A low-dose CT reconstruction method and device

Abstract

The application discloses a low-dose CT reconstruction method and device, relates to the technical field of image processing, and comprises the following steps: acquiring a low-dose CT image; adopting a trained deep learning network to process the low-dose CT image in a projection domain, a chord diagram domain and an image domain respectively; in the projection domain, the low-dose CT image is processed to obtain updated projection data; in the chord diagram domain, the updated projection data is processed to obtain updated chord diagram data; in the image domain, the updated chord diagram data is processed to obtain a reconstructed low-dose CT image; wherein the trained deep learning network takes data of a preset category as a training data set, trains an initial deep learning network, and is obtained by performing intermediate supervision in the training process and fine-tuning the deep learning network in the training. The application can realize good reconstruction of the low-dose CT image.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a low-dose CT reconstruction method and apparatus. Background Technology

[0002] Computed tomography (CT) is an important tool for medical diagnosis, but the high radiation dose of traditional CT can pose health risks. Low-dose CT (LDCT) reduces radiation hazards by lowering the X-ray dose, but the reduced dose leads to increased noise and missing information in the projection data, resulting in decreased image quality (such as artifacts and blurring). Currently, the core challenge of low-dose CT reconstruction lies in how to reconstruct high-quality images that meet clinical diagnostic needs under low-dose conditions.

[0003] In related technologies, low-dose CT reconstruction mainly includes analytical methods, iterative reconstruction, deep learning methods, and hybrid methods. Analytical methods include filtered back projection (FBP), whose core is the Radon inverse transform. It assumes that the projection data is noise-free and complete. However, in low-dose CT, photon statistical noise is significantly enhanced, and direct FBP reconstruction amplifies noise (especially high-frequency noise), leading to granular artifacts and a severe decrease in signal-to-noise ratio (SNR). Iterative reconstruction includes statistical iterative reconstruction (SIR) and model iterative reconstruction (MBIR). Iterative reconstruction requires repeated orthographic and back-projection operations, with each iteration requiring several times the computation of FBP, resulting in high computational complexity and poor real-time performance. Deep learning methods include projection domain preprocessing and image preprocessing. End-to-end reconstruction and hybrid models: Deep learning methods have made significant progress in low-dosage CT image reconstruction research. However, data-driven deep learning has certain limitations, such as data dependence, limited generalization, and lack of physical consistency. Based on this, combining deep learning with iterative reconstruction forms a hybrid approach. Traditional CT reconstruction methods can comprehensively describe the physical laws of the CT reconstruction process, and combining this with the ability of deep learning to learn complex statistical priors from large amounts of data has made hybrid methods a research trend in recent years. Although hybrid methods have achieved significant results, existing reconstruction methods are limited by data acquisition in clinical practice, resulting in low data quality and low reconstruction efficiency.

[0004] Therefore, there is an urgent need to provide a low-dose CT reconstruction method to improve the aforementioned shortcomings of existing technologies. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides a low-dose CT reconstruction method and apparatus. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] In a first aspect, the present invention provides a low-dose CT reconstruction method, comprising:

[0007] Acquire low-dose CT images;

[0008] A trained deep learning network is used to process low-dose CT images in the projection domain, chordal domain, and image domain. In the projection domain, the low-dose CT images are processed to obtain updated projection data. In the chordal domain, the updated projection data are processed to obtain updated chordal data. In the image domain, the updated chordal data are processed to obtain reconstructed low-dose CT images.

[0009] The trained deep learning network is trained using data of a preset category as the training dataset. The initial deep learning network is trained, and intermediate supervision is performed during the training process to fine-tune the deep learning network during training.

[0010] Secondly, the present invention also provides a low-dose CT reconstruction device, comprising:

[0011] Image acquisition module, used to acquire low-dose CT images;

[0012] The image processing module is used to process low-dose CT images in the projection domain, chordal domain, and image domain using a trained deep learning network. In the projection domain, the low-dose CT image is processed to obtain updated projection data; in the chordal domain, the updated projection data is processed to obtain updated chordal data; and in the image domain, the updated chordal data is processed to obtain reconstructed low-dose CT images.

[0013] The trained deep learning network is trained using data of a preset category as the training dataset. The initial deep learning network is trained, and intermediate supervision is performed during the training process to fine-tune the deep learning network during training.

[0014] The beneficial effects of this invention are:

[0015] The present invention provides a low-dose CT reconstruction method and apparatus, which processes the low-dose CT image in the projection domain to obtain updated projection data, processes the updated projection data in the chordal domain to obtain updated chordal data, and processes the updated chordal data in the image domain to obtain a reconstructed low-dose CT image. It takes into account the noise in the projection domain, chordal domain and image domain, and can preserve the tissue structure of non-artifact areas to the greatest extent, thus achieving good reconstruction of low-dose CT images.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart of a low-dose CT reconstruction method provided in an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of a chord graph domain denoising model provided in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of a first multi-scale downsampling fusion module provided in an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of a first upsampling feature fusion module provided in an embodiment of the present invention;

[0021] Figure 5 This is a schematic diagram of an image domain denoising model provided in an embodiment of the present invention;

[0022] Figure 6 This is a flowchart of an embodiment of the present invention for generating low-dose CT images corresponding to conventional dose CT based on CT scanning mechanism. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0024] Please see Figure 1 , Figure 1 This is a flowchart of a low-dose CT reconstruction method provided by an embodiment of the present invention. The low-dose CT reconstruction method provided by the present invention includes:

[0025] S101. Acquire low-dose CT images.

[0026] Specifically, in this embodiment, low-dose CT (LDCT) images are obtained by reducing the X-ray tube current, i.e. reducing the amount of photons emitted. While reducing the radiation dose, low-dose CT images increase image noise. Therefore, it is necessary to reconstruct low-dose CT images to improve the noise effect.

[0027] It should be noted that low-dose CT images involve three key data domains in the CT imaging process: the projection domain, the sinogram domain, and the image domain. These domains correspond to different stages of CT imaging, and the data characteristics and challenges of each domain are different under low-dose conditions.

[0028] The projection domain data consists of the raw signals directly received by the detector during a CT scan, i.e., the attenuation values ​​of X-rays after passing through the object (raw projection data). These data are unreconstructed one-dimensional or two-dimensional measurements. The significant characteristics of projection domain data are: significantly increased noise, i.e., reducing the X-ray dose leads to insufficient photon counts, resulting in enhanced quantum noise (Poisson noise) and electronic background noise (Gaussian noise) in the projection data; and data sparsity, i.e., certain low-dose protocols (such as sparse angle scanning) can lead to incomplete projection data, introducing artifacts during reconstruction.

[0029] Sine map data is a two-dimensional matrix (resembling a sine curve, hence the name "sine map") formed by arranging projection data according to the scanning angle and detector position. The projection values ​​of each detector element at different angles are organized into a row, reflecting the projection of the object at a certain angle. The significant characteristics of sine map data are: stripe noise and artifacts, i.e., under low-dose conditions, noise in the sine map presents a random stripe pattern, which may lead to stripe artifacts in the reconstructed image domain; and structural blurring, i.e., noise can obscure detailed features in the sine map (such as the projection of tiny lesions).

[0030] Image domain data is transformed from projection data or chordal data into cross-sectional images (two-dimensional / three-dimensional volume data) through reconstruction algorithms (such as filtered back projection (FBP), iterative reconstruction algorithms, or deep learning reconstruction). Significant characteristics of image domain data include: noise and artifacts, specifically, low-dose noise leads to mottle noise, streak artifacts, and low-contrast structural blurring in the image domain; and reduced resolution, meaning noise suppression may sacrifice image details (such as microcalcifications or lung nodules).

[0031] S102. The trained deep learning network is used to process the low-dose CT image in the projection domain, chordal domain and image domain respectively. In the projection domain, the low-dose CT image is processed to obtain updated projection data. In the chordal domain, the updated projection data is processed to obtain updated chordal data. In the image domain, the updated chordal data is processed to obtain the reconstructed low-dose CT image.

[0032] The trained deep learning network is trained using data of a preset category as the training dataset. The initial deep learning network is trained, and intermediate supervision is performed during the training process to fine-tune the deep learning network during training.

[0033] Specifically, in this embodiment, the trained deep learning network includes a projection domain denoising model; in the projection domain, low-dose CT images are processed to obtain updated projection data, including:

[0034] Construct a projection data generation model, represented as follows:

[0035] P = W + δ;

[0036] Where W represents the quantum number received by the detector. δ represents the background noise of electrons.

[0037] The background noise follows a non-stationary Gaussian distribution with zero mean, expressed as:

[0038]

[0039] Where, σ 2 Represents the variance of the noise;

[0040] Since the background noise of electrons follows a Gaussian distribution, the first conditional distribution can be obtained, which is expressed as:

[0041]

[0042] Quantum noise W i Following a Poisson distribution, it can be represented as:

[0043]

[0044] Among them, G i G represents the intensity of radiation under ideal conditions. According to Beer-Lambert's law, G i satisfy:

[0045]

[0046] Among them, Y i G represents ideal chord diagram data. 0i Indicates the actual intensity of the radiation;

[0047] According to quantum noise W i Following the Poisson distribution and combined with the Beer-Lambert law, we obtain the second conditional distribution, which is expressed as:

[0048]

[0049] Where i represents the index of the quantum;

[0050] Based on the first and second conditional distributions, the projection domain denoising model is obtained, expressed as:

[0051]

[0052] Based on the maximum a posteriori estimation theory, the posterior distribution of the projection domain denoising model is obtained, expressed as:

[0053]

[0054] Simplifying the posterior distribution of the projection domain denoising model yields the updated projection domain denoising model, expressed as:

[0055]

[0056] The projection domain denoising model is optimized and updated using the block coordinate descent method to obtain the quantum noise W. i The objective function is expressed as:

[0057]

[0058] In the first stage, the memory-limited BFGS algorithm is used to address the quantum noise W. i The objective function is solved, and in the second stage, the damped Newton method is used to solve the quantum noise W. i Solving the objective function yields the optimal quantum noise, which is the optimal quantum number W received by the detector.

[0059] Updated projection data is obtained based on the quantum number W received by the optimal detector.

[0060] It should be noted that in this embodiment, in the first stage, the memory-limited BFGS algorithm is used to process the quantum noise W. i The objective function is solved to quickly escape the flat region of the objective function, achieving efficient initial convergence using first-order and approximate second-order information; in the second stage, the damped Newton method is used to address the quantum noise W. i The objective function is solved, and when the solution is close to the optimal solution, the precise Hessian matrix is ​​used to achieve superlinear convergence and avoid getting trapped in local oscillations.

[0061] This embodiment also includes a Convolutional Block Attention Module (CBAM), which is used to enhance the attention mechanism of feature representation in deep learning networks. By dynamically adjusting the channel and spatial weights of the feature map, the deep learning network pays more attention to the feature region, thereby improving network performance.

[0062] A convolutional block attention module is used to process the updated projection data, enhancing the capture of key features.

[0063] CBAM consists of channel attention and spatial attention. Channel attention processes the input feature map F through global average pooling and global max pooling to obtain one-dimensional feature vectors. These two feature vectors are then processed through a weight-sharing multilayer perceptron (MLP), their weights are summed, and finally, a sigmoid activation function is applied to obtain the channel attention vector M. c(F), represented as:

[0064]

[0065] Where σ represents the sigmoid function, AvgPool(·) represents the average pooling operation, and MaxPool(·) represents the max pooling operation;

[0066] The spatial attention component performs global average pooling and global max pooling on the input feature map F′ along the channel dimension, resulting in two feature maps of the same size as F′ but with one channel each. These two feature maps are then merged, convolved using a 7×7 kernel, and finally activated by a sigmoid function to obtain the spatial attention vector M. s (F), represented as:

[0067]

[0068] Among them, f 7×7 This represents a 7×7 volume core.

[0069] In this embodiment, please refer to Figure 2 , Figure 2 This is a schematic diagram of a chord graph domain denoising model provided in an embodiment of the present invention. The trained deep learning network also includes a chord graph domain denoising model, which includes a noise estimation network and a first multi-scale Transformer network. In the chord graph domain, the updated projection data is processed to obtain updated chord graph data, including:

[0070] A noise estimation network is used to process the updated projection data to obtain a noise estimation feature map. A first multi-scale Transformer network is then used to process the noise estimation feature map and the updated projection data to obtain updated chord graph data.

[0071] In this embodiment, the noise estimation network includes multiple convolutional blocks. Each convolutional block includes a first convolutional layer, a first normalization layer, a first activation function, a second convolutional layer, a second normalization layer, and a second activation function. Optionally, this embodiment uses four convolutional blocks: a first convolutional block, a second convolutional block, a third convolutional block, and a fourth convolutional block. The noise estimation network processes the updated projection data to obtain a noise estimation feature map, including:

[0072] The updated projection data is processed using convolutional blocks, and the processing result is added to the updated projection data to obtain the output of the convolutional block.

[0073] In this embodiment, the first multi-scale Transformer network includes a third convolutional layer, a first Transformer module, a first multi-scale downsampling fusion module, a second Transformer module, a second multi-scale downsampling fusion module, a third Transformer module, a first upsampling feature fusion module, a fourth Transformer module, a second upsampling feature fusion module, a fifth Transformer module, and a fourth convolutional layer. Optionally, the third and fourth convolutional layers are both 3×3 convolutional kernels. The first multi-scale Transformer network is used to process the noise-estimated feature map and the updated projection data to obtain updated chordmap data, including:

[0074] The noise estimation feature map and the updated projection data are processed by a third convolutional layer to obtain high-dimensional features;

[0075] The first Transformer module is used to process the high-dimensional features to obtain the first global features;

[0076] The first global feature is processed by the first multi-scale downsampling fusion module to obtain the first detailed feature;

[0077] The second Transformer module is used to process the first detailed features to obtain the second global features;

[0078] The second global feature is processed using a second multi-scale downsampling fusion module to obtain the second detailed feature;

[0079] The third Transformer module is used to process the second detailed features to obtain the third global features;

[0080] The first upsampling feature fusion module is used to process the third global feature and the second global feature to obtain the first concatenated feature;

[0081] The fourth Transformer module is used to process the first concatenated feature to obtain the fourth global feature;

[0082] The second upsampling feature fusion module is used to process the fourth global feature and the first global feature to obtain the second concatenated feature;

[0083] The fifth Transformer module is used to process the second concatenated feature to obtain the fifth global feature;

[0084] The fourth convolutional layer is used to process the fifth global feature to obtain the global feature;

[0085] The updated chord graph data is obtained by adding the global features, the noise-estimated feature map, and the updated projection data.

[0086] It is understandable that the first, second, and third Transformer modules are for the encoding stage, while the fourth and fifth Transformer modules are for the decoding stage. In the encoding stage, the first and second multi-scale downsampling fusion modules can enhance the correlation of global features and the response of weak component features. In the decoding stage, the first upsampling feature fusion module and the second upsampling feature fusion module can gradually restore the resolution, and the channel attention gating mechanism is used to dynamically weight and fuse the features in the encoding process with the features in the decoding process across scale levels, which is beneficial for noise elimination and the construction of global information.

[0087] It should be noted that the first, second, third, fourth, and fifth Transformer modules are all composed of two normalization layers, an attention module, and a feedforward network.

[0088] In this embodiment, please refer to Figure 3 and Figure 4 , Figure 3 This is a schematic diagram of a first multi-scale downsampling fusion module provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of a first upsampling feature fusion module provided in an embodiment of the present invention. The first multi-scale downsampling fusion module and the second multi-scale downsampling fusion module have the same structure, both including a max pooling layer, a fifth convolutional layer, a first concatenation layer, and a sixth convolutional layer. Optionally, the fifth convolutional layer has a 4×4 convolutional kernel, the sixth convolutional layer has a 1×1 convolutional kernel, and the max pooling layer has a 2×2 kernel. The first multi-scale downsampling fusion module processes the first global features to obtain the first detailed features, including:

[0089] The first global feature is processed using a max pooling layer to obtain edge features;

[0090] The first global feature is processed by the fifth convolutional layer to obtain the first feature;

[0091] The first stitching layer is used to stitch the edge features and the first feature together to obtain the third stitching feature, thereby expanding the perceptual field of view;

[0092] The third concatenated feature is processed using the sixth convolutional layer to obtain the first detail feature;

[0093] The first upsampling feature fusion module and the second upsampling feature fusion module have the same structure, both including multi-level transposed convolutional layers, a second concatenation layer, a global average pooling layer, a sigmoid function, and a seventh convolutional layer; the first upsampling feature fusion module processes the third global feature and the second global feature to obtain the first concatenation feature, including:

[0094] The third global feature is upsampled through a multi-level transposed convolutional layer to obtain the second feature.

[0095] The second concatenation layer is used to concatenate the second feature and the second global feature to obtain the fourth concatenated feature;

[0096] Spatial attention weights are generated using a global average pooling layer and a sigmoid function.

[0097] Multiply the spatial attention weights by the fourth concatenation feature to obtain the third feature;

[0098] The third feature is processed using the seventh convolutional layer to obtain the first concatenated feature.

[0099] It should be noted that the first multi-scale downsampling fusion module and the second multi-scale downsampling fusion module can extract weak component region features more precisely; the first upsampling feature fusion module and the second upsampling feature fusion module adopt progressive upsampling 1 fusion, and achieve cross-level feature fine calibration through variable upsampling chain and channel attention gating.

[0100] In this embodiment, please refer to Figure 5 , Figure 5 This is a schematic diagram of an image domain denoising model provided in an embodiment of the present invention. The trained deep learning network also includes an image domain denoising model, which includes a second multi-scale Transformer network and an image thinning module. In the image domain, the updated chordogram data is processed to obtain a reconstructed low-dose CT image, including:

[0101] The updated chord graph data is processed using a second multi-scale Transformer network to obtain the fourth feature; the second multi-scale Transformer network has the same structure as the first multi-scale Transformer network.

[0102] The fourth feature was processed using an image refinement module to obtain a reconstructed low-dose CT image.

[0103] Optionally, the image thinning module can be a U-net network with CNN as its backbone, which implements encoding and decoding operations through Conv Blocks.

[0104] In this embodiment, the training process of the trained multi-domain deep learning network includes:

[0105] S1. Obtain data of a preset category as samples in the training dataset, and obtain the true labels of the samples in the training dataset; wherein, the data of the preset category is low-dose CT images, and the true labels are conventional-dose CT images.

[0106] Specifically, using publicly available clinical scan patient data from the "2016 NIH-AAPM-Mayo Clinic Low-Dose CT Challenge," conventional-dose CT images were generated using simulation methods based on the CT scanning mechanism to produce corresponding low-dose CT images, including low-dose projection data, low-dose chordal data, and low-dose image data, such as... Figure 6 As shown, Figure 6 This is a flowchart of an embodiment of the present invention for generating low-dose CT images corresponding to conventional dose CT based on CT scanning mechanisms. Integer Hönsfield unit values ​​are obtained from conventional dose CT images and then dequantized. Based on the dequantized integer Hönsfield unit values, the attenuation distribution function corresponding to the human tissue slice is calculated. The attenuation distribution function is then converted into chordal data corresponding to the conventional dose CT image. Optionally, Radon transform is used to calculate the chordal data corresponding to the conventional dose CT image. Next, a Poisson-Gaussian noise model is used to process the chordal data corresponding to the conventional dose CT image, injecting composite Gaussian and Poisson noise into the undamaged projection to obtain the projection data corresponding to the low-dose CT image. A logarithmic transform is used to convert the projection data corresponding to the low-dose CT image into low-dose CT chordal data. Finally, a filtered back-projection algorithm is used to obtain the low-dose CT image data, thus reconstructing the low-dose CT image.

[0107] S2. Input a portion of the samples in the training dataset into the deep learning network to be trained for the jth time, and obtain the prediction results output during the jth training process.

[0108] S3. Calculate the loss based on the prediction results output during the j-th training process and the true labels of the samples used to train the deep learning network for the j-th training process, and use this loss as the loss for the j-th training process.

[0109] S4. Perform backpropagation based on the loss of the j-th training process to update the network parameters of the deep learning network to be trained in the j-th training, and obtain the deep learning network to be trained in the (j+1)-th training. At the same time, during the training process, calculate the fine-tuning loss based on the intermediate iteration results and fine-tune the deep learning network in training. Iterate in this way until the number of training times or the degree of convergence meets the preset conditions, and obtain the trained deep learning network.

[0110] In this embodiment, the expression for calculating the loss is:

[0111]

[0112] Among them, X n Y represents the image data updated in the last iteration. n X represents the chord graph image updated in the last iteration. * Y represents the image data corresponding to low-dose CT images. * This represents the chord graph image corresponding to a low-dose CT image, where α, β, and γ represent hyperparameters controlling the weights of different loss terms, and L... MAE (·) represents the logarithmic mean absolute error loss function, L MSE (·) represents the logarithmic mean square error loss function, L SSIM (·) represents the structural similarity loss function.

[0113] It should be noted that in this embodiment, logarithmic mean absolute error loss is assigned to the chord graph data to prevent the processed chord graph data from being overly smoothed; then, logarithmic mean square error loss and logarithmic mean absolute error loss are used to solve the distance between the reconstructed image and the reference image; finally, structural similarity loss is used to improve structural fidelity.

[0114] In this embodiment, the fine-tuning loss is calculated. avg The expression is:

[0115]

[0116] Where n represents the number of iterations, ω i This represents the loss weights for training the i-th layer, and the output is... i denoted as the reconstructed low-dose CT image output from the i-th layer, and target denoted as the normal-dose CT image.

[0117] It should be noted that the overall network architecture is obtained by cascading denoising networks. In each iteration, the output of the previous network is input into the next network. This provides the network with a mechanism for repeated top-down and bottom-up reasoning, which increases the overall structural complexity and training difficulty of the network. By incorporating the reconstructed images of the intermediate layers of the network into the loss calculation through intermediate supervision, the error of the output layer can be mitigated. The gradient vanishing phenomenon after multiple backpropagation ensures that the underlying parameters are updated normally.

[0118] It should be noted that the deep learning network in this embodiment is implemented based on the PyTorch framework, using the Adam optimizer, with a batch size of 5 and a learning rate using a cosine return strategy from 2×10⁻⁶. -4 Gradually decrease to 1×10 -6The total training iteration cycle of the network model was set to 100 epochs. The optimal model was selected based on the best PSNR value on the validation set. This model was then used to enhance the visualization of data containing weak components. The experimental environment was completed on an NVIDIA GeForce RTX 4090 GPU cluster.

[0119] In summary, the low-dose CT reconstruction method provided by this invention has the following beneficial effects:

[0120] I. In the process of reconstructing low-dose CT images, this invention takes into account the noise in the projection domain, chordal domain, and image domain, and can preserve the tissue structure of non-artifact areas to the greatest extent, thus achieving good reconstruction of low-dose CT images.

[0121] Second, this invention combines the Unet architecture with the Transformer module to perform image processing globally and locally, achieving better detail processing capabilities.

[0122] Third, this invention proposes a structure that combines a multi-scale downsampling fusion module with a Transformer module, which can effectively focus on weak component regions and noisy regions.

[0123] Fourth, this invention proposes a structure that combines a Transformer module with an upsampling feature fusion module, which can effectively enhance weak components while reducing noise.

[0124] Based on the same inventive concept, the present invention also provides a low-dose CT reconstruction device for implementing the low-dose CT reconstruction method provided in the above embodiments of the present invention. Examples of the method are described above and will not be repeated here. The device includes:

[0125] Image acquisition module, used to acquire low-dose CT images;

[0126] The image processing module is used to process low-dose CT images in the projection domain, chordal domain, and image domain using a trained deep learning network. In the projection domain, the low-dose CT image is processed to obtain updated projection data; in the chordal domain, the updated projection data is processed to obtain updated chordal data; and in the image domain, the updated chordal data is processed to obtain the reconstructed low-dose CT image.

[0127] The trained deep learning network is trained using data of a preset category as the training dataset. The initial deep learning network is trained, and intermediate supervision is performed during the training process to fine-tune the deep learning network during training.

[0128] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0129] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0130] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A low-dose CT reconstruction method, characterized in that, include: Acquire low-dose CT images; The low-dose CT image is processed in the projection domain, chordal domain, and image domain using a trained deep learning network. In the projection domain, the low-dose CT image is processed to obtain updated projection data. In the chordal domain, the updated projection data is processed to obtain updated chordal data. In the image domain, the updated chordal data is processed to obtain a reconstructed low-dose CT image. The trained deep learning network is trained using data of a preset category as the training dataset. The initial deep learning network is trained, and intermediate supervision is performed during the training process to fine-tune the deep learning network during training. The trained deep learning network includes a chord graph domain denoising model, which includes a noise estimation network and a first multi-scale Transformer network. The first multi-scale Transformer network includes a third convolutional layer, a first Transformer module, a first multi-scale downsampling fusion module, a second Transformer module, a second multi-scale downsampling fusion module, a third Transformer module, a first upsampling feature fusion module, a fourth Transformer module, a second upsampling feature fusion module, a fifth Transformer module, and a fourth convolutional layer. The noise estimation network is used to process the updated projection data to obtain a noise estimation feature map. The first multi-scale Transformer network is used to process the noise estimation feature map and the updated projection data to obtain updated chord graph data, including: The noise estimation feature map and the updated projection data are processed by the third convolutional layer to obtain high-dimensional features; The high-dimensional features are processed using the first Transformer module to obtain the first global features; The first global feature is processed using the first multi-scale downsampling fusion module to obtain the first detailed feature; The second Transformer module is used to process the first detailed features to obtain the second global features; The second global feature is processed using the second multi-scale downsampling fusion module to obtain the second detailed feature; The third Transformer module is used to process the second detailed feature to obtain the third global feature; The first upsampling feature fusion module is used to process the third global feature and the second global feature to obtain the first concatenated feature; The fourth Transformer module is used to process the first concatenated feature to obtain the fourth global feature; The second upsampling feature fusion module is used to process the fourth global feature and the first global feature to obtain the second concatenated feature; The fifth Transformer module is used to process the second concatenated feature to obtain the fifth global feature; The fourth convolutional layer is used to process the fifth global feature to obtain the global feature; The global features, the noise estimation feature map, and the updated projection data are added together to obtain the updated chord graph data.

2. The low-dose CT reconstruction method according to claim 1, characterized in that, The trained deep learning network includes a projection domain denoising model; the process of processing the low-dose CT image in the projection domain to obtain updated projection data includes: Construct a projection data generation model, represented as follows: ； in, This represents the number of quantum numbers received by the detector. , Indicates electronic background noise. ; The electronic background noise follows a non-stationary Gaussian distribution with zero mean, expressed as: ； in, Represents the variance of the noise; Since the electronic background noise follows a Gaussian distribution, the first conditional distribution can be obtained, expressed as: ； Quantum noise Following a Poisson distribution, it can be represented as: ； in, This represents the intensity of radiation under ideal conditions, according to Beer-Lambert's law. satisfy: ； in, Represents ideal chord diagram data, Indicates the actual intensity of the radiation; According to quantum noise Following the Poisson distribution and combined with the Beer-Lambert law, we obtain the second conditional distribution, which is expressed as: ； in, Index representing a quantum; Based on the first conditional distribution and the second conditional distribution, the projection domain denoising model is obtained, expressed as: ； Based on the maximum a posteriori estimation theory, the posterior distribution of the projection domain denoising model is obtained, expressed as: ； The posterior distribution of the projection domain denoising model is simplified to obtain the updated projection domain denoising model, which is expressed as: ； The projection domain denoising model is optimized and updated using the block coordinate descent method to obtain quantum noise. The objective function is expressed as: ； In the first stage, the memory-limited BFGS algorithm is used to address the quantum noise. The objective function is solved, and in the second stage, the damped Newton method is used to address the quantum noise. The objective function is solved to obtain the optimal quantum noise, which is the optimal quantum number received by the detector. ; Based on the quantum number received by the optimal detector This yields updated projection data.

3. The low-dose CT reconstruction method according to claim 1, characterized in that, The noise estimation network includes multiple convolutional blocks, each convolutional block comprising a first convolutional layer, a first normalization layer, a first activation function, a second convolutional layer, a second normalization layer, and a second activation function. The step of processing the updated projection data using the noise estimation network to obtain a noise estimation feature map includes: The updated projection data is processed using the convolutional block, and the processing result is added to the updated projection data to obtain the output result of the convolutional block.

4. The low-dose CT reconstruction method according to claim 1, characterized in that, The first multi-scale downsampling fusion module and the second multi-scale downsampling fusion module have the same structure, both including a max pooling layer, a fifth convolutional layer, a first concatenation layer, and a sixth convolutional layer; the first multi-scale downsampling fusion module is used to process the first global features to obtain the first detailed features, including: The first global feature is processed using the max pooling layer to obtain edge features; The first global feature is processed by the fifth convolutional layer to obtain the first feature; The edge feature and the first feature are spliced ​​together using the first splicing layer to obtain the third spliced ​​feature; The third splicing feature is processed using the sixth convolutional layer to obtain the first detail feature; The first upsampling feature fusion module and the second upsampling feature fusion module have the same structure, both including multi-level transposed convolutional layers, a second concatenation layer, a global average pooling layer, a sigmoid function, and a seventh convolutional layer; the first upsampling feature fusion module is used to process the third global feature and the second global feature to obtain the first concatenation feature, including: The third global feature is upsampled using the multi-level transposed convolutional layer to obtain the second feature. The second concatenation layer is used to concatenate the second feature and the second global feature to obtain the fourth concatenated feature; Spatial attention weights are generated using the global average pooling layer and the sigmoid function; Multiply the spatial attention weight by the fourth concatenation feature to obtain the third feature; The third feature is processed using the seventh convolutional layer to obtain the first concatenated feature.

5. The low-dose CT reconstruction method according to claim 1, characterized in that, The trained deep learning network also includes an image domain denoising model, which includes a second multi-scale Transformer network and an image thinning module. In the image domain, the updated chordogram data is processed to obtain a reconstructed low-dose CT image, including: The updated chord graph data is processed using the second multi-scale Transformer network to obtain the fourth feature; wherein the second multi-scale Transformer network has the same structure as the first multi-scale Transformer network. The fourth feature is processed using the image refinement module to obtain a reconstructed low-dose CT image.

6. The low-dose CT reconstruction method according to claim 1, characterized in that, The training process of the trained deep learning network includes: Data of a preset category is obtained as samples in the training dataset, and the true labels of the samples in the training dataset are obtained; wherein, the data of the preset category is low-dose CT images, and the true labels are conventional-dose CT images; Input a portion of the samples from the training dataset into the first... The deep learning network to be trained is then trained to obtain the first... The prediction results output during this training process; According to the The prediction results output during the training process are the same as those during the training phase. The true labels of the samples for the next deep learning network to be trained are used to calculate the loss and serve as the first... Loss during each training session; According to the The loss from the training process is backpropagated to update the loss from the training process. The network parameters of the deep learning network to be trained are obtained. The deep learning network to be trained is then trained. Simultaneously, during the training process, the fine-tuning loss is calculated based on the intermediate iteration results, and the deep learning network under training is fine-tuned. This process is repeated until the number of training iterations or the degree of convergence meets the preset conditions, at which point the trained deep learning network is obtained.

7. The low-dose CT reconstruction method according to claim 6, characterized in that, The calculated loss The expression is: ； in, This represents the image data updated in the last iteration. This represents the chord graph image updated in the last iteration. This represents the image data corresponding to low-dose CT images. This represents the chordal image corresponding to a low-dose CT image. , and These represent the hyperparameters that control the weights of different loss terms. This represents the logarithmic mean absolute error loss function. This represents the log-mean squared error loss function. This represents the structural similarity loss function.

8. A low-dose CT reconstruction device, characterized in that, include: Image acquisition module, used to acquire low-dose CT images; The image processing module is used to process the low-dose CT image in the projection domain, chordal domain, and image domain using a trained deep learning network. In the projection domain, the low-dose CT image is processed to obtain updated projection data; in the chordal domain, the updated projection data is processed to obtain updated chordal data; and in the image domain, the updated chordal data is processed to obtain a reconstructed low-dose CT image. The trained deep learning network is trained using data of a preset category as the training dataset. The initial deep learning network is trained, and intermediate supervision is performed during the training process to fine-tune the deep learning network during training. The trained deep learning network includes a chord graph domain denoising model, which includes a noise estimation network and a first multi-scale Transformer network. The first multi-scale Transformer network includes a third convolutional layer, a first Transformer module, a first multi-scale downsampling fusion module, a second Transformer module, a second multi-scale downsampling fusion module, a third Transformer module, a first upsampling feature fusion module, a fourth Transformer module, a second upsampling feature fusion module, a fifth Transformer module, and a fourth convolutional layer. The noise estimation network is used to process the updated projection data to obtain a noise estimation feature map. The first multi-scale Transformer network is used to process the noise estimation feature map and the updated projection data to obtain updated chord graph data, including: The noise estimation feature map and the updated projection data are processed by the third convolutional layer to obtain high-dimensional features; The high-dimensional features are processed using the first Transformer module to obtain the first global features; The first global feature is processed using the first multi-scale downsampling fusion module to obtain the first detailed feature; The second Transformer module is used to process the first detailed features to obtain the second global features; The second global feature is processed using the second multi-scale downsampling fusion module to obtain the second detailed feature; The third Transformer module is used to process the second detailed feature to obtain the third global feature; The first upsampling feature fusion module is used to process the third global feature and the second global feature to obtain the first concatenated feature; The fourth Transformer module is used to process the first concatenated feature to obtain the fourth global feature; The second upsampling feature fusion module is used to process the fourth global feature and the first global feature to obtain the second concatenated feature; The fifth Transformer module is used to process the second concatenated feature to obtain the fifth global feature; The fourth convolutional layer is used to process the fifth global feature to obtain the global feature; The global features, the noise estimation feature map, and the updated projection data are added together to obtain the updated chord graph data.

Images