Sparse-view ct imaging method and system based on multi-level sub-imaging system

Abstract

The application discloses a sparse-view CT imaging method and system based on a multi-level sub-imaging system, and comprises the following steps: acquiring a sparse sinogram collected by a CT imaging device according to a low-dose scanning protocol; determining a sparse factor and a level number according to a complete scanning protocol as a reconstruction target; reducing the dimension of the sparse sinogram along a detector axis direction step by step to construct a plurality of levels of physical projection data and corresponding sub-imaging operators; constructing a perception projection feature through a Radon perception sinogram decomposer; performing initial reconstruction at the coarsest level; performing projection data consistency correction through a physical flow, keeping multi-channel latent features through a perception flow, and performing spatial domain and frequency domain interactive modulation in an unfolding block in the order from coarse to fine; and finally outputting a CT reconstruction image for diagnostic observation. The application can reduce the CT radiation dose while improving the imaging accuracy, stability and out-of-domain robustness of the sparse projection view CT reconstruction.

Description

Technical Field

[0001] This invention relates to the field of medical CT imaging technology, and more specifically to a sparse-view CT imaging method and system based on a multi-level sub-imaging system. Background Technology

[0002] Computed tomography (CT) is a commonly used medical imaging technique in modern clinical diagnosis. CT diagnostic imaging equipment typically acquires multi-angle projection data around the subject using an X-ray source and detector, and reconstructs tomographic images of the subject based on this projection data. Conventional CT scans usually require acquiring a relatively dense projection field of view to obtain high-quality images, but higher doses of X-rays increase the radiation exposure of the subject. Therefore, how to obtain CT reconstructed images that meet diagnostic requirements while reducing radiation dose is an important issue in the field of CT diagnostic imaging.

[0003] Sparse-view CT reduces scanning dose by decreasing the number of projection views. Under low-dose scanning protocols, the actual number of projection views acquired is significantly less than that corresponding to a full scanning protocol, resulting in incomplete projection data and underdetermined imaging characteristics. If conventional analytical reconstruction methods such as filtered backprojection are directly used, obvious fringe artifacts and loss of detail often appear in the reconstructed images, affecting lesion observation and subsequent diagnostic analysis.

[0004] In existing technologies, model-based iterative reconstruction methods can suppress noise and artifacts to some extent through regularization constraints, but they have high computational costs, complex parameter tuning, and difficulty in balancing imaging efficiency and reconstruction quality. Deep learning methods, when used for low-dose CT or sparse-view CT reconstruction, can improve image quality. Among them, deep unfolding networks, by unfolding the traditional iterative optimization process into multiple learnable network blocks and embedding projection data consistency correction in each network block, can achieve a certain balance between physical consistency and image perceptual quality.

[0005] However, existing depth unfolding networks still have the following shortcomings in sparse-view CT imaging. First, the iterative trajectory stability of existing depth unfolding networks is insufficient. Depth unfolding networks typically assume that each data consistency correction step can provide physical constraints for intermediate reconstruction results. However, in sparse-view CT, projection observations only cover a portion of the entire imaging system's spatial range, and large-scale null space components still exist. A single unfolded block struggles to stably infer null space components from limited spatial observations. Small random factors such as training noise and weight initialization are easily amplified along the unconstrained null space direction, causing oscillations such as overshoot and undershoot in the output of intermediate unfolded layers relative to the target image. These oscillations not only reduce model interpretability but also increase the risk of the model converging to a suboptimal solution.

[0006] Secondly, existing deep unfolding networks suffer from information bottlenecks between unfolded blocks. To perform sinusoidal domain correction with measured projection data, most deep unfolding networks repeatedly compress high-dimensional latent features into single-channel CT images between each iteration block, then re-input the corrected image into subsequent network blocks for feature extraction. While this process facilitates the enforcement of physical consistency constraints, it weakens the long-range propagation capability of high-dimensional perceptual features, requiring subsequent network blocks to repeatedly extract perceptual information such as texture, edge, and artifact patterns. Furthermore, since a single-channel image simultaneously carries projection data consistency information and perceptual representation information, these two elements easily compete within the limited information channels, further exacerbating the instability of the unfolding iteration.

[0007] Third, existing depth unfolding networks lack robustness in out-of-distribution (OOD) scenarios. Sparse-view CT reconstruction models are typically trained on data from specific locations or with specific noise distributions, making it easy for the network to learn the overall artifact distribution mapping associated with the training data. When the scanned location or noise level of the test subject differs from the training data—for example, training on chest and abdominal data but testing on head CT data—existing methods, especially depth unfolding methods using homogeneous unfolded blocks in series, are prone to artifact removal failures or insufficient detail recovery. Therefore, existing methods still struggle to simultaneously meet the stability, accuracy, and generalization requirements of sparse-view CT imaging. Summary of the Invention

[0008] The purpose of this invention is to provide a sparse-view CT imaging method and system based on a multi-level sub-imaging system, so as to solve the problems of unstable reconstruction iteration trajectory, limited inter-block sensing information transmission, and insufficient out-of-domain robustness under low-dose sparse-view CT acquisition.

[0009] To achieve the above objectives, this invention proposes a sparse-view CT imaging method based on a multi-level sub-imaging system, comprising the following steps:

[0010] Step S1: Control the CT diagnostic imaging equipment to perform sparse view projection acquisition on the subject according to the low-dose scanning protocol, receive the sparse projection data acquired by the CT diagnostic imaging equipment, and form a sparse sine curve based on the sparse projection data. ;

[0011] Step S2: Record the number of projection viewpoints corresponding to the complete scan protocol as The actual number of projection angles acquired by the low-dose scanning protocol is ,according to Determine the sparsity factor and according to Determine the number of levels in the null space decomposition ;

[0012] Step S3: Along the sparse sine graph The detector axis is subjected to stepwise dimensionality reduction processing to obtain physical projection data at multiple levels. Configure sub-imaging operators corresponding to the physical projection data of each level. This forms a virtual CT sub-imaging system that progresses from coarse to fine; where the hierarchical index... Indicates the target resolution level. This indicates the level with the coarsest detector grid and the lowest image resolution;

[0013] Step S4: Process the sparse sine graph using the Radon-perceptual sine graph decomposer. Latent space decomposition is performed to obtain perceptual projection features corresponding to multiple levels. ;

[0014] Step S5: Apply the physical projection data of the coarsest level respectively. and perceptual projection features Perform filtered backprojection to obtain the initial physical flow image and initial perceptual flow latent features; upsample the initial physical flow image and initial perceptual flow features to the first reconstruction level, and then, according to the coarse-to-fine level order, perform further refactoring at the current level. Layer Within each expansion block, corrections are first performed on the physical flow and the sensory flow separately, where the physical flow is addressed using sub-imaging operators. Perform a forward projection on the current physical flow image estimate, and then compare the obtained forward projection with the physical projection data of that layer. The differences between them serve as the basis for projection data consistency correction; the perceptual stream projects the current multi-channel latent features channel by channel using the same sub-imaging operator, and the resulting latent space projection is compared with the perceptual projection features of that layer. The differences between them serve as the basis for latent space correction;

[0015] Step S6: After the physical flow and sensory flow of the current unfolded block have completed the corrections, the features of the physical flow and the features of the sensory flow are subjected to spatial and frequency domain interactive modulation to fuse the projection physical constraints and sensory feature constraints, and to suppress CT reconstruction artifacts caused by sparse view acquisition. If there are still unfolded blocks that have not been executed at the current level, the next unfolded block is entered, and the processing of correcting the two flows in step S5 and the interactive modulation in step S6 are repeated. If all unfolded blocks at the current level have been executed and the current level is not the finest level, the interactively modulated physical flow image estimation and sensory flow latent features are upsampled to the next finer level, and the processing of correcting the two flows in step S5 and the interactive modulation in step S6 are repeated.

[0016] Step S7: After the finest level of processing is completed, a CT reconstructed image is generated and output as a medical image for diagnostic observation, display, transmission or storage.

[0017] Furthermore, in S3, the physical projection data of layer 0 is... , No. Layer physical projection data is derived from the formula Confirmed, among which This represents a one-dimensional average pooling operator performed along the detector axis with a factor of 2. .

[0018] Furthermore, in S3, the first Layer CT sub-imaging system meets ,in Indicates the first The corresponding CT image of the layer, This represents a sparse CT projection operator with a coarser detector grid, and the... The zero spatial dimension of the layer imaging system satisfies ,in Indicates the first Layer detector resolution, and It is a constant.

[0019] Furthermore, in S5, in the first... Layer Within each expanded block, the physical flow is estimated using the current physical flow image. As initial value For the normal equation implement Step conjugate gradient iteration yields the truncated Krylov subspace trajectory. ;according to Construct the filtered back projection approximate residual; and according to Update the physical flow image estimate, where This represents the filtering back projection operator. Indicates the first Layer The physical flow of each expanded block updates the network. , and The subscripts represent the conjugate gradient iteration, the Krylov subspace trajectory residual generated during the iteration, and the filtered back projection residual, respectively.

[0020] Furthermore, in S4, the Radon-sensing sine wave decomposer follows... as well as Generate multi-level perceptual projection features, where Indicates the first Layer-learnable Radon-aware decomposition operator, .

[0021] Furthermore, in S5, in the first... Layer Within each expanded block, the sensing flow maintains multi-channel latent features. And based on the current level, perceive the projection features. With multi-channel latent features Correction is performed on the latent space projection residuals between the projection results, and the latent features of the sensed flow are corrected according to... Update, in which This indicates that the perception flow updates the network.

[0022] Furthermore, in S6, the spatial domain and frequency domain interactive modulation includes spatial domain cross modulation and amplitude-phase cross modulation. Specifically, spatial domain cross modulation uses the spatial components of the sensing stream to modulate the spatial components of the physical stream, and uses the spatial components of the physical stream to gate the spatial components of the sensing stream. Amplitude-phase cross modulation uses the frequency-phase information of the physical stream to constrain the structural alignment of the sensing stream, and uses the frequency-amplitude information of the sensing stream to generate dynamic attenuation information for suppressing sparse view artifacts.

[0023] Furthermore, it also includes: imposing supervised constraints on the physical flow image estimation at each intermediate layer during training of the network used to perform physical flow updates, perceptual flow updates, and interactive modulation, wherein the training loss satisfies ,in Indicates the first Number of layer expandable blocks Indicates the relationship with the first Reference CT image corresponding to slice resolution, Indicates the hierarchical weight.

[0024] Furthermore, it also includes: a sparse-view CT imaging system based on a multi-level sub-imaging system, characterized in that it includes a CT diagnostic imaging device and an image reconstruction unit; the CT diagnostic imaging device includes an X-ray source, a detector, and a scanning control unit, the scanning control unit being used to control the X-ray source and the detector to perform sparse-view projection acquisition on the subject according to a low-dose scanning protocol to obtain sparse projection data; the image reconstruction unit is disposed within the CT diagnostic imaging device, or disposed within a medical image reconstruction workstation communicatively connected to the CT diagnostic imaging device; the image reconstruction unit includes a processor and a memory, the memory storing a computer program, when the computer program is executed by the processor, causing the image reconstruction unit to perform the method described in any one of claims 1 to 8 based on the sparse projection data, generating a CT reconstructed image for diagnostic observation, display, transmission, or storage.

[0025] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a virtual CT sub-imaging system that progresses from coarse to fine, this invention allows the reconstruction process to begin at a coarse level with low null space uncertainty, anchoring the unfolding trajectory to a sub-imaging system with clear physical meaning, thereby improving iterative stability and model interpretability. Furthermore, this invention no longer merely learns the overall artifact distribution mapping in the training data, but progressively refines frequency-related interpolation and artifact patterns at each level, reducing dependence on specific training sites or specific scan distributions and improving the model's robustness in reconstructing extra-domain anatomical structures. Finally, this invention sets up the physical flow and the perceptual flow in parallel, where the multi-channel latent features of the perceptual flow avoid the model repeatedly compressing and re-extracting high-dimensional perceptual information between unfolding blocks, alleviating the single-channel image domain bottleneck in data consistency correction. This invention can improve the imaging accuracy, stability, and extra-domain robustness of sparse projection perspective CT reconstruction while reducing CT radiation dose. Attached Figure Description

[0026] Figure 1 This is a flowchart of a sparse-view CT imaging method based on a multi-level sub-imaging system according to an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of a zero-space decomposition and multi-level virtual CT sub-imaging system provided according to an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of a two-stream unfolded reconstruction network structure provided according to an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram of the Radon sensing sine wave decomposer structure provided in an embodiment of the present invention.

[0030] Figure 5 This is a schematic diagram of the sparse-view CT imaging system provided in an embodiment of the present invention.

[0031] Figure 6 This is a schematic diagram comparing the reconstruction results of the sparse-view CT imaging method based on a multi-level sub-imaging system provided by the present invention with existing methods in lung and head cases.

[0032] Figure 7 This is a schematic diagram comparing the reconstruction results and lesion segmentation results of the sparse-view CT imaging method based on a multi-level sub-imaging system provided by the present invention with those of existing methods on different cases of novel coronavirus. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0034] It should be noted that, unless otherwise specified, the following embodiments and the technical features thereof can be combined with each other. Convolution operations, attention operations, interpolation operations, and CT image preprocessing operations not specifically described in the following embodiments can all be implemented using conventional methods in the art.

[0035] This invention provides a sparse-view CT imaging method based on a multi-level sub-imaging system. This method can be executed by an image reconstruction unit within the sparse-view CT imaging system. The image reconstruction unit can be located inside a CT diagnostic imaging device or within a medical image reconstruction workstation that is communicatively connected to the CT diagnostic imaging device. The following explanation uses two-dimensional tomographic image reconstruction as an example. For a fan-beam sine curve formed by rearranging helical scan data, the same processing can be performed slice by slice, and consecutive slices can be combined into a three-dimensional CT image.

[0036] like Figure 1 As shown, the scanning control unit controls the X-ray source and detector to perform sparse viewpoint projection acquisition on the object under inspection according to a low-dose scanning protocol, obtaining sparse projection data. The image reconstruction unit arranges the sparse projection data into a sparse sine curve according to the projection viewpoint and detector channel. In one embodiment, the number of projection viewpoints corresponding to the complete scanning protocol is: The actual number of projection angles acquired by the low-dose scanning protocol is Then the sparsity factor is The number of levels in the null space decomposition is .in, This indicates rounding up, therefore To meet The smallest integer of K; when K is not an integer power of 2, the grid size of the final stage detector is determined according to the actual achievable dimensionality reduction size.

[0037] like Figure 2 As shown, the image reconstruction unit for sparse sinusoids One-dimensional average pooling is performed along the detector axis to generate physical projection data at multiple levels. .therefore, Indicates the first Single-channel physical sine wave observation of the layer, retaining the original low-dose scanning protocol. The projection viewpoint, and has the same projection viewpoint as the first. The number of detector channels corresponding to the layer detector grid. Since axial average pooling of the detector can be approximated as applying a radial low-pass window to the spectrum of the imaged object, coarser layers mainly retain relatively stable low-frequency structures. Under the parallel beam approximation, the projection... The relationship between the one-dimensional Fourier transform and the two-dimensional Fourier transform of an image is as follows:

[0038]

[0039] in, Indicates the projection angle. Represents radial spatial frequency. and These represent the one-dimensional Fourier transform and the two-dimensional Fourier transform, respectively. For the complete scan protocol, along the frequency loop radius... The sampling point density is as follows:

[0040]

[0041] in, Indicates the complete scan protocol in radius The frequency domain sampling point density at a given point. With a sparsity factor of . Under the low-dose scanning protocol, the sampling point density is as follows:

[0042]

[0043] If the complete scan sampling density at the Nyquist radius is taken as the minimum sampling density without aliasing, then from The critical frequency can be obtained by sampling at a density no lower than this minimum. Therefore, even if the projection angle is sparsely extracted, a relatively stable low-frequency region still exists in the frequency domain. The detector axis is subjected to stepwise average pooling to progressively reduce the effective frequency band, ensuring that the coarsest level primarily falls within the relatively stable low-frequency region, rather than directly recovering all frequency components simultaneously in the original underdetermined system. For the fan-beam sine curve obtained from rearranged clinical spiral scan data, the following steps are performed: When the detector is subjected to axis-average pooling, its frequency domain relationship is as follows:

[0044]

[0045] in, express Axial average pooling for multiple detectors This represents the approximate low-pass frequency response corresponding to average pooling. This represents the higher-order structural residuals introduced by the fan beam geometry. The aforementioned frequencies... Normalization is performed according to the detector sampling interval, and the normalized Nyquist angular frequency is taken as... and in The sampling density is discussed within the frequency range; simultaneously, it can be utilized... The sparsity factor is obtained as The critical frequency of the imaging system at that time satisfies This invention employs a step-by-step dimensionality reduction along the detector axis, equivalent to constructing a set of progressively shrinking low-pass windows, allowing the null space component to be decomposed gradually from high to low frequencies. The higher-order structural residuals are further corrected by neural networks in the expanded blocks of each level. Notably, within the low-frequency range of interest, average pooling approximates a low-pass response, where... The term represents an approximate representation of the low-pass response, with higher-order aliasing, fan-beam geometry effects, and boundary effects caused by discrete downsampling uniformly included in the residual. For the first The cumulative dimensionality reduction of the layers relative to the original detector grid can be taken as... ; lowercase This represents the average pooling factor.

[0046] For the Layer physical projection data, and its corresponding sub-imaging operator configured with image reconstruction units. This constitutes an approximate sub-imaging system. Each sub-imaging system can be understood as a projection system obtained by a virtual CT scan system with a coarser detector grid. As the detector resolution... Reduced, the degree of freedom in the image domain is approximately Reduced in size, while the measurement dimensions are approximately... Shrink, therefore the first Zero spatial dimension of layer imaging system The conditions are as follows:

[0047]

[0048] in, and It is a constant. Representational Imaging Operator The null-space dimension. This relationship indicates that, in a coarser grid, the unobserved components to be recovered decrease faster than the image resolution, resulting in smaller null-space uncertainty for the coarsest level relative to the target resolution level, thus providing a stable starting point for coarse-to-fine reconstruction. Specifically, in this embodiment, the side length of each image grid layer is reduced proportionally to the number of detector channels, therefore the image domain degrees of freedom and measurement dimension are respectively reduced proportionally. and change, and To be related to the discretization method and scan geometry and independent of the layer The constant of change.

[0049] In one embodiment, when the sparsity factor hour, This results in four scales: target resolution, half-resolution, quarter-resolution, and eighth-resolution. The coarsest layer imaging system is only used for initialization and does not set up unwrap blocks. Image reconstruction units are arranged according to... For the coarsest level physical projection data Perform filtered backprojection to obtain the initial low-frequency CT image of a single channel. At the same time, according to Filtered backprojection is performed on the sensing projection features of each channel at the coarsest level to obtain the initial latent features of the multi-channel sensing flow. The former is the image domain starting point consistent with the low-dose measured projection, while the latter is the image domain starting point where the perceptual flow preserves the potential observation information of the sinusoid. Subsequently, for... Perform bicubic upsampling, for Perform learnable upsampling to obtain the first... Input of the first expanded block of the layer and .

[0050] like Figure 3 As shown, the image reconstruction unit follows The sequential execution of coarse-to-fine dual-stream expansion reconstruction, the first... Layers include The first unrolled block is executed sequentially. In the... Layer In each expanded block, the input includes physical flow image estimation. Multi-channel latent features of sensing flow Physical projection data of this layer The layer perceives projection features and the sub-imaging operator matched with this layer The unfolded block first performs physical flow data consistency correction and sensor flow latent space correction separately, and then performs spatial and frequency domain interactive modulation on the correction results of the two flows. The interactively modulated features of the two flows are then used as inputs for the next unfolded block or the next finer level.

[0051] For ease of explanation, the following textual descriptions and formulas, Indicates a hierarchical index. Indicates the index of the expanded block within the layer. Indicates the conjugate gradient iteration index. Indicates the first Number of layer expandable blocks Indicates the number of steps to cut off the conjugate gradient; lowercase. , Representing single-channel image and physical projection respectively, uppercase. , These represent multi-channel latent features and perceptual projection features, respectively; superscript , and These represent physical flow, sensory flow, and sparse view acquisition, respectively. and These represent the spatial components and frequency components, respectively. The reference target is indicated by the cap number, and the modulated variable is indicated by the cap number. and Let L1 norm and L2 norm be represented respectively. This indicates element-wise multiplication. This indicates splicing along the channel.

[0052] like Figure 3 and Figure 4 As shown, the Radon perceptual sinogram decomposer first decomposes the original sparse sinogram... Mapped to Layer 0 perceptual projection features Then, through the learnable Radon-aware decomposition operator Generation of projection features at each stage for perceptual flow correction Since the sine curve exhibits long-range correlation along both the projection viewpoint axis and the detector axis, for coarser levels with lower resolution, the decomposition operator... Features can be first expanded along the projection viewpoint axis, and the relationship between different angle projections can be modeled using rotational position encoding and one-dimensional self-attention. Then, the relationship between different detector positions within the same projection can be modeled in the same way along the detector axis. For... or For finer layers with higher resolution, the following can be used: and Factorized large kernel convolutions, composed of convolutions of varying sizes, expand the receptive field along both axes while reducing computational cost.

[0053] like Figure 3 As shown, in the first Layer Within each expanded block, the physical flow is estimated using the current physical flow image. As initial values, and simultaneously executed using the normal equations corresponding to the current sub-imaging operator for this layer. Step S3: Conjugate gradient iteration. The conjugate gradient iteration always uses the gradient obtained in step S3 as the starting point. Layer physical projection data With the goal of reducing the squared error between observations from the corresponding sparse viewpoint, each step is used to reduce the squared error. Obtain the truncated Krylov trajectory composed of the image estimates at each step. And form a set of residuals between the trajectory and the initial estimate. Additionally, the model calculates the difference between the predicted projection and the target projection. Then perform filtering and back projection to obtain Because too few conjugate gradient steps may result in insufficient data correction, while too many steps may introduce high-frequency projection noise into the image, this embodiment does not only use the conjugate gradient result of the last step, but also concatenates the current estimate, the trajectory of each step, the relative trajectory residual, and the filtered back-projection approximate residual along the channels and inputs them together into the physical flow update network. This enables the physical flow update network to select effective components from candidate information with different correction levels and suppress overcorrection or noise amplification.

[0054] In parallel with the physical flow, the perceptual flow receives the current multi-channel image domain latent features. and the Layer-aware projection features The perceptual flow will use the same sub-imaging operator. Acting independently For each feature channel, obtain the multi-channel predicted projection. Then perceive the projection features at the current level. Calculate the latent space projection residual for the target The residual represents the difference between the predicted projections of the current sensing latent features on each feature channel and the Radon sensing latent observations at that layer. Channel-by-channel filtering and backprojection of this residual yields the image domain latent space correction cue. And input the perceptual flow to update the network. Since performing multiple conjugate gradient iterations within a multi-channel latent space significantly increases computational cost, the perceptual stream employs the filtered backprojection approximation residual for lightweight correction. Because the perceptual stream does not require repeated compression to a single-channel CT image between each unfolded block, it can continuously transmit high-dimensional perceptual information such as texture, edges, and artifact patterns across multiple unfolded blocks.

[0055] Within each expanded block, physical flow features and sensing flow features further perform spatial and frequency domain interactive modulation on the corrected physical flow and sensing flow. This module operates according to a preset channel ratio. The two flow characteristics are divided into spatial components and frequency components, respectively; for example... This indicates that one-quarter of the channels enter the frequency domain branch, and the remaining channels enter the spatial domain branch. Let... and Let the spatial components of the physical flow and the sensory flow be represented respectively. Then, the spatial domain cross-modulation can be expressed as follows (since the representation of this modulation is consistent across all levels, the current level is omitted unless otherwise specified). and expand block (subscript)

[0056] in, and These represent the scaling and offset parameters generated from the spatial components of the perceptual stream, respectively. Represents a convolution mapping. This represents the Sigmoid function. This represents element-wise multiplication. The first term injects texture information maintained by the perceptual flow into the physical flow, while the second term uses the geometric structure constrained by the projected data in the physical flow to gate the perceptual flow, thereby reducing the generation of structures by perceptual features that are inconsistent with the actual imaging geometry. And for the assumption... and Representing the frequency components of the two streams respectively, performing a two-dimensional fast Fourier transform on them yields the physical amplitude. Physical phase Perceived amplitude and perceived phase Then, amplitude-phase cross-modulation can be expressed as follows:

[0057]

[0058] in, Indicates feature splicing, This represents a phase modulation network used to align the phase structure of the sensed flow with the structure in the physical flow determined by the scanning geometry. The amplitude mask generation network is used to identify frequency-dependent sparse view artifact patterns based on the amplitude of the sensed flow and generate a dynamic attenuation mask acting on the amplitude of the physical flow. The modulated amplitude and phase are returned to the spatial domain via a two-dimensional inverse Fourier transform, and then merged with the output of the spatial branch to obtain the interactive modulation result of the current unfolded block. Thus, sensed information can only selectively enter the physical flow through affine parameters, gating masks, phase corrections, and amplitude attenuation masks, rather than directly replacing the image content in the physical flow constrained by measured projections. In the above interactive modulation, both the two-dimensional fast Fourier transform and inverse transform are performed independently by feature channel; at each point... This represents the set of learnable parameters for the corresponding network or decomposition operator.

[0059] The loop at the current level and the loop between levels are as follows: [The loop continues with the first...] Layer The inputs to each expanded block are denoted as follows: and First based on and The independent correction results of the two streams are obtained, and then the results are sent to subsequent processing via the interactive modulation. and Intermediate supervision is applied to the physical flow image estimation output of each expanded block; the perceptual flow latent features are not directly used to calculate the L1 loss with the reference CT image. The training loss satisfies... ,in Indicates the first Number of layer expandable blocks Indicates the relationship with the first Reference CT image corresponding to slice resolution, This indicates the hierarchy weight. If the current expanded block is not the last expanded block of this layer, then it remains unchanged. If the current expandable block remains unchanged and proceeds to the next expandable block at the same level, and the current expandable block is the last expandable block at that level, then... Then, learnable upsampling is used to send the final physical flow image estimate and perceived flow latent features of this layer to the next layer. Layer; if the current expanded block is the last expanded block of layer 0, then... As the final CT reconstructed image output.

[0060] like Figure 5 As shown, in one system embodiment, the sparse-view CT imaging system includes a CT diagnostic imaging device, an image reconstruction unit, and a medical image output unit. The scanning control unit of the CT diagnostic imaging device controls the X-ray source and detector to acquire sparse projection data according to a low-dose scanning protocol; the image reconstruction unit converts the sparse projection data into a sparse sine wave and executes the aforementioned sparse-view CT imaging method based on a multi-level sub-imaging system; the medical image output unit is used to display, transmit, or store CT reconstructed images. This image reconstruction unit can be a built-in reconstruction module of the CT diagnostic imaging device or a software module in a medical image reconstruction workstation. The CT reconstructed images can be displayed on a medical image display terminal to provide diagnostic information to doctors and transmitted to a hospital image information system for archiving, or stored in the CT diagnostic imaging device or reconstruction workstation for further downstream tasks such as lesion segmentation and automatic diagnosis.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] 1. This invention constructs a virtual CT sub-imaging system from coarse to fine through null space decomposition, enabling the reconstruction process to start from a coarse layer with low null space uncertainty and then gradually recover to the target resolution. This method can anchor the unfolding trajectory to a sub-imaging system with clear physical meaning, reducing overshoot, undershoot, and random oscillations of intermediate unfolding layers relative to the target image, thereby improving iterative stability and model interpretability.

[0063] 2. This invention sets up the physical flow and the perception flow in parallel. The physical flow is used to perform projection data consistency correction, while the perception flow is used to maintain high-dimensional perception features within the multi-channel latent space, thereby avoiding the repeated compression of latent features into single-channel CT images between each unfolded block. This approach can alleviate the information bottleneck between unfolded blocks, allowing perception information such as long-range structural information and artifact patterns to continuously propagate in the reconstruction network, thus improving the accuracy of sparse-view CT reconstruction.

[0064] 3. This invention utilizes the Radon perceptual sine wave decomposer, a coarse-to-fine hierarchical reconstruction trajectory, and interactive modulation in the spatial and frequency domains. This allows the model to no longer simply learn the overall artifact distribution mapping in the training data, but rather progressively refine the frequency-related interpolation and artifact patterns at each level. This approach reduces dependence on specific training sites or specific scan distributions, improving the model's robustness to out-of-domain anatomical structures.

[0065] 4. In the unfolded block, the present invention selectively fuses the physical flow and the perceptual flow through spatial domain cross-modulation and amplitude-phase cross-modulation, so that the geometric constraints provided by the physical flow and the texture details provided by the perceptual flow can complement each other, which is beneficial to suppressing sparse view artifacts and preserving anatomical details.

[0066] 5. This invention can be deployed in CT diagnostic imaging equipment or medical image reconstruction workstations to generate CT reconstruction images for diagnostic observation under low-dose sparse view acquisition conditions. It can improve the imaging accuracy, stability and out-of-domain robustness of sparse projection view CT reconstruction while reducing CT radiation dose.

[0067] To demonstrate the effectiveness of this invention, comparative and ablation experiments were conducted.

[0068] This embodiment compares the reconstruction of different models under different sparse projection viewpoints. The American Association of Physicists in Medicine (AAPM) clinical low-dose CT dataset was used for validation. Chest and abdominal CT scans were divided into a training set of 136 cases, a validation set of 30 cases, and a test set of 30 cases, containing 17294, 3542, and 4021 tomographic images respectively; an additional 1136 slices of head CT images not used in training were used as an out-of-domain test set. The size of the complete fan-beam sine curve is 512 projection viewpoints × 256 detector channels, and the corresponding reference CT image size is [missing information]. Resolution. The complete fan beam sine map is uniformly extracted along the projection viewing angle direction into sparse sine maps of 64, 32, and 16 viewing angles, which are used to represent different degrees of sparse viewing angle acquisition conditions.

[0069] In this embodiment, a sparse sine wave is used as input, and the CT image corresponding to the complete view projection data is used as the reference target image. The physical flow update network, the perceptual flow update network, the Radon perceptual sine wave decomposer, and the interactive modulation module are trained end-to-end. The layer weights are set to... Except for the coarsest initialization level, each sub-imaging level is set... The number of expansion blocks and the maximum number of iterations for the conjugate gradient are set to [number]. All methods employ the same random rotation and flipping data augmentation techniques, using a learning rate of... The AdamW optimizer was used to train the model for 50 epochs with a batch size of 8. The above values ​​are specific to this embodiment and do not constitute a limitation on the technical solution of this invention.

[0070] This embodiment selects the filtered back projection method FBP, the image domain post-processing network FBPConvNet, the depth unfolding networks FISTA-Net, LEARN++, DREAM-Net and PAIL, the dual-domain networks DuDoTrans, DDPTransformer and FreeSeed, and the diffusion generation network CvG-Diff as comparison methods. Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), learned perceptual patch similarity (LPIPS), and data consistency error (on the observed sine curve) are used as comparison methods. error, The higher the PSNR and SSIM, the higher the reconstruction quality; the lower the LPIPS and data consistency error, the higher the reconstruction quality.

[0071] Table 1 presents the quantitative test results. As shown in Table 1, the method of the present invention achieves the highest PSNR and SSIM and the lowest LPIPS under 16-view, 32-view, and 64-view conditions. This result indicates that the multi-level sub-imaging system can first recover the relatively stable low-frequency structure under sparse angle sampling, and then supplement higher-frequency anatomical details layer by layer, enabling the method of the present invention to maintain high reconstruction quality under different sparsity levels.

[0072] Table 1. Reconstruction results of the present invention and the comparative method on AAPM chest and abdomen data under different sparse projection views.

[0073] FBP 18.00 / 0.410 / 41.5 / 60.6 21.47 / 0.505 / 33.4 / 33.4 26.26 / 0.640 / 23.0 / 13.5 FBPConvNet 29.84 / 0.837 / 7.27 / 6.38 32.53 / 0.875 / 5.10 / 4.46 35.78 / 0.915 / 3.07 / 1.67 FISTA-Net 28.24 / 0.801 / 9.32 / 7.04 31.62 / 0.863 / 5.97 / 5.14 34.99 / 0.913 / 3.25 / 2.49 DuDoTrans 28.22 / 0.808 / 9.15 / 7.81 32.02 / 0.867 / 5.62 / 3.67 35.76 / 0.915 / 3.01 / 2.01 DDPTransformer 30.17 / 0.845 / 8.04 / 5.66 33.27 / 0.885 / 5.40 / 3.13 36.62 / 0.922 / 3.14 / 1.58 FreeSeed 30.02 / 0.831 / 7.17 / 6.82 32.35 / 0.872 / 5.33 / 6.58 35.06 / 0.909 / 3.64 / 1.93 LEARN++ 30.02 / 0.833 / 7.53 / 2.44 35.25 / 0.909 / 3.23 / 0.98 38.43 / 0.939 / 1.83 / 0.70 DREAM-Net 31.56 / 0.862 / 6.32 / 2.88 34.73 / 0.904 / 3.87 / 2.14 36.85 / 0.927 / 2.55 / 1.48 CvG-Diff 30.97 / 0.856 / 7.32 / 4.25 33.23 / 0.886 / 5.45 / 2.47 34.90 / 0.906 / 4.28 / 2.01 PAIL 32.89 / 0.880 / 5.21 / 2.13 36.08 / 0.918 / 2.99 / 2.16 38.98 / 0.943 / 1.62 / 1.39 This invention 34.42 / 0.898 / 4.25 / 1.69 37.47 / 0.928 / 2.58 / 0.89 40.17 / 0.949 / 1.42 / 0.51

[0074] To verify the robustness of the model to anatomical structures outside the training distribution, this embodiment uses 1136 layers of untrained head CT images as out-of-domain test data. Each method still uses the network parameters trained on chest and abdominal data, without retraining or fine-tuning using the aforementioned head CT images, and inputs sparse sine waves of 16, 32, and 64 views obtained by uniformly extracting from complete projections, respectively. The evaluation metrics and the numerical order in the table are the same as in Table 1. The out-of-domain reconstruction results of this invention and the comparative methods are shown in Table 2 (quantitative test metrics). Figure 6 The visualization comparison is shown below.

[0075] Table 2. Reconstruction results of the present invention and the comparative method on AAPM head out-of-domain data under different sparse projection perspectives.

[0076] FBP 22.05 / 0.521 / 39.4 / 48.2 25.58 / 0.602 / 31.3 / 29.7 31.38 / 0.723 / 19.9 / 12.0 FBPConvNet 29.95 / 0.863 / 7.88 / 8.64 34.46 / 0.919 / 3.75 / 8.01 40.26 / 0.960 / 1.26 / 2.31 FISTA-Net 30.06 / 0.872 / 7.30 / 8.92 33.74 / 0.914 / 4.19 / 7.98 39.03 / 0.961 / 1.73 / 4.65 DuDoTrans 29.17 / 0.850 / 8.16 / 10.3 34.13 / 0.921 / 4.03 / 4.07 41.18 / 0.971 / 1.23 / 1.65 DDPTransformer 30.51 / 0.892 / 6.54 / 6.40 35.88 / 0.947 / 2.73 / 2.77 42.72 / 0.982 / 0.80 / 1.17 FreeSeed 31.38 / 0.883 / 6.53 / 6.08 34.22 / 0.895 / 4.65 / 7.88 40.89 / 0.972 / 1.14 / 1.98 LEARN++ 31.66 / 0.892 / 5.91 / 2.76 40.88 / 0.974 / 0.96 / 1.10 46.75 / 0.991 / 0.29 / 0.61 DREAM-Net 33.19 / 0.910 / 4.11 / 3.24 38.08 / 0.954 / 1.77 / 2.61 42.41 / 0.978 / 0.87 / 1.35 CvG-Diff 31.44 / 0.890 / 7.18 / 7.16 34.37 / 0.924 / 4.93 / 5.12 37.53 / 0.949 / 3.26 / 2.02 PAIL 36.53 / 0.948 / 2.40 / 1.54 41.46 / 0.977 / 0.89 / 1.94 47.64 / 0.992 / 0.23 / 1.17 This invention 37.85 / 0.959 / 1.90 / 1.01 44.38 / 0.987 / 0.50 / 0.51 50.65 / 0.996 / 0.10 / 0.22

[0077] From Table 2 and Figure 6 It is evident that the advantages of the method of this invention over data within the same domain are further amplified on data outside the head domain. This result demonstrates that the method of this invention does not solely rely on learning the overall artifact distribution of a specific scanned area, but rather progressively refines frequency-related interpolation and artifact patterns along the trajectory of multi-level null space decomposition, thus reducing artifact removal failure and anatomical boundary shift caused by domain offset.

[0078] On the NVIDIA RTX A6000 graphics processor, a 32-viewpoint depth is supported. The tomographic images were tested in batches of 16. The average result for 50 batches was as follows: the method of this invention has 4.2M parameters, 108.2G floating-point operations, and an average inference time of 29.4 ms per tomographic image, where G represents... This involves floating-point operations. Under the same conditions, PAIL has 7.5M parameters and an average inference time of 275.5 ms per tomographic image; LEARN++ has 6.1M parameters and an average inference time of 112.6 ms per tomographic image. These results demonstrate that the method of this invention maintains high reconstruction accuracy while possessing a model size and image processing speed suitable for CT diagnostic imaging workflows.

[0079] To verify the effectiveness of this invention in downstream diagnosis, this embodiment uses the MosMedData dataset to perform quality assessment and lesion segmentation verification on the reconstructed images. The MosMedData dataset includes 1110 chest CT scans with imaging manifestations of the novel coronavirus, divided into a training set of 888 cases, a validation set of 111 cases, and a test set of 111 cases, comprising 37158, 4658, and 4595 tomographic images, respectively. Twenty cases were randomly selected from the test set for downstream lesion segmentation evaluation. Each reconstruction method uses the same sparse projection data as input and outputs CT reconstructed images. The CT reconstructed images from each method are then input into the same pre-trained medical image segmentation model, using the segmentation results corresponding to the complete view reference CT image as a reference to evaluate the consistency of lesion segmentation between ground-glass opacities and solid regions. Reconstruction metrics include PSNR, SSIM, and LPIPS; segmentation metrics include weighted Dice coefficient, 95% Hausdorff distance (HD95), and relative volume difference (RVD). A larger Dice indicates a higher degree of overlap between the lesion segmentation region and the reference segmentation region in the reconstructed image; a smaller HD95 indicates a smaller boundary deviation between the two; and the closer the absolute value of RVD is to 0, the closer the lesion volume estimation is to the reference result. The evaluation results are shown in Table 3, where the reconstruction indicators are expressed as mean ± standard deviation, and the unit of LPIPS is [missing value]. .

[0080] Table 3. Reconstruction and lesion segmentation results of the present invention and the comparative method on the MosMedData dataset.

[0081] FBP 21.05±1.51 / 0.478±0.033 / 36.47±2.71 0.066 ∞ 3055% FBPConvNet 32.65±1.70 / 0.906±0.022 / 4.04±1.42 0.701 11.0 -16.6% FISTA-Net 31.73±1.40 / 0.895±0.018 / 4.66±1.53 0.699 10.78 -27.9% DuDoTrans 32.40±1.90 / 0.910±0.023 / 4.41±1.57 0.665 16.0 -21.7% DDPTransformer 32.55±2.23 / 0.903±0.061 / 5.10±3.97 0.712 10.1 -17.2% FreeSeed 34.66±1.56 / 0.926±0.019 / 3.20±1.43 0.794 5.57 -8.9% LEARN++ 35.03±1.56 / 0.933±0.020 / 2.71±1.36 0.788 8.33 -10.0% DREAM-Net 34.93±1.60 / 0.932±0.019 / 3.00±1.57 0.797 6.35 -9.4% CvG-Diff 34.05±1.59 / 0.924±0.020 / 3.77±1.60 0.745 9.64 -9.5% PAIL 36.06±1.38 / 0.942±0.016 / 2.32±1.28 0.827 4.09 -10.3% This invention 37.68±1.42 / 0.951±0.016 / 1.88±1.24 0.855 2.67 -4.5%

[0082] As shown in Table 3, the PSNR, SSIM, LPIPS, Dice, and HD95 of the method of this invention are all superior to all the comparative methods, and the absolute value of RVD is also closest to 0. This indicates that the improvement in reconstruction quality can be transferred to downstream lesion segmentation tasks, thereby improving the accuracy in downstream clinical diagnostic tasks.

[0083] like Figure 7 As shown, representative cases of different progression stages of the novel coronavirus from the MosMedData dataset are presented sequentially, including complete-view reference CT images, reconstructed images from various comparison methods and the method of this invention, and corresponding lesion segmentation results obtained using the same pre-trained medical image segmentation model. Ground-glass opacities are marked in green, and consolidation areas are marked in orange. All CT images are displayed using a lung window with a window width of 1400 HU and a window level of -300 HU. For cases with lung parenchymal involvement greater than 75% that were not included in the training set, [further details are needed]. Figure 7 Cases marked as outside the region. (Through...) Figure 7 The method allows for a direct comparison of whether different methods result in missed lesions or false positive areas. The lesion segmentation contour obtained by the method of this invention is closer to the segmentation contour corresponding to the complete view reference CT image, which is consistent with the results of Dice, HD95 and RVD in Table 3.

[0084] Table 4 shows the iterative stability comparison of this invention with other depth unfolding models. Under the 32 sparse view sampling condition of the AAPM dataset, this embodiment records the SSIM, LPIPS, and average CT value of the reconstructed image at each unfolding stage in three independent training sessions for FISTA-Net, LEARN++, DREAM-Net, PAIL, and the method of this invention, and summarizes the main stability information in the original iterative curves in Table 4. In the table, "maximum SSIM drop between adjacent iterations" represents the maximum decrease in SSIM between any two adjacent unfolding stages; "maximum LPIPS rebound between adjacent iterations" represents the maximum increase in LPIPS; "maximum deviation of average CT value" represents the maximum absolute deviation of the average CT value of the intermediate reconstructed image relative to the average CT value of the reference image, approximately -553.5 HU; "maximum half-width of SSIM envelope between training sessions" is estimated from the upper and lower envelopes of the three independent training results and is used to characterize the trajectory dispersion caused by training randomness. To ensure the fairness of the comparison, the method of this invention used in the test only performed supervised learning on the final output of the model and the full-view CT image during training to specifically eliminate the influence of intermediate supervision on the experiment.

[0085] Table 4. Iterative stability statistics of the present invention and various depth unfolding methods under AAPM 32 sparse view sampling conditions

[0086] FISTA-Net 0.08 0.05 89 0.18 LEARN++ 0.33 0.27 220 0.25 DREAM-Net 0.12 0.03 143 0.19 PAIL 0.22 0.07 65 0.12 This invention 0.008 0.002 17 0.02

[0087] As shown in Table 4, the SSIM fallback and LPIPS bounce of the method of the present invention are both less than 0.01 within each resolution level, and the maximum half width of the SSIM envelope formed by three independent trainings is about 0.02, indicating that the intermediate reconstruction results are smoothly updated along the preset coarse to fine trajectory, and have low sensitivity to training initialization, batch data changes and random noise.

[0088] To verify the effectiveness of the multi-level sub-imaging system and the dual-flow structure, this embodiment sets up four comparative examples under 32-view conditions of AAPM data: Comparative example 1 removes the multi-level null space decomposition while maintaining the same number of unfolding iterations as the method of this invention at the target resolution; Comparative example 2 similarly removes the decomposition and reduces the number of unfolding iterations to make the inference time close to that of the method of this invention; Comparative example 3 retains only the physical flow and increases its number of feature channels; Comparative example 4 retains only the sensory flow and increases its number of feature channels. The quantitative results of each comparative example are shown in Table 5, where the parameter M in the table represents... There are 1 trainable parameters.

[0089] Table 5. Ablation experimental results of multi-level sub-imaging system and dual-flow structure

[0090] Comparative Example 1: Removing null space decomposition, same number of iterations 37.29 / 0.9261 43.53 / 0.9846 2.0M 75.5 ms Comparative Example 2: Removing null space decomposition brings inference time closer to the target time. 36.37 / 0.9198 41.06 / 0.9758 0.8M 30.7 ms Comparative Example 3: Only physical flow is retained 36.06 / 0.9174 40.91 / 0.9760 4.6M 31.1 ms Comparative Example 4: Only the perceptual stream is retained 36.97 / 0.9243 42.84 / 0.9809 4.7M 30.2 ms Method of the present invention 37.47 / 0.9280 44.38 / 0.9867 4.2M 29.4 ms

[0091] As shown in Table 5, while maintaining the same number of unfolding iterations, the inference time of a single tomographic image after removing multi-level decomposition increased from 29.4 ms to 75.5 ms, and both the in-domain and head-domain PSNR decreased. When the inference times were similar, the in-domain PSNR after removing multi-level decomposition decreased by 1.10 dB. This result indicates that the coarse-to-fine multi-level sub-imaging system does not simply increase the number of unfolding iterations, but rather reduces the null space uncertainty of the coarse-level sub-imaging system, providing a more stable initial path with lower computational cost for subsequent iterations. Comparative Examples 3 and 4, with higher parameter counts than the method of this invention, still yielded lower intra-domain and extra-domain reconstruction results than the method of this invention. This result demonstrates that the physical flow and the perceptual flow have different and complementary functions. The physical flow maintains the consistency of the projection data, while the perceptual flow bypasses the bottleneck of the single-channel image domain and maintains high-dimensional perceptual information. The parallel setup and selective interaction of the two cannot be replaced by simply increasing the parameter count of either single flow.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A sparse-view CT imaging method based on a multi-level sub-imaging system, characterized in that, Includes the following steps: Step S1: Control the CT diagnostic imaging equipment to perform sparse view projection acquisition on the subject according to the low-dose scanning protocol, receive the sparse projection data acquired by the CT diagnostic imaging equipment, and form a sparse sine curve based on the sparse projection data. ; Step S2: Record the number of projection viewpoints corresponding to the complete scan protocol as The actual number of projection angles acquired by the low-dose scanning protocol is ,according to Determine the sparsity factor and according to Determine the number of levels in the null space decomposition ; Step S3: Along the sparse sine graph The detector axis is subjected to stepwise dimensionality reduction processing to obtain physical projection data at multiple levels. Configure sub-imaging operators corresponding to the physical projection data of each level. This forms a virtual CT sub-imaging system that progresses from coarse to fine; where the hierarchical index... Indicates the target resolution level. This indicates the level with the coarsest detector grid and the lowest image resolution; Step S4: Process the sparse sine graph using the Radon-perceptual sine graph decomposer. Latent space decomposition is performed to obtain perceptual projection features corresponding to multiple levels. ; Step S5: Apply the physical projection data of the coarsest level respectively. and perceptual projection features Perform filtered backprojection to obtain the initial physical flow image and initial perceptual flow latent features; upsample the initial physical flow image and initial perceptual flow features to the first reconstruction level, and then, according to the coarse-to-fine level order, perform further refactoring at the current level. Layer Within each expansion block, corrections are first performed on the physical flow and the sensory flow separately, where the physical flow is addressed using sub-imaging operators. Perform a forward projection on the current physical flow image estimate, and then compare the obtained forward projection with the physical projection data of that layer. The differences between them serve as the basis for projection data consistency correction; the perceptual stream projects the current multi-channel latent features channel by channel using the same sub-imaging operator, and the resulting latent space projection is compared with the perceptual projection features of that layer. The differences between them serve as the basis for latent space correction; Step S6: After the physical flow and sensory flow of the current unfolded block have completed the corrections, the features of the physical flow and the features of the sensory flow are subjected to spatial and frequency domain interactive modulation to fuse the projection physical constraints and sensory feature constraints, and to suppress CT reconstruction artifacts caused by sparse view acquisition. If there are still unfolded blocks that have not been executed at the current level, the next unfolded block is entered, and the processing of correcting the two flows in step S5 and the interactive modulation in step S6 are repeated. If all unfolded blocks at the current level have been executed and the current level is not the finest level, the interactively modulated physical flow image estimation and sensory flow latent features are upsampled to the next finer level, and the processing of correcting the two flows in step S5 and the interactive modulation in step S6 are repeated. Step S7: After the finest level of processing is completed, a CT reconstructed image is generated and output as a medical image for diagnostic observation, display, transmission or storage.

2. The method according to claim 1, characterized in that, In step S3, the physical projection data of layer 0 is: , No. Layer physical projection data is derived from the formula Confirmed, among which This represents a one-dimensional average pooling operator performed along the detector axis with a factor of 2. .

3. The method according to claim 1, characterized in that, In step S3, the first Layer CT sub-imaging system meets ,in Indicates the first The corresponding CT image of the layer, This represents a sparse CT projection operator with a coarser detector grid, and the... The zero spatial dimension of the layer imaging system satisfies ,in Indicates the first Layer detector resolution, and It is a constant.

4. The method according to claim 1, characterized in that, In step S5, at the first Layer Within each expanded block, the physical flow is estimated using the current physical flow image. as initial value For the normal equation implement Step conjugate gradient iteration yields the truncated Krylov subspace trajectory. ;according to Construct the filtered back projection approximate residual; and according to Update the physical flow image estimate, where This represents the back projection operator that matches the discrete projection operator. This represents the filtering back projection operator. Indicates the first Layer The physical flow of each expanded block updates the network. , and The subscripts represent the conjugate gradient iteration, the Krylov subspace trajectory residual generated during the iteration, and the filtered back projection residual, respectively.

5. The method according to claim 1, characterized in that, In step S4, the Radon sensing sine wave decomposer follows... as well as Generate multi-level perceptual projection features, where Indicates the first Layer-learnable Radon-aware decomposition operator, .

6. The method according to claim 1, characterized in that, In step S5, at the first Layer Within each expanded block, the sensing flow maintains multi-channel latent features. And based on the current level, perceive the projection features. With multi-channel latent features Correction is performed on the latent space projection residuals between the projection results, and the latent features of the sensed flow are corrected according to... Update, in which This indicates that the perception flow updates the network.

7. The method according to claim 1, characterized in that, In step S6, the spatial domain and frequency domain interactive modulation includes spatial domain cross modulation and amplitude-phase cross modulation. Specifically, spatial domain cross modulation uses the spatial components of the sensing stream to modulate the spatial components of the physical stream and uses the spatial components of the physical stream to gate the spatial components of the sensing stream. Amplitude-phase cross modulation uses the frequency-phase information of the physical stream to constrain the structural alignment of the sensing stream and uses the frequency-amplitude information of the sensing stream to generate dynamic attenuation information for suppressing sparse view artifacts.

8. The method according to claim 1, characterized in that, When training the network to perform physical flow updates, perceptual flow updates, and interactive modulation, supervised constraints are imposed on the physical flow image estimation at each intermediate layer, and the training loss satisfies... ,in Indicates the first Number of layer expandable blocks Indicates the relationship with the first Reference CT image corresponding to slice resolution, Indicates the hierarchical weight.

9. A sparse-view CT imaging system based on a multi-level sub-imaging system, characterized in that, The invention includes a CT diagnostic imaging device and an image reconstruction unit. The CT diagnostic imaging device includes an X-ray source, a detector, and a scanning control unit. The scanning control unit controls the X-ray source and detector to perform sparse view projection acquisition on the subject according to a low-dose scanning protocol to obtain sparse projection data. The image reconstruction unit is located within the CT diagnostic imaging device or within a medical image reconstruction workstation that is communicatively connected to the CT diagnostic imaging device. The image reconstruction unit includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it causes the image reconstruction unit to perform the method described in any one of claims 1 to 8 based on the sparse projection data to generate a CT reconstructed image for diagnostic observation, display, transmission, or storage.

Images