relative position of the target angle between adjacent sampled angles, it performs trajectory-aware distortion on the rendered coordinates and fuses the bidirectional residuals with the basic interpolation results to finally obtain a continuous dense projection under the target angle.

[0055] like Figure 3 As shown, compared with the prior art, the present invention does not simply perform fixed grid interpolation on the missing projection, nor does it merely learn the discrete mapping between viewpoints. Instead, it uses a continuous representation to learn the intrinsic law of projection changing with angle, and connects this law with the real scanning geometry of CT through sensor offset modeling.

[0056] The method of this application will be further described below with reference to a specific embodiment.

[0057] In one example implementation, the reconstructed slices are first uniformly adjusted to a size of 256×256 and normalized. Then, based on two-dimensional fan-beam imaging geometry, the image domain is discretized within the range of [-128, 128] × [-128, 128], and a dense supervisory sine curve is constructed using 360 uniform viewing angles and 512 detector channels; both the source radius and detector radius can be set to 500. Preferably, forward projection can be achieved using the ODL library combined with the ASTRA backend, and image domain reconstruction evaluation can be performed using Ram-Lak filtered FBP.

[0058] In a preferred training implementation, the AAPM-Mayo chest CT dataset and the Brain Stroke CT dataset can be used for training and validation, with the AAPM dataset containing 8 training samples and 2 test samples. Training can be performed using the PyTorch framework, the Adam optimizer, an initial learning rate of 1×10^-4, a batch size of 6, and 50 epochs. The preferred evaluation metrics are PSNR and SSIM, with the image domain PSNR of the reconstructed image obtained from the restored projection being used as a supplementary evaluation metric.

[0059] like Figure 5 , Figure 6 , Figure 7 and Figure 8 As shown in the existing experiments, the present invention achieves excellent results on various settings of the AAPM and Brain Stroke datasets. In existing scale recovery tasks, CoinS outperforms in settings such as 45→180, 45→360, and 30→90. In cross-scale generalization tasks, especially when trained from 45→90 and tested to 180 and 360 viewpoints, and trained from 30→90 and tested to 120 / 180 / 360 viewpoints, the present invention can significantly improve the PSNR in both the projection domain and the image domain. This indicates that the present invention does not memorize interpolation relationships with fixed multiples, but rather truly learns stable continuous representations.

[0060] Further ablation experiments showed that the edge enhancement module mainly improves the recovery of local details at the known scale, the bidirectional rendering module mainly enhances cross-angle continuity, and the sensor offset learning module is the key factor in achieving cross-scale generalization capability. In tests where no 360-degree view target was seen, after introducing sensor offset modeling, the projection domain PSNR and image domain PSNR on the AAPM data increased from 41.3883 / 28.4940 to 45.5263 / 31.4271, and the corresponding indicators on the Brain Stroke data also showed significant improvement.

[0061] The quantitative experimental results on the AAPM and Brain Stroke datasets are shown in Table 1 (PSNR: mean / variance; SSIM: mean / standard deviation × 100).

[0062] Table 1

[0063] The results of the continuous representation cross-scale generalization experiment are shown in Table 2, where the first index is the projection domain PSNR and the second index is the image domain PSNR.

[0064] Table 2

[0065] The ablation experimental results of the method disclosed herein under the known scale settings are shown in Table 3.

[0066] Table 3

[0067] The ablation experimental results of the method disclosed herein under no-scale settings are shown in Table 4, where each result consists of projection domain PSNR and image domain PSNR.

[0068] Table 4

[0069] Table 5 shows the comparison results of inference FLOPs and PSNR for different projection restoration models under the 45->360 viewing angle settings.

[0070] Table 5

[0071] Compared with the prior art, the present invention has at least the following beneficial effects: Based on the real continuous scanning mechanism of CT, a more physically consistent continuous representation model can be established for sparse view projection recovery.

[0072] Explicit edge enhancement improves the recovery of abrupt structures, high-frequency details, and boundary transitions, while reducing over-smoothing.

[0073] By using sensor offset learning and trajectory perception distortion, the detector position drift of the same structure as the angle changes is explicitly modeled, improving the angular continuity and structural consistency of the recovery results.

[0074] By employing a two-way residual fusion mechanism, the accuracy of local interval interpolation and global continuity are balanced, reducing grid-like artifacts and discontinuous transitions.

[0075] It exhibits strong generalization performance at both seen and unseen scales, and is particularly suitable for continuous query scenarios where the target angle density is not completely consistent with that of the training phase.

[0076] Based on existing experimental results, compared with the comparison method, the AAPM and Brain Stroke datasets can achieve better projection domain PSNR and image domain PSNR, and the advantage is more obvious in cross-scale tasks without target viewpoints.

[0077] This disclosed method addresses the shortcomings of existing projection restoration schemes, such as lack of continuous scan modeling, neglect of detector coordinate drift, and tendency to over-smooth. It proposes a combined approach utilizing explicit edge enhancement coding, sensor offset learning, trajectory-aware coordinate distortion, and Fourier residual bidirectional rendering to achieve continuous projection restoration at any target angle. This method improves the detail fidelity, angular continuity, and downstream image reconstruction quality of sparse-view CT projection restoration, demonstrating significant engineering application value.

[0078] Corresponding to the above method embodiments, this disclosure also provides a continuous implicit neural super-resolution system for sparse-view CT sinusoidal images, mainly comprising: Data input module: used to receive sparse sine graphs or projection sequences formed by original projection sampling.

[0079] Edge enhancement coding module: used to extract multi-directional mutation features and form high-frequency enhanced representations.

[0080] Continuous parameter decoding module: used to output Fourier residual parameters and detector axis offset field.

[0081] The trajectory distortion rendering module is used to generate trajectory-aware query coordinates based on the target angle and offset field, and to complete bidirectional residual rendering.

[0082] Projection fusion module: used to fuse the basic interpolation with the two-way residual to output the restored projection of the target angle.

[0083] Image reconstruction module: Used to feed the restored dense projection into FBP or other reconstruction algorithms to generate the target CT image.

[0084] Optionally, the edge extraction module can be replaced with other directional filter banks or learnable edge prior modules that can explicitly enhance high-frequency information in the projection domain; the backbone network can be replaced with other residual networks, convolutional networks, or Transformer backbones; the continuous residual modeling can be replaced with other periodic basis functions, orthogonal basis functions, or locally explicit representations; and the image reconstruction module can be replaced with FBP, iterative reconstruction, or deep reconstruction networks. As long as the core idea of ​​"continuous angle representation + trajectory-aware offset modeling + high-frequency enhancement" is retained, it should be considered to fall within the protection scope of this application.

[0085] The system can execute the contents of the above method embodiments accordingly. For the parts not described in detail in this embodiment, please refer to the contents recorded in the above method embodiments, and they will not be repeated here.

[0086] It should be understood that the various parts of this disclosure can be implemented in hardware, software, firmware, or a combination thereof.

[0087] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A continuous implicit neural super-resolution method for sparse-view CT sinusoidal images, characterized in that, include: Step 1: Obtain sparse view CT projection data of the target object and construct the input sparse sine curve according to the order of the sampled angles; Step 2: Input the input sparse sinusoidal graph into the edge-aware encoder to extract multi-directional high-frequency mutation features and obtain high-frequency enhanced feature representation; Step 3: Input the high-frequency enhancement feature representation into the decoder, predict the Fourier residual parameters by the first branch of the decoder, and predict the detector axis offset field at each sampled angle position by the second branch of the decoder. Step 4: Based on the relative position of the target angle within the adjacent sampled angle intervals, and combined with the detector axis offset field, calculate the forward trajectory sensing distortion coordinates and the backward trajectory sensing distortion coordinates respectively. Step 5: Based on the forward trajectory sensing twisted coordinates and the backward trajectory sensing twisted coordinates, generate the forward residual and the backward residual respectively using the Fourier residual parameters; Step 6: Perform basic interpolation on the projection values ​​of adjacent sampled angles to obtain the basic interpolation results; Step 7: The basic interpolation results are weighted and fused with the forward and backward residuals according to their relative positions to generate the restored projection at the target angle; Step 8: Repeat steps 1 to 7 for multiple target angles to output a continuous dense sine curve.

2. The method according to claim 1, characterized in that, The edge-aware encoder includes an explicit edge extraction layer and a residual backbone network. The explicit edge extraction layer uses multiple anisotropic convolutional masks with different directions and aspect ratios to extract multi-directional abrupt responses. The residual backbone network fuses and enhances the multi-directional abrupt responses to obtain high-frequency enhanced feature representations.

3. The method according to claim 2, characterized in that, The anisotropic convolution mask includes various combinations of 7×3, 3×7, 7×7, 5×5, 5×3 and 3×5 directional masks.

4. The method according to claim 1, characterized in that, The first branch of the decoder adopts a 1×1 convolution structure, outputting several sets of cosine coefficients and sine coefficients for each local anchor point, which constitute the Fourier residual parameters. The second branch of the decoder adopts a 1×1 convolution structure and outputs the detector axis offset field.

5. The method according to claim 1, characterized in that, The forward trajectory sensing distortion coordinates and the backward trajectory sensing distortion coordinates are determined in the following way: Set target angle Located at adjacent sampled angles and Between, relative position The detector coordinates are The detector axis offset field is and ,but: The forward trajectory perception distortion coordinates are: ； The backward trajectory perception distortion coordinates are: 。 6. The method according to claim 5, characterized in that, The Fourier residual parameters are modeled on a local angle interval basis. The residual term at any relative position t within each local angle interval is explicitly calculated, and the expression for the residual term is: in, This represents the magnitude of the cosine of the Fourier series. This represents the magnitude of the sine wave in the Fourier series. Indicates a predefined frequency reference; When calculating the forward residual, the detector coordinates in the residual term expression are used. Replace with forward trajectory-aware warped coordinates ,get: ； When calculating the backward residual, the detector coordinates in the residual term expression are... Replace with the backward trajectory-aware distortion coordinates and relative position Replace with ,get: 。 7. The method according to claim 6, characterized in that, The expression for the basic interpolation result is: in, This represents the projected value of the previously measured angle. This represents the projected value of the next measured angle; The expression for the restored projection at the target angle is: in, This represents the repaired residual from the projection of the previous measurement angle. This represents the repaired residual from the projection of the next measurement angle.

8. The method according to claim 1, characterized in that, After step 8, the method further includes: A continuous dense sine wave is input into a filtered back-projection reconstruction algorithm or an iterative reconstruction algorithm to generate a CT image of the target object.

Images