Low-dose ct imaging method and apparatus based on semantic prior knowledge guidance

Abstract

The application discloses a low-dose CT imaging method and device based on semantic prior knowledge guidance. In view of the problems of image over-smoothing, detail loss and the like in low-dose CT imaging, the application utilizes the significant difference of different anatomical structures (such as bones, soft tissues, gas regions and the like) in noise level, extracts semantic prior information through a pre-trained visual model, so as to effectively distinguish different tissue types and capture the noise mode of specific regions. Meanwhile, a semantic perception module is proposed, the detail features focused by the noise reduction network are aligned and fused with the high-level anatomical semantic features captured by the semantic network, so that the anatomical structure is more accurately perceived in the noise reduction process. Further, a segmentation network is introduced after the noise reduction network, and a multi-task learning strategy is adopted, so as to realize the joint optimization of noise reduction and semantic segmentation. This design not only improves the noise reduction performance of the image and effectively retains the key anatomical details, but also provides clearer image support for clinical diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the cross field of medical image processing and computer vision, in particular to a low-dose CT imaging method and device based on semantic prior knowledge guidance. BACKGROUND

[0002] CT imaging plays an irreplaceable role in medical diagnosis, which can provide accurate anatomical and pathological information for tumor screening, cardiovascular disease and lung disease, etc. However, the radiation dose produced by CT scanning has long been a concern for patient safety, especially in the case of multiple scans, the cumulative effect of radiation may increase the risk of cancer, etc. In order to reduce the harm of radiation, low-dose CT imaging is gradually adopted in the clinic, for example, reducing the tube current and tube voltage, but this method inevitably increases the image noise, which affects the accuracy of diagnosis. Although the existing denoising methods can improve the image quality to a certain extent, they often cannot balance noise removal and image detail preservation, and are strongly dependent on manual regularization and hyperparameters. The direct mapping relationship between normal dose CT (NDCT) and low dose CT (LDCT) does not follow a clear rule, for example, the strip artifact caused by sparse sampling, and the mixed Gaussian noise and Poisson noise caused by low tube current.

[0003] In recent years, low-dose CT image post-processing algorithms based on deep learning have been widely studied and have made significant progress. The earliest studies mostly optimize the encoder-decoder network by minimizing the pixel-level loss between the denoised image and the normal dose CT image, and typical models such as residual encoder-decoder convolutional neural network (RED-CNN). Although these methods perform outstandingly in denoising performance, due to the use of pixel-level loss functions (such as L1 or L2 loss), it is easy to produce over-smoothed images, resulting in the loss of some details. In order to solve this problem, researchers introduced generative adversarial networks (GAN) and structural similarity loss and perceptual loss to generate more realistic denoised images, and typical methods such as WGAN-VGG model. However, GAN, due to its adversarial training characteristics, usually requires careful design of the optimization process and network architecture to ensure the convergence and stability of the model. In addition, some model-based methods try to combine iterative reconstruction methods and neural networks to improve the interpretation of deep learning models.

[0004] Although deep learning-based methods greatly improve the image quality of low-dose CT, most of these methods ignore the differences between different anatomical regions, resulting in blurred details in regions with low noise levels. In addition, existing studies have shown that the noise level in CT images varies significantly depending on the type of tissue. Therefore, in recent years, some studies have begun to explore the use of anatomical semantics to guide CT image denoising. For example, by utilizing the absorption of X-rays in different tissues to generate anatomical feature domains (such as subcutaneous fat, muscle, and visceral fat), and combining semantic features and image features in a semantic fusion module, while introducing a structural semantic loss function to measure segmentation differences. However, these methods obtain semantic features that are too rough, resulting in suboptimal denoising results. SUMMARY

[0005] The present application aims to address the shortcomings of the prior art by providing a low-dose CT imaging method and device based on prior knowledge guidance. The method aims to introduce precise semantic priors through a visual large model to distinguish different anatomical structures, allowing the denoising network to implicitly distinguish noise levels in different regions. At the same time, a semantic perception module is proposed to align and fuse the detailed features focused on by the denoising network with the high-level anatomical semantic features captured by the semantic network, ensuring more accurate perception of anatomical structures during the denoising process. Finally, a segmentation network is introduced after the denoising network, using a multi-task learning strategy to achieve joint optimization of denoising and semantic segmentation, resulting in better denoising results.

[0006] The low-dose CT imaging method based on semantic prior knowledge guidance designed by the present application includes the following steps:

[0007] Obtaining semantic features of low-dose CT images using a large model as prior knowledge of anatomical structures;

[0008] Designing a joint network structure including a denoising network and a semantic segmentation network, the denoising network is used to reduce noise in CT images, and the semantic segmentation network uses extracted semantic features to identify anatomical structures;

[0009] Using mean square error as the loss in the denoising process, and using Dice coefficient and cross-entropy as the loss in the segmentation process;

[0010] Using the above losses, the denoising network and the segmentation network are trained using an alternating optimization method;

[0011] Using the trained joint network structure for low-dose CT imaging.

[0012] Further, the visual branch in the pre-trained CLIP-Driven general model is used to extract semantic features.

[0013] Further, the training data set adopts two public low-dose CT data sets released by "NIHAAPM-Mayo Clinic Low-Dose CT Grand Challenge" in 2016 and 2020. The data set can also be constructed according to the actual situation for network training.

[0014] Further, the denoising network sequentially comprises a shallow feature extraction layer, a deep feature extraction layer and an output layer; the shallow feature extraction layer and the output layer are both composed of convolution; the deep feature extraction layer is composed of three semantic perception modules, and the semantic perception module is composed of cross attention.

[0015] Preferably, the denoising process is specifically as follows:

[0016] The shallow feature extraction is expressed as:

[0017] f l =conv(X)

[0018] Wherein, conv represents a convolution layer, X represents a low-dose CT image,

[0019] The semantic feature is first taken as a query key of the cross attention module through the self-attention module, and the formula is expressed as:

[0020] Q=LN(SA(LN(GN(S))))

[0021] Wherein, LN, GN, SA and S represent layer normalization, group normalization, a self-attention module and a semantic feature, then the low-dose CT feature is taken as K and V values through convolution, and finally the semantic feature transformed through the cross attention and the original feature f are spliced as the final feature:

[0022]

[0023] Wherein, f s and f respectively represent the image feature after fusing the semantic feature and the output of the last layer, d k and GELU respectively represent a scale factor and a Gaussian error linear unit, and concat represents a connection operation.

[0024] The semantic feature can be expressed as:

[0025] S=conv(GELU(conv(Y s )))

[0026] Wherein, Y s represents a semantic label extracted by a pre-trained visual model from the NDCT,

[0027] The output layer can be expressed as:

[0028]

[0029] wherein and f d respectively represent the output of the low-dose CT with noise reduction and the deep network feature extraction layer, and the learning process of the network is accelerated by using a skip connection.

[0030] Further, the segmentation network comprises an encoder-decoder structure, the encoder and the decoder both adopt a residual module, and the encoder and the decoder use a skip connection to retain low-level features of the image.

[0031] Further, in each training cycle, the alternately optimized training method first uses the output of the pre-trained large model as prior knowledge, which is fused into the denoising network through a semantic perception module to realize guided optimization of the parameters of the denoising network; then, the denoised image is used as input, and the semantic prior knowledge is used as a prediction target to optimize the parameters of the segmentation network.

[0032] Further, after the joint network structure is trained, an independent validation data set is used for performance evaluation, and image quality evaluation indicators are used for quantitative evaluation of the denoising performance, including peak signal-to-noise ratio, structural similarity index and root mean square error, and the best network parameters are selected for testing.

[0033] Further, the alternately optimized training process is as follows:

[0034] First, a pair of low-dose CT image and normal-dose CT image is selected from the data set, then the normal-dose CT image is output to the pre-trained visual model to obtain a semantic label, the semantic label and the low-dose CT image are output to the denoising network to obtain a denoised low-dose CT image, and the loss function is calculated according to the following formula

[0035] wherein Y represents a standard-dose CT image, denotes a denoised low-dose CT image; then the parameters of the denoising network are optimized by back propagation, then the denoised normal-dose CT image is output to the segmentation network to obtain a segmented semantic label, and the loss function is calculated according to the following formula

[0036] L s =L Dice +αL CE

[0037] wherein p CE is a cross-entropy objective function, L Dice denotes a Dice coefficient, and a denotes a weight factor for balancing different loss functions; then the parameters of the denoising network and the segmentation network are optimized by back propagation; the cycle is repeated until the loss function converges.

[0038] Based on the same inventive concept, the present application also discloses an electronic device comprising:

[0039] one or more processors;

[0040] a storage device for storing one or more programs;

[0041] When the one or more programs are executed by the one or more processors, the one or more processors implement the low-dose CT imaging method guided by semantic prior knowledge.

[0042] Based on the same inventive concept, the present application also discloses a computer readable medium having a computer program stored thereon, the program being executed by a processor to implement the low-dose CT imaging method guided by semantic prior knowledge.

[0043] The present application has the following advantages:

[0044] 1. Utilize the significant differences in noise levels among different anatomical structures (such as bones, soft tissues, gas regions, etc.) to extract semantic prior information through pre-trained visual models. These features can effectively distinguish different tissue types and capture noise patterns specific to different regions.

[0045] 2. Propose a semantic perception module that aims to align the denoising features with the semantic features. In the denoising process, the denoising network mainly focuses on the detailed information of the image to remove noise and preserve fine structures; while the semantic network focuses on capturing high-level anatomical semantic information. This module effectively fuses these two features to organically align the detailed features with the semantic features, ensuring that the denoising network can better perceive anatomical structures while denoising.

[0046] 3. Introduce a segmentation network after the denoising network, using a multi-task learning strategy to achieve joint optimization of denoising and semantic segmentation. This design not only improves the denoising network's perception of noise levels in different anatomical structures, but also better preserves key anatomical details, providing clearer and more reliable image support for clinical diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a schematic diagram of the overall network structure of an embodiment of the present application.

[0048] Figure 2 is a schematic diagram of the denoising network of an embodiment of the present application.

[0049] Figure 3 is a schematic diagram of the segmentation network of an embodiment of the present application.

[0050] Figure 4 is a schematic diagram of the semantic perception module of an embodiment of the present application.

[0051] Figure 5 is a general flowchart of embodiments of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0053] Embodiment one

[0054] The present embodiment mainly introduces semantic prior information through a semantic segmentation model based on the difference characteristics of different anatomical structure noise levels, and proposes a CT imaging method and device based on semantic prior guidance. The differences of different anatomical structures are fully considered, and the segmentation network is used to supervise the denoising process. The results obtained by the present application are more scientific and more accurate.

[0055] Step 1: Data preparation: Collect a large number of low-dose CT (LDCT) images and their corresponding normal-dose CT (NDCT) images to form a training data set. The implementation process is as follows:

[0056] Two public low-dose CT data sets released in 2016 and 2020 by "NIHAAPM-Mayo Clinic Low-Dose CT Grand Challenge" are used. The Mayo 2016 data set includes normal-dose abdominal CT images and corresponding simulated quarter-dose CT images of 10 anonymous patients. In this embodiment, the images of 8 patients are selected as the training set, the images of 1 patient are selected as the validation set, and the images of 1 patient are selected as the test set. The Mayo 2020 data set provides abdominal CT image data of 100 patients, and this embodiment randomly selects 20 patients for experiments, of which 16 patients' images are selected as the training set, 2 patients' images are selected as the validation set, and 2 patients' images are selected as the test set. For subsequent processing, the data is preprocessed by truncating the original intensity value in the range of 2% to 98% of the initial Hounsfield unit (HU) and normalizing each original CT case to have a mean of zero and a variance of one. Those skilled in the art can construct the data set according to the actual situation. Generally speaking, the more data samples, the more diseases and anatomical structures covered, and the stronger the generalization ability of the model.

[0057] Step 2: Semantic extraction: forward propagate the preprocessed NDCT images using a pre-trained visual large model to extract the semantics of the anatomical structure. The implementation process is described as follows:

[0058] The visual branch in the pre-trained CLIP-Driven general model is used as the pre-trained visual model, which is hereinafter abbreviated as "CLIP Vision". Specifically, "CLIP-Vision" is composed of a visual encoder and a text-driven segmenter. The visual encoder adopts Swin UNETR. Swin UNETR is a "U-shaped" network, and the encoder part adopts SwinTransformer, directly using 3D blocks, connected to a CNN-based decoder through skip connections of different resolutions. The preprocessed NDCT image is first passed through the four stages of the encoder, each stage containing two transformer blocks, and the resolution of the features is reduced by half between each stage using a block merging layer. Then pass through the four stages of the decoder, input the representation extracted by each stage into the residual block, and then upsample the processing features of each stage through the deconvolution layer, and connect them with the features output by the residual block of the previous stage. Finally, the semantic label of the NDCT is obtained through the text classifier. The text-driven segmenter is composed of three consecutive convolutional layers, and finally outputs the probability of each class foreground and background. The pre-trained visual model is represented as g, and the semantic features extracted by inputting the NDCT image Y into g are represented as Y s . The formula is represented as:

[0059] Y s = g(Y)

[0060] A person skilled in the art can select a suitable pre-trained visual model according to the actual effect.

[0061] Step 3: Multi-task learning model construction: design a joint network structure, including a denoising network and a segmentation network. The denoising network aims to reduce the noise in the image, while the semantic segmentation network measures the semantic distance between the predicted normal dose CT and the standard dose CT. Inspired by the pre-trained visual model, this scheme takes the standard dose CT as the input of the pre-trained visual model, and its segmentation result as the semantic label of the standard dose CT. The semantic segmentation network adopts an encoder-decoder structure, and both the encoder and the decoder part adopt residual modules and use skip connections to retain the low-level features of the image. Down-sampling is realized by convolution with a stride of 2, and up-sampling is realized by deconvolution operation. The denoising network also adopts an encoder-decoder structure, unlike the encoder of the semantic segmentation network, the encoder of the denoising network adopts a semantic perception module to fuse the extracted semantic prior and image features to distinguish different regions. In addition, a semantic loss function is introduced to measure the semantic distance between the standard dose CT and the predicted standard dose CT. This scheme can adaptively adjust the denoising strategy for different anatomical structures, better preserve the local details of the image, especially in low-noise areas, ensure that important anatomical structure information is not lost, and thus improve the diagnostic value of the CT image.

[0062] The denoising network includes a shallow feature extraction layer, a deep feature extraction layer, and an output layer. The shallow feature extraction layer and the output layer are both composed of a 3x3 convolution, which is used to expand or squeeze the channel to adapt to the size of the input image. The deep feature extraction layer is composed of 3 semantic perception modules, which are composed of cross-attention, allowing the denoising network to focus on the alignment of semantics.

[0063] The shallow feature extraction can be represented as

[0064] f l = conv(X)

[0065] where conv represents a convolution layer, and X represents an LDCT image.

[0066] The semantic feature is first passed through a self-attention module as the query key of the cross-attention module, which is represented by the formula:

[0067] Q = LN(SA(LN(GN(S))))

[0068] where LN, GN, SA, and S represent layer normalization, group normalization, a self-attention module, and a semantic feature. Then the LDCT feature is passed through a convolution as the K and V values. Finally, the semantic feature transformed by the cross-attention is concatenated with the original feature f as the final feature

[0069]

[0070] where f s and f represent the image feature after fusing the semantic feature and the output of the previous layer, respectively. d k and GELU represent a scale factor and a Gaussian error linear unit, respectively. concat represents a concatenation operation.

[0071] The semantic feature can be represented as

[0072] S = conv(GELU(conv(Y s )))

[0073] where Y s represents a semantic label extracted by a pre-trained visual model from an NDCT.

[0074] The output layer can be represented as

[0075]

[0076] where and f d represent the output of the denoised LDCT and the deep network feature extraction layer, respectively. A skip connection is used to speed up the learning process of the network.

[0077] The segmentation network structure is mainly composed of an encoder, a bottleneck layer, and a decoder. The encoder is composed of multiple convolutional layers and pooling layers. Each convolutional layer usually contains two 3x3 convolution operations and a ReLU activation function, which is used to extract image features and maintain spatial structure. The pooling layer reduces the spatial resolution of the feature map through a 2x2 max pooling operation, increases the receptive field, and captures more context information in a larger range. Between the encoder and the decoder, the bottleneck layer further extracts deep features through a convolution block. The up-sampling layer in the decoder uses deconvolution or bilinear interpolation to increase the spatial resolution of the feature map, and through a jump connection, the corresponding encoder feature map is spliced with the decoder feature map to retain detailed information and enhance segmentation accuracy. Finally, the output of the decoder is passed through a 1x1 convolutional layer to map the feature map to the required number of output channels, generating the final segmentation map. The formula is as follows

[0078]

[0079] where Seg represents the segmentation network.

[0080] The number of features in the set convolutional layer is 64. The number of attention features in the semantic perception module is 64. Those skilled in the art can select these parameters according to the actual effect.

[0081] Step 4: Loss function design: design a comprehensive loss function, including denoising loss and segmentation loss, to ensure that the image details and anatomical structure information are preserved as much as possible during the denoising process. The implementation process is described as follows:

[0082] The objective function of the denoising network uses mean square error, which is expressed as:

[0083]

[0084] The objective function of the segmentation network uses Dice coefficient and cross-entropy. For multi-class segmentation, the Dice coefficient formula is:

[0085]

[0086] The cross-entropy objective function using focal loss can be expressed as:

[0087]

[0088] The total objective function of the segmentation model is:

[0089] L s =L Dice +αL CE

[0090] where C denotes the number of classes, γ is a tunable parameter, and a represents a weight factor balancing different loss functions. In the experiment, C = 32, γ = 2, and a = 0.5 are set.

[0091] These parameters can be selected by a person skilled in the art according to actual effects.

[0092] Step 5: Alternating optimization strategy: The denoising network and the segmentation network are trained using the alternating optimization method. In each training cycle, first, the output of the pre-trained visual model is used as prior information, which is fused into the denoising network through the semantic perception module to guide the optimization of the parameters of the denoising network; then, the denoised image is used as input, and the semantic prior knowledge is used as the prediction target to optimize the parameters of the segmentation network. Through this method, the two networks can promote each other, thereby improving the overall performance. The implementation process is as follows:

[0093] First, pairs of LDCT and NDCT data are selected from the data set, and then the NDCT is output to the pre-trained visual model to obtain semantic labels. The semantic labels and the LDCT are output to the denoising network to obtain the denoised LDCT. The loss function is calculated according to the total loss of the segmentation model, and the parameters of the denoising network are optimized through back propagation. Then, the denoised LDCT is output to the segmentation network to obtain the segmented semantic labels. The loss function is calculated according to the total loss of the segmentation model, and the parameters of the denoising network and the segmentation network are optimized through back propagation. This cycle is repeated until the loss function converges. The training hyperparameters are as follows: In this embodiment, the AdamW optimizer with a batch size of 4 is used to train the network, and β1, β2, and ε are set to 0.9, 0.99, and 1e-8, respectively. The input image size is set to 512x512, and the network is trained for 100 epochs. The initial learning rate of the denoising network and the segmentation network is 1e-4, and the cosine annealing strategy is used to gradually decay to 1e-6. This method is implemented based on PyTorch, and the training and testing are completed on an NVIDIA Tesla V100 GPU. Since the segmentation network only assists the denoising network in learning during training, the denoising network does not introduce additional computational cost during testing.

[0094] Step 6: Model evaluation: After training, the model performance is evaluated using an independent validation set, including the denoising effect. By comparing with traditional denoising methods, the advantages of the present application are verified. The implementation process is as follows:

[0095] After training the model, the best model parameters are selected according to the evaluation index for testing on the test set to observe the performance of the present application.

[0096] The application adopts three commonly used objective image quality evaluation indexes to quantitatively evaluate the denoising performance: peak signal-to-noise ratio (PSNR), structural similarity index and root mean square error (RMSE). Among them, the higher the PSNR and SSIM, the lower the RMSE, indicating better performance. The definitions of PSNR, SSIM and RMSE are as follows

[0097]

[0098] Embodiment two

[0099] Based on the same inventive concept, the application further provides an electronic device comprising one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in embodiment one.

[0100] Since the device introduced in embodiment two of the application is the electronic device used to implement the low-dose CT imaging method based on semantic prior knowledge guidance in embodiment one of the application, the specific structure and modifications of the electronic device can be understood by those skilled in the art based on the method introduced in embodiment one of the application, and therefore will not be described here again. Any electronic device used in the method of embodiment one of the application belongs to the scope of the application.

[0101] Embodiment three

[0102] Based on the same inventive concept, the application further provides a computer readable medium having a computer program stored thereon, wherein the program is executed by a processor to implement the method described in embodiment one.

[0103] Since the device introduced in embodiment three of the application is the computer readable medium used to implement the low-dose CT imaging method based on semantic prior knowledge guidance in embodiment one of the application, the specific structure and modifications of the electronic device can be understood by those skilled in the art based on the method introduced in embodiment one of the application, and therefore will not be described here again. Any electronic device used in the method of embodiment one of the application belongs to the scope of the application.

[0104] The specific embodiments described herein are merely illustrative of the spirit of the application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or replace them with similar ways without departing from the spirit of the application or exceeding the scope defined by the appended claims.

Claims

1. A method for low-dose CT imaging based on semantic prior knowledge guidance, characterized in that, The method comprises the following steps: obtaining semantic features of the low-dose CT image by using a large model as prior knowledge of anatomical structures; designing a joint network structure, including a denoising network and a semantic segmentation network, the denoising network being used to reduce noise in the CT image, and the semantic segmentation network being used to identify anatomical structures by using the extracted semantic features; the denoising network sequentially comprising a shallow feature extraction layer, a deep feature extraction layer, and an output layer; the denoising process using mean square error as a loss, and the segmentation process using Dice coefficient and cross-entropy as a loss; wherein the denoising process is specifically as follows: the shallow feature extraction is expressed as: wherein conv represents a convolution layer, and X represents a low-dose CT image, the semantic feature is first passed through a self-attention module as a query key of a cross-attention module, and the formula is expressed as: where LN, GN, SA, represent layer normalization, group normalization, self-attention module and semantic feature, and then the low-dose CT feature is convolved as and value, and finally the semantic feature transformed by cross-attention and the original feature spliced as the final feature: wherein and respectively represent the image feature after fusing the semantic feature and the output of the previous layer, and respectively represent the scale factor and the Gaussian error linear unit, and concat represents the connection operation. the semantic feature is expressed as: wherein denotes semantic labels extracted by the pre-trained visual model from the NDCT, the output layer is expressed as: wherein and respectively represent the output of the low-dose CT and the deep network feature extraction layer of the noise reduction, and the skip connection is adopted to accelerate the learning process of the network. the denoising network and the segmentation network are trained by using the above loss and an alternating optimization method; the trained joint network structure is used for low-dose CT imaging. 2.The low-dose CT imaging method based on semantic prior knowledge guidance according to claim 1, characterized in that: The semantic feature is extracted by using a visual branch in a pre-trained CLIP-Driven general model. 3.The low-dose CT imaging method based on semantic prior knowledge guidance according to claim 1, characterized in that: The shallow feature extraction layer and the output layer are both composed of convolution; the deep feature extraction layer is composed of three semantic perception modules, and the semantic perception module is composed of cross-attention. 4.The low-dose CT imaging method based on semantic prior knowledge guidance according to claim 1, characterized in that: The segmentation network comprises an encoder-decoder structure, the encoder and the decoder both adopt a residual module, and the encoder and the decoder use a skip connection. 5.The low-dose CT imaging method based on semantic prior knowledge guidance according to claim 1, characterized in that: In each training cycle, the alternating optimization training method first uses a pre-trained large model output as prior knowledge, which is fused into the denoising network through a semantic perception module to realize guiding optimization of parameters of the denoising network; then, the denoised image is used as input, and the semantic prior knowledge is used as a prediction target to optimize parameters of the segmentation network.

6. The method of claim 5, wherein: The alternating optimization training process is specifically as follows: first, paired low-dose CT images and normal-dose CT images are selected from a data set, then the normal-dose CT images are output to a pre-trained visual model to obtain semantic labels, the semantic labels and the low-dose CT images are output to the denoising network to obtain denoised low-dose CT images, and a loss function is calculated according to the following formula wherein represents a standard dose CT image; represents a low dose CT image after denoising; then the parameters of the denoising network are optimized by back propagation, and then the normal dose CT image after denoising is output to the segmentation network to obtain the segmented semantic label, and the loss function is calculated according to the following formula: wherein a cross-entropy objective function, denotes the Dice coefficient, denotes a weight factor balancing the different loss functions; then, back propagation is performed to optimize parameters of the denoising network and the segmentation network; the cycle is repeated until the loss function converges. 7.The low-dose CT imaging method based on semantic prior knowledge guidance according to claim 1, characterized in that: After the joint network structure is trained, an independent verification data set is used for performance evaluation, and image quality evaluation indexes are used for quantitative evaluation of the denoising performance, including peak signal-to-noise ratio, structural similarity index, and root mean square error.

8. An electronic device, comprising: comprise: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method as claimed in any one of claims 1-7.

9. A computer readable medium having stored thereon a computer program, characterized in that: The program is executed by the processor to implement the method as claimed in any one of claims 1-7.

Images