A low-dose CT image reconstruction method and system based on flow matching

Abstract

The application provides a low-dose CT image reconstruction method and system based on flow matching, and the method comprises the following steps: estimating a signal-related noise parameter of a low-dose CT image through a physical noise encoder, and constructing a noise prior distribution conforming to the physical law of CT imaging; embedding an anatomical transport module in a U-Net coding path, estimating an anatomical structure displacement field by using adjacent slice features, and realizing cross-slice feature alignment; introducing an interlayer staggered attention mechanism at a bottleneck layer, and fusing global anatomical information of adjacent slices; and generating a reconstructed image from the noise prior by ordinary differential equation integration based on a flow matching framework. Through the introduction of a physical noise model and a three-dimensional anatomical continuity constraint, high fidelity and high efficiency are realized in low-dose CT reconstruction, and the method can be widely applied to clinical CT imaging systems.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method and system for low-dose CT image reconstruction based on flow matching. Background Technology

[0002] Computed tomography (CT) is an indispensable tool in modern medical imaging diagnosis, widely used in clinical settings such as tumor screening, cardiovascular disease diagnosis, and emergency trauma assessment. However, the X-ray radiation generated during CT scans poses a potential risk to patient health, especially for patients requiring multiple follow-up examinations and sensitive populations such as children, where the cumulative effect of radiation dose may increase the risk of cancer. Therefore, minimizing radiation dose while ensuring image diagnostic quality has long been a core issue of concern in the field of CT imaging.

[0003] Low-dose CT scans reduce radiation dose by decreasing tube current or tube voltage, but this significantly lowers the signal-to-noise ratio of the projected data, leading to severe quantum noise and streak artifacts in the reconstructed images. The noise exhibits signal correlation, meaning areas with higher signal intensity in the image also have higher noise levels, while artifacts manifest as striped interference along the probe channel. This noise and artifacts can mask small lesions and blur tissue boundaries, severely impacting the accuracy of diagnosis.

[0004] Extensive research has been conducted in academia and industry to address the image quality issues of low-dose CT. Traditional iterative reconstruction methods, such as statistical iterative reconstruction and model-based iterative reconstruction, establish accurate CT imaging physical models and noise statistical models, and then perform iterative optimization in the projection domain. These methods can suppress noise and artifacts to some extent, but they suffer from high computational complexity and long reconstruction times, making them difficult to meet the needs of real-time clinical diagnosis. Furthermore, iterative reconstruction methods require high accuracy in system modeling; any model error can lead to new artifacts in the reconstructed image.

[0005] In recent years, deep learning-based post-processing methods have become a research hotspot in the field of low-dose CT reconstruction. These methods typically employ convolutional neural networks, such as U-Net and its variants, to learn the mapping relationship from low-dose images to full-dose images in an end-to-end manner. These methods offer fast inference speed, simple implementation, and significant improvements in quantitative metrics. However, existing deep learning methods have the following shortcomings: First, most methods treat CT reconstruction as a general image-to-image translation problem, failing to fully utilize the prior physical knowledge of CT imaging, especially the dependence of noise on signal; second, existing methods typically process individual slices independently, ignoring the anatomical continuity between adjacent slices in the CT image sequence, leading to inconsistencies in the reconstruction results in three-dimensional space; third, existing methods are sensitive to the distribution of training data, and reconstruction performance significantly degrades when the dose level or scanning protocol of the test data differs from that of the training data.

[0006] Generative modeling methods, including diffusion models and flow matching models, have made groundbreaking progress in image generation in recent years and have also been introduced into low-dose CT reconstruction tasks. These methods can generate high-quality, diverse images by learning the transformation process from a simple prior distribution to the target data distribution. However, existing generative methods still face challenges in low-dose CT reconstruction: First, they do not incorporate the CT imaging physical model into the generation process, and the prior distribution usually adopts a standard Gaussian distribution, which differs significantly from the noise distribution of low-dose CT images; second, the generation process has limited ability to preserve three-dimensional anatomical structures, making it difficult to ensure anatomical consistency between adjacent slices; third, the sampling process of diffusion models requires hundreds to thousands of iterations, resulting in slow inference speeds that are difficult to meet the needs of clinical applications.

[0007] In summary, existing low-dose CT reconstruction techniques still have shortcomings in terms of physical model utilization, anatomical continuity maintenance, and inference efficiency. There is an urgent need for a low-dose CT reconstruction method that can integrate physical priors, maintain three-dimensional anatomical consistency, and achieve fast reconstruction speed. Summary of the Invention

[0008] Purpose of the Invention: The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a low-dose CT (computed tomography) image reconstruction method and system based on flow matching. By modeling the image reconstruction process as a geodesic flow matching process from the prior distribution of physical noise to the target distribution, the invention utilizes a physical noise encoder to adaptively estimate the signal-related noise parameters to construct a prior distribution that conforms to the laws of CT imaging, achieves multi-scale anatomical feature alignment between adjacent slices through an anatomical transport module, fuses global anatomical structure information through an inter-slice staggered attention mechanism, and finally generates high-quality reconstructed images rapidly through integration of ordinary differential equations. This solves the technical problems of insufficient utilization of physical models, lack of anatomical continuity, and slow reconstruction speed in the prior art.

[0009] The method includes the following steps:

[0010] Step 1: Estimate signal-related noise parameters of low-dose CT images using a physical noise encoder: Input a low-dose CT image, extract image features through a convolutional neural network, output noise standard deviation and inverse gain parameters, establish a signal-related noise variance model, and construct a physical noise prior distribution based on the estimated noise parameters;

[0011] Step 2: Use the anatomical transport module to align anatomical features between adjacent slices: Input the current slice features, the previous slice features, and the next slice features into the anatomical transport module, estimate the two-dimensional anatomical structure displacement field between adjacent slices, use the estimated displacement field to perform bidirectional spatial deformation on the previous slice features and the next slice features respectively, and use an adaptive gating mechanism to perform weighted fusion of the current slice features and the deformed adjacent slice features to output the aligned anatomical features.

[0012] Step 3: Use interlayer staggered attention mechanism to fuse global anatomical information of adjacent slices: Input the current slice feature and adjacent slice feature pairs into the bottleneck layer, calculate the attention weight through multi-head attention mechanism, introduce an anatomical structure perception mask based on the gray-level difference between adjacent slices to modulate the attention weight, and output the fused bottleneck features.

[0013] Step 4: Generate reconstructed images from low-dose images using a physical guided flow matching training framework: Construct a linear interpolation path from the prior distribution of physical noise to the target distribution based on the estimated noise parameters, predict the velocity field through a neural network, solve ordinary differential equations to generate reconstructed images, and train loss functions including flow matching loss, physical consistency loss, trajectory anchor point loss and divergence regularization loss.

[0014] Step 1 includes: The physical noise encoder includes a convolutional layer, a pooling layer and a fully connected layer. A single-channel low-dose CT image is input, and spatial features are extracted through two or more layers of convolution and pooling operations. The features are compressed into a feature vector through global average pooling, and after mapping through a fully connected layer, two scalar values ​​are output, which are denoted as the first original feature and the second original feature, respectively.

[0015] The first original feature is mapped to a space between a preset lower limit and a preset upper limit of noise standard deviation using a Sigmoid activation function to obtain the noise standard deviation.

[0016] The second original feature is mapped using the Sigmoid function to a value between a preset lower limit and a preset upper limit of inverse gain, resulting in the inverse gain parameter, which is specifically expressed as follows:

[0017] ,

[0018] ,

[0019] in, Indicates the standard deviation of noise. This indicates the lower limit of the preset noise standard deviation. This indicates the upper limit of the preset noise standard deviation. Indicates the first primitive feature, Represents the inverse gain parameter. This indicates the preset lower limit of the inverse gain. This indicates the preset upper limit of inverse gain. Indicates the second primitive feature;

[0020] Based on the estimated noise standard deviation and inverse gain parameter, a signal-related noise variance model is established, specifically expressed as follows:

[0021] ,

[0022] in, Indicates the noise variance. The normalized positive image value is obtained from a low-dose CT image through a linear transformation. , Represents low-dose CT images normalized to the [-1,1] interval;

[0023] Based on the signal-correlated noise variance model, the prior distribution of physical noise is constructed, specifically represented as follows:

[0024] ,

[0025] in, This represents the sampled value of the prior distribution of physical noise. This represents random noise that follows a standard normal distribution. The signal-to-noise ratio correlation coefficient is calculated using the local signal-to-noise ratio. , This represents the local signal-to-noise ratio.

[0026] Step 2 includes:

[0027] Let the current slice features be denoted as the first feature tensor. The first piece of features is denoted as the second feature tensor. All subsequent features are denoted as the third feature tensor. The three feature tensors have the same spatial size and number of channels;

[0028] The second, first, and third feature tensors are concatenated along the channel dimension to obtain a concatenated feature tensor. This concatenated feature tensor is then input into the displacement field estimation network, which outputs a two-dimensional anatomical structure displacement field, specifically represented as follows:

[0029] ,

[0030] in, This represents the displacement field of a two-dimensional anatomical structure, including horizontal and vertical displacement components. This represents the displacement field estimation network. This indicates a splicing operation along the channel dimension;

[0031] The estimated two-dimensional anatomical displacement field is used to perform bidirectional spatial deformation on the features of adjacent slices, and feature resampling is achieved through a spatial transformation network, as specifically expressed as follows:

[0032] ,

[0033] ,

[0034] in, This indicates the features of the previous segment after deformation. This indicates the post-deformation slice characteristics. This represents a spatial transformation function that uses bilinear interpolation to resample the feature map. This indicates an inversion operation on the displacement field, used to achieve reverse deformation;

[0035] By fusing the features of the current slice with the features of the deformed adjacent slices through an adaptive gating mechanism, the learnable gating weights are first calculated, specifically as follows:

[0036] ,

[0037] in, This represents the gate weight tensor, which has the same spatial dimensions as the feature tensor. This represents a gated weight generation network consisting of convolutional layers;

[0038] The average feature is obtained by averaging the features of the previous slice after deformation, the current slice feature, and the subsequent slice feature. Then, the current slice feature and the average feature are weighted and fused using gating weights. Specifically, this is represented as follows:

[0039] ,

[0040] in, This indicates the output characteristics of the anatomical transport module. This indicates the average characteristic.

[0041] In step 2, the displacement field estimation network adopts a grouped convolutional structure. The number of channels of the input spliced ​​feature tensor is three times the number of current feature channels, and the number of channels of the output two-dimensional anatomical structure displacement field is two. The intermediate layer of the network adopts a group normalization layer and SiLU activation function, and the output layer adopts a zero-initialization convolutional module, so that the output of the network is zero displacement field in the initial state.

[0042] Step 3 includes:

[0043] Let the features of the current slice be denoted as the fourth feature tensor, and the features of the previous slice be denoted as the fifth feature tensor. The subsequent features are denoted as the sixth feature tensor. The three feature tensors have the same spatial size and number of channels. The fifth and sixth feature tensors are concatenated along the channel dimension to obtain the context feature pair.

[0044] The average context feature is obtained by averaging the two feature tensors in the context feature pair, specifically represented as:

[0045] ,

[0046] in, Represents average contextual features;

[0047] The query features, key features, and value features are calculated separately using convolutional projection, as follows:

[0048] ,

[0049] ,

[0050] ,

[0051] in, Indicates query characteristics, Indicates key features, Indicates value characteristics, These represent the query projection matrix, key projection matrix, and value projection matrix, respectively, all implemented using 1×1 convolutional layers. Represents the fourth characteristic tensor;

[0052] The query features and key features are reshaped into two-dimensional spatial sequence forms, and the attention score matrix is ​​calculated, as follows:

[0053] ,

[0054] in, Represents the attention score matrix. This represents matrix multiplication of query features and key features. This represents the feature dimension of each attention head, and Softmax represents the Softmax normalization operation along the last dimension.

[0055] An anatomical structure-aware mask is introduced, and the mask matrix is ​​calculated based on the gray-level differences between adjacent slices, specifically represented as follows:

[0056] ,

[0057] in, The anatomical structure perception mask matrix represents the first... Line number The element values ​​of the column, Indicates the spatial position of the first and last segments. The difference in gray values ​​at each location Indicates the spatial position of the first and last segments. The difference in gray values ​​at each location This represents a learnable temperature parameter; exp is an exponential function.

[0058] The anatomical structure perception mask is logarithmically superimposed onto the attention score matrix to obtain the modulated attention score matrix, specifically represented as follows:

[0059] ,

[0060] in, This represents the modulated attention score matrix. Represents the natural logarithm operation;

[0061] The modulated attention score matrix is ​​multiplied by the value features, projected onto the output, and then added to the original input features to obtain the fused bottleneck features, specifically represented as follows:

[0062] ,

[0063] in, This represents the output features of the interlayer staggered attention module. This represents the output projection matrix, implemented by a 1×1 convolutional layer.

[0064] Step 4 includes the following steps:

[0065] Step 6.1: Constructing a linear interpolation path: Let the initial point sampled from the prior distribution of physical noise be denoted as the first sampling point, and the target full-dose image be denoted as the second sampling point. Define a linear interpolation path within the time interval [0,1], and add a small amount of Gaussian noise during the interpolation process to ensure numerical stability, specifically as follows:

[0066] ,

[0067] in, Indicates time Interpolation point at that location, This represents a time variable, where t takes values ​​in the range [0,1]. Indicates the first sampling point. Indicates the second sampling point. This indicates the preset minimum noise figure;

[0068] Step 6.2, Define the velocity field: Let the velocity field predicted by the neural network be denoted as the predicted velocity field, which satisfies the ordinary differential equation ,in This represents the velocity field predicted by the neural network. Represents the trainable parameters of a neural network. This represents the input low-dose CT image; d represents the differential symbol.

[0069] Step 6.3, Calculate the flow matching loss: Calculate the mean square error between the predicted velocity field and the true velocity field. The true velocity field is defined as the difference between the endpoint and the starting point, specifically expressed as:

[0070] ,

[0071] in, Indicates the stream matching loss. This means averaging over all possible paths. Represents the mathematical expectation. Denotes the square of the L2 norm. Represents the true velocity field;

[0072] Step 6.4, Calculate the physical consistency loss: The reconstructed image predicted by the constraint neural network conforms to the signal-correlated noise model. First, the target image is reconstructed based on the predicted velocity field, specifically as follows:

[0073] ,

[0074] in, This represents the target image predicted by the neural network.

[0075] The noise-to-signal ratio between the predicted target image and the input low-dose image is calculated and specifically expressed as follows:

[0076] ,

[0077] Wherein, NSR represents the noise-to-signal ratio. Indicates local spatial average;

[0078] The physical consistency loss is constructed based on the noise-to-signal ratio, and is specifically expressed as follows:

[0079] ,

[0080] in, Indicates the loss of physical consistency. This represents a time-weighted function that varies over time, used to enhance the constraint strength of the intermediate path;

[0081] Step 6.5, Calculate trajectory anchor point loss: Randomly select two adjacent time points, constraining the neural network to maintain consistency in the target images predicted at the two time points. First, randomly generate the offset, specifically represented as follows:

[0082] ,

[0083] in, Indicates a random offset. This indicates the preset maximum offset. Indicates from 0 to Uniform sampling between Indicates a random positive or negative sign;

[0084] The anchor point time point is calculated as follows:

[0085] ,

[0086] in, This indicates the anchor point in time; `clip` represents the clipping operation. This indicates the preset minimum time value. This indicates the preset maximum time value;

[0087] Calculate the predicted target images at the original time point and the anchor time point respectively, and then calculate the mean square error, which is specifically expressed as:

[0088] ,

[0089] in, Indicates trajectory anchor point loss. This represents a comprehensive optimization of all possible combinations of path start points, path end points, master time points, and anchor time points. Indicates time The predicted target image at that location, Indicates time The predicted target image at the location;

[0090] Step 6.6: Calculate the divergence regularization loss: Calculate the divergence of the two-dimensional anatomical structure displacement field output by each anatomical transport module in Step 2, and take the absolute value as the regularization term, specifically expressed as:

[0091] ,

[0092] in, This represents the divergence regularization loss. This represents the horizontal component of the displacement field of a two-dimensional anatomical structure. This represents the vertical component of the displacement field of a two-dimensional anatomical structure. This represents the partial derivative of the horizontal component in the horizontal direction. This represents the partial derivative of the vertical component in the vertical direction;

[0093] Step 6.7: Construct the total loss function: The total loss function is calculated by weighted summation of all losses, specifically expressed as:

[0094] ,

[0095] in, Represents the total loss function. This represents the weighting coefficient for the physical consistency loss. This represents the trajectory anchor point loss weighting coefficient. This represents the divergence regularization loss weight coefficient.

[0096] Step 4 includes:

[0097] Initialization: Set the total number of steps to N, and the time step size to N. initial point ;

[0098] Iterative solution: For From 0 to Perform the following operations:

[0099] ,

[0100] ,

[0101] ,

[0102] in, Indicates the first The timing of the step This represents the predicted velocity field at the current time point. Indicates the state of the current iteration step;

[0103] Output: After the iteration is complete, The image is reconstructed by cropping it to a preset value range.

[0104] The present invention also provides a CT image reconstruction system based on the method described above, comprising:

[0105] The physical noise encoder module is used to take low-dose CT images as input, estimate signal-related noise parameters, including noise standard deviation and inverse gain parameters, through a convolutional neural network, and construct a prior distribution of physical noise based on the estimated parameters.

[0106] The anatomical transport module array consists of multiple anatomical transport modules. Each anatomical transport module is embedded into different levels of the U-Net encoding path. Each anatomical transport module receives the current slice features, the previous slice features, and the next slice features at the current level, estimates the displacement field of the two-dimensional anatomical structure, and performs bidirectional spatial deformation. It fuses features through an adaptive gating mechanism to achieve multi-scale anatomical feature alignment.

[0107] The interlayer staggered attention module is set in the bottleneck layer of U-Net. It receives the current slice feature and the feature pairs of adjacent slices, calculates the attention weight through a multi-head attention mechanism, introduces an anatomical structure perception mask based on the gray-level difference of adjacent slices to modulate the attention weight, and outputs the fused bottleneck features.

[0108] The flow matching training module is used to construct a linear interpolation path from the prior distribution of physical noise to the target distribution. It predicts the velocity field through a neural network, calculates the total loss function including flow matching loss, physical consistency loss, trajectory anchor point loss and divergence regularization loss, and optimizes the network parameters through the backpropagation algorithm.

[0109] The inference sampling module is used to iteratively generate a reconstructed image from the prior distribution of physical noise sampling points by integrating ordinary differential equations.

[0110] The present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described thereon.

[0111] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described thereon.

[0112] The core idea of ​​this application is to transform the low-dose CT image reconstruction problem into a geodesic flow matching problem guided by physical priors. Specifically, this application assumes that noise in low-dose CT images follows signal-related physical laws, rather than being simple independent and identically distributed noise. Therefore, a physical noise encoder is constructed to adaptively estimate noise parameters from the low-dose input, generating a noise prior distribution that conforms to the CT imaging physical model. Based on this, this application models the reconstruction process as a flow matching process from this physical noise prior distribution to the full-dose image distribution. A neural network is used to learn the velocity field and solve ordinary differential equations to achieve rapid generation.

[0113] Meanwhile, this application recognizes the three-dimensional anatomical continuity of CT image sequences, with spatial correspondences between anatomical structures in adjacent slices. To address this, this application proposes an anatomical transport module that estimates the displacement field of anatomical structures between adjacent slices to achieve anatomical alignment at the feature level, effectively transferring anatomical information from adjacent slices to the current slice. Furthermore, this application introduces an interlayer staggered attention mechanism at the bottleneck layer, enhancing the feature transfer of key anatomical structures through anatomical structure-aware masks, thereby achieving the fusion of global anatomical information. Through these technical means, this application, while maintaining the efficient inference advantages of the flow matching method, incorporates CT imaging physical priors and three-dimensional anatomical continuity constraints, achieving high-quality reconstruction of low-dose CT images.

[0114] The present invention has the following beneficial effects:

[0115] First, it offers high physical fidelity. This invention adaptively estimates the signal-related noise parameters of low-dose CT images using a physical noise encoder, constructing a noise prior distribution that conforms to the physical laws of CT imaging. Furthermore, a physical consistency loss is introduced during training to constrain the reconstructed image to conform to the signal-related noise model. This makes the reconstruction results more consistent with the physical mechanisms of CT imaging, avoiding the generation of artifacts that contradict physical laws, and improving the realism and reliability of the reconstructed images.

[0116] Second, it ensures good anatomical structural integrity. This invention estimates the anatomical displacement field between adjacent slices through an anatomical transport module, achieving multi-scale feature alignment and effectively transferring anatomical information from adjacent slices to the current slice. Simultaneously, through an inter-slice staggered attention mechanism, it fuses global anatomical information from adjacent slices at the bottleneck layer, introducing an anatomical structure-aware mask to enhance the feature transfer of key anatomical structures. This 2.5D processing strategy fully utilizes the three-dimensional anatomical continuity of CT image sequences, significantly improving the reconstruction of small structures such as blood vessels and trabeculae, and reducing inconsistencies across slices.

[0117] Third, it exhibits strong robustness. This invention adaptively estimates noise parameters through a physical noise encoder, demonstrating excellent generalization ability for low-dose images with different dose levels and scanning protocols. The trajectory anchor point loss enhances the consistency of the model's predictions at different time points, improving the stability of training and the reliability of the generated results. Attached Figure Description

[0118] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0119] Figure 1 This is a flowchart of the present invention.

[0120] Figure 2 This is a schematic diagram of the overall network architecture of the present invention.

[0121] Figure 3 This is a comparison chart of the prediction results of this invention.

[0122] Figure 4 This is a diagram showing the physical prior starting point for the visualization model estimation in this invention.

[0123] Figure 5 This is a visualization of the anatomical motion field and divergence of the present invention.

[0124] Figure 6 This provides a 2.5D spatial slicing context visualization for the present invention.

[0125] Figure 7 This is a trajectory diagram generated by flow matching for ordinary differential equations (ODE) in this invention. Detailed Implementation

[0126] This invention provides a low-dose CT image reconstruction method based on flow matching, such as... Figure 1 As shown, it includes the following steps:

[0127] Step 1: Estimate signal-related noise parameters of low-dose CT images using a physical noise encoder: Input a low-dose CT image, extract image features through a convolutional neural network, output noise standard deviation and inverse gain parameters, establish a signal-related noise variance model, and construct a physical noise prior distribution based on the estimated noise parameters;

[0128] Step 2: Use the anatomical transport module to align anatomical features between adjacent slices: Input the current slice features, the previous slice features, and the next slice features into the anatomical transport module, estimate the two-dimensional anatomical structure displacement field between adjacent slices, use the estimated displacement field to perform bidirectional spatial deformation on the previous slice features and the next slice features respectively, and use an adaptive gating mechanism to perform weighted fusion of the current slice features and the deformed adjacent slice features to output the aligned anatomical features.

[0129] Step 3: Use interlayer staggered attention mechanism to fuse global anatomical information of adjacent slices: Input the current slice feature and adjacent slice feature pairs into the bottleneck layer, calculate the attention weight through multi-head attention mechanism, introduce an anatomical structure perception mask based on the gray-level difference between adjacent slices to modulate the attention weight, and output the fused bottleneck features.

[0130] Step 4: Generate reconstructed images from low-dose images using a physical guided flow matching training framework: Construct a linear interpolation path from the prior distribution of physical noise to the target distribution based on the estimated noise parameters, predict the velocity field through a neural network, solve ordinary differential equations to generate reconstructed images, and train loss functions including flow matching loss, physical consistency loss, trajectory anchor point loss and divergence regularization loss.

[0131] Step 1 includes: The physical noise encoder includes a convolutional layer, a pooling layer and a fully connected layer. A single-channel low-dose CT image is input, and spatial features are extracted through two or more layers of convolution and pooling operations. The features are compressed into a feature vector through global average pooling, and after mapping through a fully connected layer, two scalar values ​​are output, which are denoted as the first original feature and the second original feature, respectively.

[0132] The first original feature is mapped to a space between a preset lower limit and a preset upper limit of noise standard deviation using a Sigmoid activation function to obtain the noise standard deviation.

[0133] The second original feature is mapped using the Sigmoid function to a value between a preset lower limit and a preset upper limit of inverse gain, resulting in the inverse gain parameter, which is specifically expressed as follows:

[0134] ,

[0135] ,

[0136] in, Indicates the standard deviation of noise. This indicates the lower limit of the preset noise standard deviation. This indicates the upper limit of the preset noise standard deviation. Indicates the first primitive feature, Represents the inverse gain parameter. This indicates the preset lower limit of the inverse gain. This indicates the preset upper limit of inverse gain. Indicates the second primitive feature;

[0137] Based on the estimated noise standard deviation and inverse gain parameter, a signal-related noise variance model is established, specifically expressed as follows:

[0138] ,

[0139] in, Indicates the noise variance. The normalized positive image value is obtained from a low-dose CT image through a linear transformation. , Represents low-dose CT images normalized to the [-1,1] interval;

[0140] Based on the signal-correlated noise variance model, the prior distribution of physical noise is constructed, specifically represented as follows:

[0141] ,

[0142] in, This represents the sampled value of the prior distribution of physical noise. This represents random noise that follows a standard normal distribution. The signal-to-noise ratio correlation coefficient is calculated using the local signal-to-noise ratio. , This represents the local signal-to-noise ratio.

[0143] Step 2 includes:

[0144] Let the current slice features be denoted as the first feature tensor. The first piece of features is denoted as the second feature tensor. All subsequent features are denoted as the third feature tensor. The three feature tensors have the same spatial size and number of channels;

[0145] The second, first, and third feature tensors are concatenated along the channel dimension to obtain a concatenated feature tensor. This concatenated feature tensor is then input into the displacement field estimation network, which outputs a two-dimensional anatomical structure displacement field, specifically represented as follows:

[0146] ,

[0147] in, This represents the displacement field of a two-dimensional anatomical structure, including horizontal and vertical displacement components. This represents the displacement field estimation network. This indicates a splicing operation along the channel dimension;

[0148] The estimated two-dimensional anatomical displacement field is used to perform bidirectional spatial deformation on the features of adjacent slices, and feature resampling is achieved through a spatial transformation network, as specifically expressed as follows:

[0149] ,

[0150] ,

[0151] in, This indicates the features of the previous segment after deformation. This indicates the post-deformation slice characteristics. This represents a spatial transformation function that uses bilinear interpolation to resample the feature map. This indicates an inversion operation on the displacement field, used to achieve reverse deformation;

[0152] By fusing the features of the current slice with the features of the deformed adjacent slices through an adaptive gating mechanism, the learnable gating weights are first calculated, specifically as follows:

[0153] ,

[0154] in, This represents the gate weight tensor, which has the same spatial dimensions as the feature tensor. This represents a gated weight generation network consisting of convolutional layers;

[0155] The average feature is obtained by averaging the features of the previous slice after deformation, the current slice feature, and the subsequent slice feature. Then, the current slice feature and the average feature are weighted and fused using gating weights. Specifically, this is represented as follows:

[0156] ,

[0157] in, This indicates the output characteristics of the anatomical transport module. This indicates the average characteristic.

[0158] In step 2, the displacement field estimation network adopts a grouped convolutional structure. The number of channels of the input spliced ​​feature tensor is three times the number of current feature channels, and the number of channels of the output two-dimensional anatomical structure displacement field is two. The intermediate layer of the network adopts a group normalization layer and SiLU activation function, and the output layer adopts a zero-initialization convolutional module, so that the output of the network is zero displacement field in the initial state.

[0159] Step 3 includes:

[0160] Let the features of the current slice be denoted as the fourth feature tensor, and the features of the previous slice be denoted as the fifth feature tensor. The subsequent features are denoted as the sixth feature tensor. The three feature tensors have the same spatial size and number of channels. The fifth and sixth feature tensors are concatenated along the channel dimension to obtain the context feature pair.

[0161] The average context feature is obtained by averaging the two feature tensors in the context feature pair, specifically represented as:

[0162] ,

[0163] in, Represents average contextual features;

[0164] The query features, key features, and value features are calculated separately using convolutional projection, as follows:

[0165] ,

[0166] ,

[0167] ,

[0168] in, Indicates query characteristics, Indicates key features, Indicates value characteristics, These represent the query projection matrix, key projection matrix, and value projection matrix, respectively, all implemented using 1×1 convolutional layers. Represents the fourth characteristic tensor;

[0169] The query features and key features are reshaped into two-dimensional spatial sequence forms, and the attention score matrix is ​​calculated, as follows:

[0170] ,

[0171] in, Represents the attention score matrix. This represents matrix multiplication of query features and key features. This represents the feature dimension of each attention head, and Softmax represents the Softmax normalization operation along the last dimension.

[0172] An anatomical structure-aware mask is introduced, and the mask matrix is ​​calculated based on the gray-level differences between adjacent slices, specifically represented as follows:

[0173] ,

[0174] in, The anatomical structure perception mask matrix represents the first... Line number The element values ​​of the column, Indicates the spatial position of the first and last segments. The difference in gray values ​​at each location Indicates the spatial position of the first and last segments. The difference in gray values ​​at each location This represents a learnable temperature parameter; exp is an exponential function.

[0175] The anatomical structure perception mask is logarithmically superimposed onto the attention score matrix to obtain the modulated attention score matrix, specifically represented as follows:

[0176] ,

[0177] in, This represents the modulated attention score matrix. Represents the natural logarithm operation;

[0178] The modulated attention score matrix is ​​multiplied by the value features, projected onto the output, and then added to the original input features to obtain the fused bottleneck features, specifically represented as follows:

[0179] ,

[0180] in, This represents the output features of the interlayer staggered attention module. This represents the output projection matrix, implemented by a 1×1 convolutional layer.

[0181] Step 4 includes the following steps:

[0182] Step 4.1: Constructing a linear interpolation path: Let the initial point sampled from the prior distribution of physical noise be denoted as the first sampling point, and the target full-dose image be denoted as the second sampling point. Define a linear interpolation path within the time interval [0,1], and add a small amount of Gaussian noise during the interpolation process to ensure numerical stability, specifically as follows:

[0183] ,

[0184] in, Indicates time Interpolation point at that location, This represents a time variable, where t takes values ​​in the range [0,1]. Indicates the first sampling point. Indicates the second sampling point. This indicates the preset minimum noise figure;

[0185] Step 4.2, Define the velocity field: Let the velocity field predicted by the neural network be denoted as the predicted velocity field, which satisfies the ordinary differential equation ,in This represents the velocity field predicted by the neural network. Represents the trainable parameters of a neural network. This represents the input low-dose CT image; d represents the differential symbol.

[0186] Step 4.3, Calculate the flow matching loss: Calculate the mean square error between the predicted velocity field and the true velocity field. The true velocity field is defined as the difference between the endpoint and the starting point, specifically expressed as:

[0187] ,

[0188] in, Indicates the stream matching loss. This means averaging over all possible paths. Represents the mathematical expectation. Denotes the square of the L2 norm. Represents the true velocity field;

[0189] Step 4.4, Calculate the physical consistency loss: The reconstructed image predicted by the constraint neural network conforms to the signal-correlated noise model. First, the target image is reconstructed based on the predicted velocity field, specifically as follows:

[0190] ,

[0191] in, This represents the target image predicted by the neural network.

[0192] The noise-to-signal ratio between the predicted target image and the input low-dose image is calculated and specifically expressed as follows:

[0193] ,

[0194] Wherein, NSR represents the noise-to-signal ratio. Indicates local spatial average;

[0195] The physical consistency loss is constructed based on the noise-to-signal ratio, and is specifically expressed as follows:

[0196] ,

[0197] in, Indicates the loss of physical consistency. This represents a time-weighted function that varies over time, used to enhance the constraint strength of the intermediate path;

[0198] Step 4.5, Calculate trajectory anchor point loss: Randomly select two adjacent time points, constraining the neural network to maintain consistency in the target image predicted at the two time points. First, randomly generate the offset, specifically as follows:

[0199] ,

[0200] in, Indicates a random offset. This indicates the preset maximum offset. Indicates from 0 to Uniform sampling between Indicates a random positive or negative sign;

[0201] The anchor point time point is calculated as follows:

[0202] ,

[0203] in, This indicates the anchor point in time; `clip` represents the clipping operation. This indicates the preset minimum time value. This indicates the preset maximum time value;

[0204] Calculate the predicted target images at the original time point and the anchor time point respectively, and then calculate the mean square error, which is specifically expressed as:

[0205] ,

[0206] in, Indicates trajectory anchor point loss. This represents a comprehensive optimization of all possible combinations of path start points, path end points, master time points, and anchor time points. Indicates time The predicted target image at that location, Indicates time The predicted target image at the location;

[0207] Step 4.6: Calculate the divergence regularization loss: Calculate the divergence of the two-dimensional anatomical structure displacement field output by each anatomical transport module in Step 2, and take the absolute value as the regularization term, specifically expressed as:

[0208] ,

[0209] in, This represents the divergence regularization loss. This represents the horizontal component of the displacement field of a two-dimensional anatomical structure. This represents the vertical component of the displacement field of a two-dimensional anatomical structure. This represents the partial derivative of the horizontal component in the horizontal direction. This represents the partial derivative of the vertical component in the vertical direction;

[0210] Step 4.7: Construct the total loss function: The total loss function is calculated by weighting and summing the various losses, specifically as follows:

[0211] ,

[0212] in, Represents the total loss function. This represents the weighting coefficient for the physical consistency loss. This represents the trajectory anchor point loss weighting coefficient. This represents the divergence regularization loss weight coefficient.

[0213] Step 5 includes:

[0214] Initialization: Set the total number of steps to N, and the time step size to N. initial point ;

[0215] Iterative solution: For From 0 to Perform the following operations:

[0216] ,

[0217] ,

[0218] ,

[0219] in, Indicates the first The timing of the step This represents the predicted velocity field at the current time point. Indicates the state of the current iteration step;

[0220] Output: After the iteration is complete, The image is reconstructed by cropping it to a preset value range.

[0221] This embodiment provides a low-dose CT image reconstruction system based on flow matching, such as... Figure 2 As shown, its network structure parameters are configured as follows:

[0222] The physical noise encoder consists of three convolutional layers, three pooling layers, and two fully connected layers. The convolutional layers have 32, 64, and 128 channels respectively, with a 3×3 kernel size, a stride of 1, and padding of 1. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The pooling layers use 2×2 max pooling. The fully connected layers map the feature vectors to 128 dimensions, then to a 2-dimensional output. The preset noise standard deviation lower limit is 0.001, and the upper limit is 0.5. The preset inverse gain lower limit is 0.001, and the upper limit is 0.1.

[0223] The U-Net backbone network has 4 input channels and 1 output channel, with an initial total of 96 channels. The encoding path contains 4 downsampling levels with channel multiplication factors of 1, 1, 2, 2, and 4, respectively. Each downsampling level contains 2 residual blocks, which employ group normalization and the SiLU activation function. The output channel count is the same as the current level's channel count. The attention mechanism is implemented at spatial resolutions of 16×16 and 8×32, with 4 attention heads. The decoding path uses skip connections, fusing encoder features through channel concatenation.

[0224] Ten anatomical transport modules are deployed after each residual block of the encoding path is embedded. The displacement field estimation network in each anatomical transport module consists of three sets of convolutional layers: the first set is a grouped convolution with three times the number of current feature channels as input and the same number of output channels as current feature channels, and a kernel size of 1×1; the second set is a standard convolution with half the number of input and output channels as input and half the number of current feature channels, a kernel size of 3×3, and padding of 1; the third set is a zero-initialization convolution with half the number of current feature channels as input, two output channels, a kernel size of 3×3, and padding of 1.

[0225] The interlayer staggered attention module is placed in the bottleneck layer of U-Net, with 4 attention heads. The query, key-value projection matrices, and output projection matrices are all 1×1 convolutional layers. The temperature parameter of the anatomical structure-aware mask is initialized to 1.0 and used as a learnable parameter in training.

[0226] In the flow matching training framework, the minimum noise coefficient is set to 1e-4, the preset maximum offset is set to 0.1, the preset minimum time value is set to 0.01, and the preset maximum time value is set to 0.99. The weight coefficients for physical consistency loss, trajectory anchor point loss, and divergence regularization loss are set to 0.10, 0.05, and 0.01 respectively.

[0227] The training process includes the following steps:

[0228] The first step is data preparation. Acquire CT image sequences, including low-dose scans and corresponding full-dose scans. Normalize the images to the [-1,1] interval and construct training triplets, each containing the current low-dose slice, the previous low-dose slice, the next low-dose slice, and the corresponding full-dose slice.

[0229] The second step is warm-up training. In the first 500 training steps, only the physical noise encoder is trained. Low-dose images are input, and the local image variance is calculated as a supervision signal. The parameters of the physical noise encoder are optimized through mean squared error loss so that it can accurately estimate the signal-related noise parameters.

[0230] The third step is joint training. After warm-up training, the physical noise encoder, the anatomical transport module, the interlayer staggered attention module, and the U-Net backbone network are trained simultaneously. The AdamW optimizer is used with a learning rate of 1e-4, and the physical noise encoder's learning rate is set to 1e-3. The weight decay coefficient is set to 0. The batch size is set to 4. To adapt to memory limitations, each batch can be split into sub-batches of size 1 for gradient accumulation. An exponential moving average strategy is used with a decay rate of 0.9999 to stabilize training and improve inference performance.

[0231] Step 4: Loss Calculation. In each training iteration, the time variable t is randomly sampled, a linear interpolation path is constructed, and the flow matching loss is calculated. The target image is reconstructed based on the predicted velocity field, and the physical consistency loss is calculated. Anchor point time points are randomly sampled, and the trajectory anchor point loss is calculated. The displacement fields of all dissected transport modules are collected, such as... Figure 5 As shown, the divergence regularization loss is calculated. The total loss is obtained by weighted summation of the various losses and then updated through backpropagation.

[0232] Step 5: Model saving. A model checkpoint is saved every 5000 steps, including network parameters, exponential moving average parameters, and the current training step number.

[0233] The reasoning process specifically includes the following steps:

[0234] Step 1: Input Processing. Load the pre-trained model and exponential moving average parameters, and set the model to evaluation mode. Input a low-dose CT image sequence, select the current slice to be reconstructed and its adjacent slices, and normalize them to the [-1,1] interval, as shown below. Figure 6 As shown.

[0235] The second step is noise prior generation. The low-dose image is input into the physical noise encoder to estimate the noise standard deviation and inverse gain parameter, constructing a signal-correlated noise variance model, and generating physical noise prior distribution sampling points, such as... Figure 4 As shown.

[0236] The third step is to solve the ordinary differential equations. The total number of steps is set to 50, and the time step size is 0.02. Iteration begins with sampling points from the prior distribution of physical noise. For each time step, the current state, time step, and conditional inputs (low-dose image, adjacent slices) are input into the network to predict the velocity field. The state is updated using the Euler method, as follows: Figure 7 As shown.

[0237] The fourth step is post-processing. After iteration, the state is cropped to the [-1, 1] interval, linearly transformed to the 0-255 grayscale range, and the reconstructed image is output, as shown below. Figure 3 As shown.

[0238] This embodiment verifies the effectiveness of the method of the present invention on a clinical CT dataset. The dataset contains CT scans of 10 patients, each containing approximately 200 slices, with the low-dose scan dose being 25% of the full dose. The dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio.

[0239] Peak signal-to-noise ratio (PSNR) and structural similarity index were used as evaluation metrics. Experimental results show that the method of this invention achieves an average PSNR of 42.73 dB and a structural similarity index of 0.9677 on the test set.

[0240] Comparative experiment:

[0241] To evaluate the performance of this invention, it is compared with the following representative methods:

[0242] RED-CNN: A convolutional model that uses a hierarchical encoder-decoder architecture with fast connections.

[0243] N2N: A convolutional model originally proposed for self-supervised learning on noisy CT images. For fair comparison, the N2N architecture is used for supervised learning.

[0244] EDCNN: Considers a convolutional model that uses trainable Sobel kernels for edge detection and dense connections.

[0245] WGAN: Considers an adversarial model that uses convolutional generator and discriminator subnetworks.

[0246] DU-GAN: An adversarial model using convolutional generator and discriminator subnetworks is considered. IDDPM: A diffusion model with a convolutional backbone and an attention mechanism is considered, which generates NDCT images from Gaussian noise images and provides additional guidance with LCDT images as input.

[0247] UFormer: Considers an efficient Transformer model that uses a hierarchical encoder-decoder architecture and a self-attention mechanism based on local windows.

[0248] LIT-Former: Considers an efficient Transformer model originally proposed for processing 3D images, which uses separate Transformer modules for in-plane and through-plane dimensions. By removing the through-plane module, LIT-former is applied to 2D images.

[0249] ViMEDNet: Considers a state-space model that uses a hierarchical encoder-decoder architecture equipped with a spatial SSM module.

[0250] The evaluation metrics used were Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity to Sound (SSIM). The quantitative results on the test set are shown in the table below:

[0251] Table 1: Comparison of quantitative results of different methods on the test set (mean ± standard deviation)

[0252]

[0253] Experimental results show that the flow-matching-based low-dose CT image reconstruction method proposed in this invention achieves optimal levels in both PSNR and SSIM, with a PSNR of 42.73 dB significantly surpassing existing methods and exhibiting the lowest standard deviation, demonstrating the effectiveness and stability of the method. This method achieves the best balance between pixel-level fidelity and structural preservation capability by fusing CT physical priors and three-dimensional anatomical continuity constraints, providing reliable technical support for the clinical application of low-dose CT images.

[0254] This embodiment proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method steps of this application.

[0255] This embodiment proposes a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the method described in this application, which will not be repeated here.

[0256] This embodiment proposes a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in this application, and will not be repeated here.

[0257] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0258] This invention provides a method and system for low-dose CT image reconstruction based on flow matching. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A low-dose CT image reconstruction method based on flow matching, characterized in that, Includes the following steps: Step 1: Estimate signal-related noise parameters of low-dose CT images using a physical noise encoder: Input a low-dose CT image, extract image features through a convolutional neural network, output noise standard deviation and inverse gain parameters, establish a signal-related noise variance model, and construct a physical noise prior distribution based on the estimated noise parameters; Step 2: Use the anatomical transport module to align anatomical features between adjacent slices: Input the current slice features, the previous slice features, and the next slice features into the anatomical transport module, estimate the two-dimensional anatomical structure displacement field between adjacent slices, use the estimated displacement field to perform bidirectional spatial deformation on the previous slice features and the next slice features respectively, and use an adaptive gating mechanism to perform weighted fusion of the current slice features and the deformed adjacent slice features to output the aligned anatomical features; Step 3: Use interlayer staggered attention mechanism to fuse global anatomical information of adjacent slices: Input the current slice feature and adjacent slice feature pairs into the bottleneck layer, calculate the attention weight through multi-head attention mechanism, introduce an anatomical structure perception mask based on the gray-level difference between adjacent slices to modulate the attention weight, and output the fused bottleneck features. Step 4: Generate reconstructed images from low-dose images using a physical guided flow matching training framework: Construct a linear interpolation path from the prior distribution of physical noise to the target distribution based on the estimated noise parameters, predict the velocity field through a neural network, solve ordinary differential equations to generate reconstructed images, and train loss functions including flow matching loss, physical consistency loss, trajectory anchor point loss and divergence regularization loss.

2. The method according to claim 1, characterized in that, Step 1 includes: The physical noise encoder includes a convolutional layer, a pooling layer and a fully connected layer. A single-channel low-dose CT image is input, and spatial features are extracted through two or more layers of convolution and pooling operations. The features are compressed into a feature vector through global average pooling, and after mapping through a fully connected layer, two scalar values ​​are output, which are denoted as the first original feature and the second original feature, respectively. The first original feature is mapped to a space between a preset lower limit and a preset upper limit of noise standard deviation using a Sigmoid activation function to obtain the noise standard deviation. The second original feature is mapped using the Sigmoid function to a value between a preset lower limit and a preset upper limit of inverse gain, resulting in the inverse gain parameter, which is specifically expressed as follows: ， ， in, Indicates the standard deviation of noise. This indicates the lower limit of the preset noise standard deviation. This indicates the upper limit of the preset noise standard deviation. Indicates the first primitive feature, Represents the inverse gain parameter. This indicates the preset lower limit of the inverse gain. This indicates the preset upper limit of inverse gain. Indicates the second primitive feature; Based on the estimated noise standard deviation and inverse gain parameter, a signal-related noise variance model is established, specifically expressed as follows: ， in, Indicates the noise variance. The normalized positive image value is obtained from a low-dose CT image through a linear transformation. , Represents low-dose CT images normalized to the [-1,1] interval; Based on the signal-correlated noise variance model, the prior distribution of physical noise is constructed, specifically represented as follows: ， in, This represents the sampled value of the prior distribution of physical noise. This represents random noise that follows a standard normal distribution. The signal-to-noise ratio correlation coefficient is calculated using the local signal-to-noise ratio. , This represents the local signal-to-noise ratio.

3. The method according to claim 2, characterized in that, Step 2 includes: Let the current slice features be denoted as the first feature tensor. The first piece of features is denoted as the second feature tensor. All subsequent features are denoted as the third feature tensor. The three feature tensors have the same spatial size and number of channels; The second, first, and third feature tensors are concatenated along the channel dimension to obtain a concatenated feature tensor. This concatenated feature tensor is then input into the displacement field estimation network, which outputs a two-dimensional anatomical structure displacement field, specifically represented as follows: ， in, This represents the displacement field of a two-dimensional anatomical structure, including horizontal and vertical displacement components. This represents the displacement field estimation network. This indicates a splicing operation along the channel dimension; The estimated two-dimensional anatomical displacement field is used to perform bidirectional spatial deformation on the features of adjacent slices, and feature resampling is achieved through a spatial transformation network, as specifically expressed as follows: ， ， in, This indicates the features of the previous segment after deformation. This indicates the post-deformation slice characteristics. This represents a spatial transformation function that uses bilinear interpolation to resample the feature map. This indicates an inversion operation on the displacement field, used to achieve reverse deformation; By fusing the features of the current slice with the features of the deformed adjacent slices through an adaptive gating mechanism, the learnable gating weights are first calculated, specifically as follows: ， in, This represents the gate weight tensor, which has the same spatial dimensions as the feature tensor. This represents a gated weight generation network consisting of convolutional layers; The average feature is obtained by averaging the features of the previous slice after deformation, the current slice feature, and the subsequent slice feature. Then, the current slice feature and the average feature are weighted and fused using gating weights. Specifically, this is represented as follows: ， in, This indicates the output characteristics of the anatomical transport module. This indicates the average characteristic.

4. The method according to claim 3, characterized in that, In step 2, the displacement field estimation network adopts a grouped convolutional structure. The number of channels of the input spliced ​​feature tensor is three times the number of current feature channels, and the number of channels of the output two-dimensional anatomical structure displacement field is two. The intermediate layer of the network adopts a group normalization layer and SiLU activation function, and the output layer adopts a zero-initialization convolutional module, so that the output of the network is zero displacement field in the initial state.

5. The method according to claim 4, characterized in that, Step 3 includes: Let the features of the current slice be denoted as the fourth feature tensor, and the features of the previous slice be denoted as the fifth feature tensor. The subsequent features are denoted as the sixth feature tensor. The three feature tensors have the same spatial size and number of channels. The fifth and sixth feature tensors are concatenated along the channel dimension to obtain the context feature pair. The average context feature is obtained by averaging the two feature tensors in the context feature pair, specifically represented as: ， in, Represents average contextual features; The query features, key features, and value features are calculated separately using convolutional projection, as follows: ， ， ， in, Indicates query characteristics, Indicates key features, Indicates value characteristics, These represent the query projection matrix, key projection matrix, and value projection matrix, respectively, all implemented using 1×1 convolutional layers. Represents the fourth characteristic tensor; The query features and key features are reshaped into two-dimensional spatial sequence forms, and the attention score matrix is ​​calculated, as follows: ， in, Represents the attention score matrix, This represents matrix multiplication of query features and key features. This represents the feature dimension of each attention head, and Softmax represents the Softmax normalization operation along the last dimension. An anatomical structure-aware mask is introduced, and the mask matrix is ​​calculated based on the gray-level differences between adjacent slices, specifically represented as follows: ， in, The anatomical structure perception mask matrix represents the first... Line number The element values ​​of the column, Indicates the spatial position of the first and last segments. The difference in gray values ​​at each location Indicates the spatial position of the first and last segments. The difference in gray values ​​at each location This represents a learnable temperature parameter; exp is an exponential function. The anatomical structure perception mask is logarithmically superimposed onto the attention score matrix to obtain the modulated attention score matrix, specifically represented as follows: ， in, This represents the modulated attention score matrix. Represents the natural logarithm operation; The modulated attention score matrix is ​​multiplied by the value features, projected onto the output, and then added to the original input features to obtain the fused bottleneck features, specifically represented as follows: ， in, This represents the output features of the interlayer staggered attention module. This represents the output projection matrix, implemented by a 1×1 convolutional layer.

6. The method according to claim 5, characterized in that, Step 4 includes the following steps: Step 6.1: Constructing a linear interpolation path: Let the initial point sampled from the prior distribution of physical noise be denoted as the first sampling point, and the target full-dose image be denoted as the second sampling point. Define a linear interpolation path within the time interval [0,1], and add a small amount of Gaussian noise during the interpolation process to ensure numerical stability, specifically as follows: ， in, Indicates time Interpolation point at that location, This represents a time variable, where t takes values ​​in the range [0,1]. Indicates the first sampling point. Indicates the second sampling point. This indicates the preset minimum noise figure; Step 6.2, Define the velocity field: Let the velocity field predicted by the neural network be denoted as the predicted velocity field, which satisfies the ordinary differential equation ,in This represents the velocity field predicted by the neural network. Represents the trainable parameters of a neural network. This represents the input low-dose CT image; d represents the differential symbol. Step 6.3, Calculate the flow matching loss: Calculate the mean square error between the predicted velocity field and the true velocity field. The true velocity field is defined as the difference between the endpoint and the starting point, specifically expressed as: ， in, Indicates the stream matching loss. This means averaging over all possible paths. Represents the mathematical expectation. Denotes the square of the L2 norm. Represents the true velocity field; Step 6.4, Calculate the physical consistency loss: The reconstructed image predicted by the constraint neural network conforms to the signal-correlated noise model. First, the target image is reconstructed based on the predicted velocity field, specifically as follows: ， in, This represents the target image predicted by the neural network. The noise-to-signal ratio between the predicted target image and the input low-dose image is calculated and specifically expressed as follows: ， Wherein, NSR represents the noise-to-signal ratio. Indicates local spatial average; The physical consistency loss is constructed based on the noise-to-signal ratio, and is specifically expressed as follows: ， in, Indicates the loss of physical consistency. This represents a time-weighted function that varies over time, used to enhance the constraint strength of the intermediate path; Step 6.5, Calculate trajectory anchor point loss: Randomly select two adjacent time points, constraining the neural network to maintain consistency in the target images predicted at the two time points. First, randomly generate the offset, specifically represented as follows: ， in, Indicates a random offset. This indicates the preset maximum offset. Indicates from 0 to Uniform sampling between Indicates a random positive or negative sign; The anchor point time point is calculated as follows: ， in, This indicates the anchor point in time; `clip` represents the clipping operation. This indicates the preset minimum time value. This indicates the preset maximum time value; Calculate the predicted target images at the original time point and the anchor time point respectively, and then calculate the mean square error, which is specifically expressed as: ， in, Indicates trajectory anchor point loss. This represents a comprehensive optimization of all possible combinations of path start points, path end points, master time points, and anchor time points. Indicates time The predicted target image at that location, Indicates time The predicted target image at the location; Step 6.6: Calculate the divergence regularization loss: Calculate the divergence of the two-dimensional anatomical structure displacement field output by each anatomical transport module in Step 2, and take the absolute value as the regularization term, specifically expressed as: ， in, Represents the divergence regularization loss. This represents the horizontal component of the displacement field of a two-dimensional anatomical structure. This represents the vertical component of the displacement field of a two-dimensional anatomical structure. This represents the partial derivative of the horizontal component in the horizontal direction. This represents the partial derivative of the vertical component in the vertical direction; Step 6.7: Construct the total loss function: The total loss function is calculated by weighted summation of all losses, specifically expressed as: ， in, Represents the total loss function. This represents the weighting coefficient for the physical consistency loss. This represents the trajectory anchor point loss weighting coefficient. This represents the divergence regularization loss weight coefficient.

7. The method according to claim 6, characterized in that, Step 4 includes: Initialization: Set the total number of steps to N, and the time step size to N. initial point ; Iterative solution: For From 0 to Perform the following operations: ， ， ， in, Indicates the first The timing of the step This represents the predicted velocity field at the current time point. Indicates the state of the current iteration step; Output: After the iteration is complete, The image is cropped to a preset value range to obtain the reconstructed image.

8. A CT image reconstruction system based on the method described in any one of claims 1 to 7, characterized in that, include: The physical noise encoder module is used to take low-dose CT images as input, estimate signal-related noise parameters, including noise standard deviation and inverse gain parameters, through a convolutional neural network, and construct a prior distribution of physical noise based on the estimated parameters. The anatomical transport module array consists of multiple anatomical transport modules. Each anatomical transport module is embedded into different levels of the U-Net encoding path. Each anatomical transport module receives the current slice features, the previous slice features, and the next slice features at the current level, estimates the displacement field of the two-dimensional anatomical structure, and performs bidirectional spatial deformation. It fuses features through an adaptive gating mechanism to achieve multi-scale anatomical feature alignment. The interlayer staggered attention module is set in the bottleneck layer of U-Net. It receives the current slice feature and the feature pairs of adjacent slices, calculates the attention weight through a multi-head attention mechanism, introduces an anatomical structure perception mask based on the gray-level difference of adjacent slices to modulate the attention weight, and outputs the fused bottleneck features. The flow matching training module is used to construct a linear interpolation path from the prior distribution of physical noise to the target distribution. It predicts the velocity field through a neural network, calculates the total loss function including flow matching loss, physical consistency loss, trajectory anchor point loss and divergence regularization loss, and optimizes the network parameters through the backpropagation algorithm. The inference sampling module is used to iteratively generate a reconstructed image from the prior distribution of physical noise sampling points by integrating ordinary differential equations.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1 to 7.

Images