Texture feature guided texture preserving low dose ct image denoising

Abstract

The application discloses a texture feature guided texture preserving low dose CT image denoising method, and belongs to the field of medical image processing.The application specifically discloses a multi-scale deep residual attention network model with texture feature guidance, which is applied to low dose CT imaging.The main network model comprises four sub-models, one is a multi-scale initial denoising network model for denoising low dose CT, and the other network is used for extracting texture details after the initial denoising network, and the two network parts work cooperatively.The extracted texture details and the initial low dose CT are fused through a multi-scale image and texture feature fusion network model, and then enter a multi-scale main denoising network for further denoising of the low dose CT, which is beneficial to the main denoising network to learn more unobvious details.The low dose CT image denoising method disclosed by the application efficiently removes the noise and stripe artifacts in the low dose CT image, and meanwhile, the structural information and texture feature detail information in the image are preserved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image processing, and more specifically relates to a texture preserving low-dose CT image denoising based on texture feature guidance. BACKGROUND

[0002] X-ray computed tomography (CT) has made great progress in basic technology and clinical medical applications, and has become an indispensable imaging technology in medical diagnosis and treatment. However, during the CT scanning process, high-dose ionizing radiation will be generated, thereby causing certain health risks to patients. In view of this health problem, people have put forward the basic concept of low-dose CT. Reducing the radiation dose can reduce the flux of X-rays by reducing the working voltage and working current of the X-ray tube and shortening the exposure time. However, with the reduction of the radiation dose, noise and stripe artifacts will be generated in the reconstructed image, thereby causing the decline of the image quality and the adverse effects on clinical diagnosis. We guarantee the quality of the reconstructed image while effectively reducing the harm to the human body caused by the radiation dose, which is of great significance to clinical diagnosis, so that the low-dose CT image has become one of the hot issues in the medical field. In order to improve the denoising ability of the low-dose CT image and improve the quality of the low-dose CT image, many algorithms have been designed to solve the above problems. These algorithms can be divided into three categories (a) chord diagram filtering technology, (b) iterative reconstruction algorithm and (c) image post-processing.

[0003] The pre-reconstruction filtering techniques refer to the processing of the raw data pairs before image reconstruction, such as filtered back-projection (FBP). The optimization object of these techniques is the projection image, and a suitable filtering algorithm is designed according to the characteristics of the projection object to remove the noise in the image domain, and then the image is reconstructed according to the filtered back-projection algorithm. The results show that the filtered back-projection technique can complete the denoising task of low-dose CT image with less computational overhead. However, in removing the noise in the projection domain, new noise or artifacts will be generated in the reconstructed image. With the attention of the iterative reconstruction (IR) algorithm, especially in the field of low-dose CT image. This method integrates the prior information in the image domain into a unified objective function. Although the iterative reconstruction technique has achieved satisfactory denoising effect, the mainstream iterative reconstruction algorithm has the shortcomings of large computational overhead and slow calculation speed. Another choice for low-dose CT image is to post-process the reconstructed image. The image post-processing technique does not depend on the original data. Therefore, these techniques can be directly applied to low-dose CT images. Common image post-processing techniques include the non-local mean (NLM) method to complete the denoising task, and the block-matching (BM3D) method which is proved to be computationally efficient for various X-ray imaging tasks and can significantly improve the image quality. Although the image post-processing technique has completed the denoising task of low-dose CT image, the texture details, structural information of the image will be lost and new noise will be introduced in the image post-processing process.

[0004] Recently, deep learning has achieved good results in the application of low-dose CT images. Chen et al. proposed a shallow convolutional neural network framework RED-CNN, the first 5 layers of the network are used to extract deep features of the image, and the last 5 layers are used to gradually realize the reconstruction of the image through deconvolution. The network removes the downsampling operation and retains more structural information, thereby obtaining a higher peak signal-to-noise ratio. Yang et al. proposed a generative adversarial network WGAN with perceptual loss, which first improves the denoising performance of the network through the adversarial learning of the generator and the discriminator, and then optimizes the visual effect of the CT image by adding perceptual loss in the architecture.

[0005] However, we find that the above-mentioned networks still have some limitations in network structure. These methods are to suppress the unobvious features in order to mine the obvious features in the image. No matter how many layers the network architecture design has, although these methods can reconstruct most of the details of the image, they usually have poor performance in maintaining some small and irregular texture details. The main reason for this phenomenon is that the texture detail information is not obvious compared with the obvious features on the CT image, so during the feature extraction process in the deep layer of the neural network, the texture detail information is often extracted less, resulting in the neglect of the texture information and the structural edge information in the LDCT image denoising process. In order to solve these problems, the present application proposes a texture feature guided texture preserving low dose CT image denoising method, and a texture feature guided multiscale deep residual attention network model is constructed. The present application proposes a multiscale convolutional coding network based on the advantages of the powerful representation ability of the multiscale convolutional coding and the deep convolutional network, and is used for the denoising of low dose CT image. The method of the present application combines the advantages of multiscale convolutional coding and deep convolutional network, establishes an interpretable multiscale convolutional coding network model, enhances the feature learning ability, fully removes the noise and stripe artifacts in the low dose CT image, improves the quality of the denoised CT image, realizes the high contrast, high resolution, less noise, less artifact, low dose CT image, reduces the additional radiation of patients, and increases the diagnosis and treatment benefit. SUMMARY

[0006] 1. Technical problems to be solved by the application

[0007] The present application aims to overcome the problems in the prior art that the texture detail information is not obvious relative to the features on the CT image, the texture information is lost, the detail preservation and noise artifact residual are difficult to unify after the low dose CT image is denoised, and provides a texture feature guided texture preserving low dose CT image denoising method, which is called a texture feature guided multiscale deep residual attention network model (MDRAN). The method improves the feature information perception, coding and decoding ability through convolution feature learning in multiple scales without changing the existing CT hardware cost, obtains rich prior knowledge from a large amount of data, is applied to low dose CT image denoising, improves the quality of the denoised CT image, realizes low dose CT image with less noise, less artifact and more texture details, and makes a certain contribution to reducing the additional radiation risk of patients.

[0008] 2. Technical scheme

[0009] To achieve the above-mentioned purpose, the technical scheme provided by the present application is:

[0010] The method for texture preserving low-dose CT image denoising based on texture feature guidance comprises the following steps:

[0011] Step 1, a plurality of matched low-dose CT image and normal-dose CT image data are collected, and then each pair of matched low-dose CT image and normal-dose CT image forms a training sample, denoted as {I LD ,I ND}, wherein I LD is a low-dose CT image, I ND is a normal-dose CT image, a network training data set is constructed and used for network training;

[0012] Step 2, a texture feature guided multi-scale deep residual attention network model is established for mapping between low-dose CT image and normal-dose CT image;

[0013] Step 3, the sub-networks in the texture feature guided multi-scale deep residual attention network model are respectively trained using the data set established in step 1, so that the trained network model parameters are obtained;

[0014] Step 4, the low-dose CT image to be processed is input into the multi-scale deep residual attention network model with the obtained network model parameters, and a CT image with removed noise and stripe artifacts is output.

[0015] As a further improvement of the present application, the method for texture preserving low-dose CT image denoising based on texture feature guidance in step 2 constructs a texture feature guided multi-scale deep residual attention network model, characterized in that the constructed texture feature guided multi-scale deep residual attention network model comprises four sub-network models, which are respectively a multi-scale initial denoising network model, a multi-scale texture feature prediction network model, a multi-scale image and texture feature fusion network model and a multi-scale main denoising network model.

[0016] As a further improvement of the present application, the method for texture preserving low-dose CT image denoising based on texture feature guidance in step 2 constructs a multi-scale initial denoising network model, characterized in that a multi-scale initial denoising model is first used to preliminarily denoise the input low-dose CT image to generate a group of output images, denoted as I output .

[0017] As a further improvement of the present application, the method for texture preserving low-dose CT image denoising based on texture feature guidance in step 2 constructs a multi-scale texture feature prediction network model, characterized in that I output output by the multi-scale initial denoising network is used to generate a pair of data sets, namely a texture detail data set, denoted as I output_Edge and a corresponding difference set IDiff , difference set I Diff is the real texture detail information that the primary denoising network cannot learn. Among them, I output_Edge is generated by I output through the Sobel operator, and then input to the texture prediction model to learn the mapping relationship between I output_Edge and I Diff , so as to predict the missing texture detail information I Edge .

[0018] As a further improvement of the application, the texture feature guided texture preserving low dose CT image denoising method in step 2 constructs a multi-scale image and texture feature fusion network model, characterized in that: the texture detail information I Edge predicted by the texture prediction model is output as a texture detail information guided image, which is subjected to multi-scale spatial pyramid attention and further uses a Sigmoid function to obtain attention weights, and then is fused with the input low dose CT image I L through a multi-scale image and texture feature fusion model.

[0019] As a further improvement of the application, the texture feature guided texture preserving low dose CT image denoising method in step 2 constructs a multi-scale main denoising network model, characterized in that: after the low dose CT image and the texture detail information I Edge are fused through the multi-scale image and texture feature fusion model, they are input into the multi-scale main denoising network to further train the multi-scale main denoising network model, so as to output the denoised CT image.

[0020] As a further improvement of the application, the texture feature guided texture preserving low dose CT image denoising method constructs a texture feature guided multi-scale deep residual attention network model, characterized in that: the multi-scale initial denoising network model, the multi-scale texture feature prediction network model and the multi-scale main denoising network model in the constructed multi-scale deep residual attention network model all use Figure 1 the rectangular dashed box structure part as the denoising module of the network model, named multi-scale convolution residual attention network (MRAN). The four sub-networks constructed work cooperatively, two of which contribute more texture detail features to the main denoising network, which is conducive to the main network learning more texture details and achieving the purpose of removing noise. The model establishes a loss function between the denoised CT image and the label data in the form of absolute error loss (L1), iteratively updates the network model parameters through stochastic gradient descent, reduces the loss value, and thus determines the network model parameters.

[0021] As a further improvement of the present application, the method of texture feature guided texture preserving low dose CT image denoising in step 2, the loss function between the denoised CT image and the label data is established in the form of absolute error loss (L1 norm), characterized by: the objective function of low dose CT image denoising is represented as:

[0022]

[0023] Where f(x) represents the denoising function, I ND(i) represents the corresponding normal dose CT image, N is the total number of training samples. Specifically, f(x) = F Mden (F Fusion (F Text (F PreDen (I LD )), here F Mden , F Fusion , F Text and F PreDen respectively represent a multi-scale main denoising network model, a multi-scale image and texture feature fusion network model, a multi-scale texture feature prediction network model and a multi-scale initial denoising network model. Generally, the denoising algorithm based on convolution uses L2 norm loss, but L2 norm loss often produces over-smoothed images, resulting in loss of structural details. Therefore, the present application uses L1 norm loss as the loss function of the whole network training, that is, the loss function of the whole network training.

[0024] 3. Beneficial effects

[0025] Compared with the prior art, the technical scheme provided by the present application has the following beneficial effects:

[0026] The present application discloses a texture feature guided multi-scale deep residual attention network model for low dose CT image denoising, which includes four sub-network models: a multi-scale initial denoising network model, a multi-scale texture feature prediction network model, a multi-scale image and texture feature fusion network model and a multi-scale main denoising network model. Among them, the multi-scale initial denoising network model is used to preliminarily denoise the input low dose CT image, and obtain the preliminarily denoised CT image I output After that, the texture boundary is extracted by the sobel operator, and the real texture information I Diff that cannot be learned in the preliminary denoising network is obtained. output_Edge Then input into the texture prediction model, further learn the mapping relationship of I Diff and I Edge , so as to predict the missing texture detail information I EdgeAs the texture detail information guide map, after the attention weight obtained through the multi-scale space pyramid attention, the low dose CT image is fused through the multi-scale image and texture feature fusion network model, and then input into the multi-scale main denoising network model, so that the multi-scale main denoising network model is further trained, and the denoised CT image is output. The method can realize the denoising of the low dose CT image, establish the mapping between the low dose CT image and the denoised image, so that the denoised CT image has less noise and pseudo noise, high structure detail contrast, contains more texture detail information, and has better image quality, and the results are more conducive to clinical diagnosis and analysis. The experimental results verify that in the denoising of the CT image with about 1 / 4 normal dose, compared with the existing RED-CNN, WGAN-VGG, Non-Local and SACNN, the method (MDRAN) can effectively remove the noise and stripe pseudo noise in the low dose CT image, improve the original low dose CT image quality, increase the texture detail information, and does not damage the anatomical structure information, and the denoised image has good visual effect and contrast. The method can provide an advanced and practical low dose CT image processing framework for domestic hospital imaging departments and CT manufacturers, reduce the additional radiation of patients, increase the diagnosis and treatment benefits, and has high application and popularization prospect. Based on the application, a texture feature guide texture preserving low dose CT image denoising for low dose CT image denoising is disclosed, and a texture feature guide multi-scale deep residual attention network model is invented. Based on the previous research, the multi-scale convolutional coding form is used to increase the richness of feature extraction. Finally, an interpretable network model is established, and the high-quality denoising of the low dose CT image is realized. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 It is a network structure diagram of the low dose CT image denoising method based on the multi-scale deep residual attention network (MDRAN) in the embodiment of the application.

[0028] Figure 2 It is a network structure diagram of the multi-scale convolutional residual attention network (MRAN) in the embodiment of the application.

[0029] Figure 3 It is a network structure diagram of the multi-scale convolutional residual group (MRG) in the embodiment of the application.

[0030] Figure 4 It is a network structure diagram of the multi-scale convolutional residual attention block (MRAB) in the embodiment of the application.

[0031] Figure 5 It is a multi-scale texture feature prediction network model combined with the initial denoising network in the embodiment of the application.

[0032] Figure 6This is a multi-scale image and texture feature fusion network model in the embodiments of the present invention.

[0033] Specific implementation measures

[0034] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings.

[0035] The present invention will be further described below with reference to embodiments.

[0036] This example constructs a texture feature-guided multi-scale deep residual attention network model for low-dose CT image denoising. The network structure diagram is shown below. Figure 1 As shown, the specific steps are as follows:

[0037] Step 1: Collect multiple sets of matched low-dose CT images and normal-dose CT images, construct a dataset, and use it for network training;

[0038] Specifically, the training dataset is designed to use a simulated phantom of a specific body part for scanning. Except for the scanning dose, all other scanning parameters are identical. The goal is to ensure that the scanning positions of the low-dose CT images and the normal-dose CT images are matched. For example, when performing a low-dose CT scan on the abdomen, the dose can be reduced by lowering the current. That is, the strategy of keeping all parameters the same (e.g., scan tube voltage, scan angle, voxel size) except for the scanning current parameter is used for data acquisition. In the constructed dataset, the low-dose CT images constitute the training dataset, and the normal-dose CT images constitute the label dataset.

[0039] Step 2: Establish a multi-scale deep residual attention network model guided by texture features.

[0040] Specifically, the texture feature-guided multi-scale deep residual attention network model includes four sub-network models: a multi-scale initial denoising network model, a multi-scale texture feature prediction network model, a multi-scale image and texture feature fusion network model, and a multi-scale main denoising network model. Among these, the multi-scale initial denoising network model, the multi-scale texture feature prediction network model, and the multi-scale main denoising network model all use... Figure 1 The rectangular dashed box section serves as the denoising module of the network model, named Multi-Scale Convolutional Residual Attention Network (MRAN).

[0041] like Figure 2 As shown, in MRAN, using multi-scale convolution to achieve low-dose CT image denoising helps the network integrate features learned from the image at different scales. Therefore, multi-scale convolution is encapsulated as a multi-scale convolutional residual attention block (MRAB, e.g.) Figure 4 As shown, important channel features can be extracted, facilitating deep feature extraction. Then, multiple MRABs are cleverly embedded into multi-scale convolutional residual groups (MRGs, such as...).Figure 3 Thus together with the denoising module MRAN we need. Due to the effective combination of this attention mechanism and the residual network, the balance ability of MRAN network in denoising and keeping texture details is improved, so that the MRAN denoising network can better learn image features.

[0042] As shown in Figure 3 In the multi-scale convolution residual group (MRG), the network structure is composed of six multi-scale convolution residual attention blocks (MRAB), convolution operations and short skip connections, so that the low resolution and high resolution feature information are combined, which helps to reduce the important feature details lost during downsampling.

[0043] In the multi-scale convolution residual group (MRG), the basic module of MRG is the multi-scale convolution residual attention block (MRAB), as shown in Figure 4As shown in the figure. Among them, in the MRAB, it is composed of three 3x3 convolutions, one multi-scale selective convolutional neural network and global context attention block GC block. We decompose the larger and more demanding 5X5 and 7X7 convolutional layers into a set of smaller and more lightweight 3X3 convolutional operations instead. The output of the 2nd and 3rd 3X3 convolutional layers is closer to the 5X5 and 7X7 convolutional operations, so we extract the output from the three convolutional layers and connect them together to extract spatial features of different scales. Compared with traditional convolutional neural networks, the multi-scale convolutional neural network achieves better results while reducing the need for memory operations. Using a multi-scale convolutional neural network to denoise low-dose CT images helps the denoising network learn CT image features at different scales. In order to further improve the feature expression ability and enhance the image quality, we connect a multi-scale selective kernel convolution (MSKConv) after the multi-scale convolution. The multi-scale selective kernel convolution is composed of two operators, fusion operation and selection operation. The fusion operation combines and summarizes the information of multiple paths to obtain a global and comprehensive selection weight representation. The selection operation summarizes the kernel feature maps of different sizes according to the selection weight. We first fuse the feature map information generated by the previous multi-scale splicing through element addition, and then embed global information by generating data information in the channel range through global average pooling. Then a simple fully connected layer is used to obtain a compact feature to reduce the dimension and improve the operation efficiency. In the selection operation, the information of different spatial scales guided by the compact feature is adaptively selected through soft attention, and finally the final feature map is obtained through the attention weight of each convolution kernel. In addition, in the traditional U-net network architecture, the encoder and the decoder are directly connected through the skip connection. The skip connection can combine the feature information of low resolution and high resolution, which helps to reduce the loss of important feature details during downsampling. However, the fast link transmits all the features extracted in the encoder to the decoder. For the encoder, the first few layers perform feature extraction, while the features in the decoder are more advanced, and the two sets of features are more different. In order to solve the above problems, we add a global context attention block (GC block) after the multi-scale convolution residual attention block (MRAB). Compared with the traditional SE, CBAM attention module, the global context attention block (GC block) can implicitly learn rich feature details in the feature space, thereby making up for the difference between the features. We effectively combine the attention block with the multi-scale selective kernel convolution, so as to make more full use of the global context information and improve the overall learning ability of the network.

[0044] Unlike other low-dose CT denoising networks, a multi-scale texture feature prediction network model is added in the invention proposed in this paper. The texture feature prediction network structure combined with the multi-scale initial denoising network model is as followsFigure 5 The training set used in the training stage is the Sobel edge extraction I output_Edge , and the Sobel edge extraction I Diff of the difference between the normal dose CT image and the generated image. The added multi-scale texture feature prediction model makes the denoised CT image contain more texture feature detail information that the network itself cannot learn.

[0045] In order to better capture the structural information and channel information between the attention modules, and make up for the semantic gap between the deep features. Our texture fusion model is based on the spatial pyramid model (SPANet) as shown in Figure 6 Different from the spatial pyramid model, we add a multi-scale residual group (MRG) to the model to extract channel relationships, and introduce a multi-scale spatial pyramid attention structure (MSPA) into the model to encode the intermediate features. The multi-scale spatial pyramid attention structure mainly consists of three parts: channel adapter, spatial pyramid structure and channel information structure. Two convolutions are introduced in the channel adapter to integrate channel information and match the channel number of the output feature map. The multi-scale spatial pyramid structure uses four different sizes of feature contexts, which combines global, coarse, fine and detailed features to extract global information, structural information and detailed information. As shown in Figure Six The global average pooling provides global information and regularizes the network structure, and the coarse and fine global average pooling is to provide more structural information. The original feature map is passed through a 1x1 convolution to make the output feature map contain all the feature details, and then the four output upsampling operations are summed and input into the channel information structure. Two convolutions and BN normalization in the channel information structure capture the channel relationship between the attention, learn the attention feature map output from the multi-scale spatial pyramid structure, and further use the Sigmoid function to generate attention weights. Then the generated attention weight matrix is multiplied element by element with the feature map generated after the multi-scale residual group of the low dose CT, to obtain the spatial attention feature, and finally the residual connection is made with the feature map generated after the multi-scale residual group of the low dose CT to obtain the output.

[0046] The application constructs a multi-scale deep residual attention network model for texture feature guided texture preserving low dose CT image denoising, which can mine deep information and predict texture feature detail information through the cooperative work of four sub-networks, and fuse the texture feature detail information with the low dose CT image as a guide image to participate in the training of the main network, so as to obtain a denoised CT image containing more texture details.

[0047] Step 3, combine the texture feature guided multi-scale deep residual attention network model, and the specific training stage is as follows.

[0048] First, the low-dose CT image I LD is input into the preliminary denoising model F PreDen , and the output is I output . The expression is as follows:

[0049] I output = F Pre (I LD ) (1)

[0050] Then, the Sobel operator is used to generate some dense edge information, which is named I output_Edge , which contains significant features and weak texture features, and further enhances the real texture detail information I Diff . In Figure 5 , a network named texture prediction model is constructed to learn the correlation between I output_Edge and the real texture information I Diff . I output_Edge is input into the texture prediction model, denoted as F Text , and the predicted texture detail information I Edge can be obtained, and the expression is as follows:

[0051] I Edge = F Text (I output_Edge ) (3)

[0052] In the multi-scale texture feature prediction model, a convolution layer is used to extract shallow features from I output_Edge , and then a residual module containing multiple multi-scale convolution residual groups is input to extract deep features, and then a convolution layer is input, which is used for feature encoding. Therefore, formula (3) can be further expressed as:

[0053] I Edge = F conv (F MRG (F conv (I output ))) (4)

[0054] Where F MRG and F conv represent residual operation and convolution operation respectively. I Edge is the output of the multi-scale texture feature prediction model, and then I Edge is input into the multi-scale spatial pyramid attention mechanism, and the channel relationship between the attention mechanisms is obtained through the attention mechanism. Further, the attention weight is obtained using the Sigmoid function, and then it is fused with the original low-dose CT image to form a new feature map, denoted as F Fusion . The expression is as follows:

[0055] I{I LD I Edge}=F Fusion (I LD I Edge (5)

[0056] The new feature maps are used as training data for the main denoising network {I} LD ,I Edge Then, it is input into the multi-scale master denoising network for further denoising, denoted as F. Mden Its expression is as follows:

[0057] I H =F Mden (I{I LD ,I Edge})

[0058] In the main denoising network, a convolutional layer is used to extract shallow features, which are then fed into several multi-scale convolutional residual groups. Each multi-scale convolutional residual group consists of several residual blocks and short-slap connections. Two pooling layers are used for dimensionality reduction, denoted as F. pool Then, after upsampling and concat connection operations, F up The data is then input into a convolutional layer for feature encoding, and finally output as a denoised CT image containing more texture details. Therefore, formula (6) can be further expressed as:

[0059] I H =F up (F pool (F MRG (F conv (I{I LD ,I Edge})))) (7).

[0060] Step 4: Use the trained multi-scale deep residual attention network to process low-dose CT images and output denoised CT images.

[0061] Specifically, the low-dose CT images to be processed are input into a multi-scale depth residual attention network model whose network model parameters have been obtained, and the denoised CT images are output.

[0062] Invention Conclusion

[0063] The application constructs a multi-scale deep residual attention network model for texture feature guided texture preserving low-dose CT image denoising. Compared with the traditional U_net denoising network, in MDRAN, we use a multi-scale convolution residual attention network (MRAN) as a denoising module, which can better suppress noise and greatly improve the denoising performance of the network. At the same time, in order to better eliminate the noise and artifacts existing in the low-dose CT image and preserve more texture feature detail information, we design a unique noise removal and texture detail information prediction subnetwork, and the two subnetworks work together to remove noise while predicting texture feature information, so that the texture feature information is better preserved. Then the predicted texture feature information and the low-dose CT image enter the fusion model, and the training of the training pair as the guide image is input into the main denoising network for further training. The four subnetworks work together, which is conducive to the main subnetwork to remove noise and learn more texture details. From the above experiments, it can be seen that the method (MDRAN) can effectively remove the noise and stripe artifacts in the low-dose CT image, improve the quality of the original low-dose CT image, increase the texture detail information, and not destroy the anatomical structure information, and the denoised image has good visual effect and contrast. The method of the application is expected to provide an advanced and practical low-dose CT image processing framework for domestic hospital imaging departments and CT manufacturers, reduce the additional radiation for patients, increase the diagnosis and treatment benefits, and has high application and promotion prospects.

[0064] The above describes the application and its embodiments in a schematic manner, which is not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. Therefore, if a person of ordinary skill in the art is inspired thereby, without departing from the spirit of the application, similar structural modes and embodiments can be designed without creativity, which should all belong to the protection scope of the application.

Claims

1. A method for texture preserving low dose CT image denoising based on texture feature guided, characterized in that, Comprising the following steps: Step 1, collect multiple sets of matched low-dose CT images and normal-dose CT image data, then each pair of matched low-dose CT images and normal-dose CT images constitutes a training sample, denoted as {I LD , ND} LD low-dose CT image I ND normal-dose CT image, construct a network training data set and use it for network training; Step 2: Construct a texture feature-guided multi-scale deep residual attention network (MDRAN) model for mapping between low-dose CT images and normal-dose CT images. The constructed texture feature-guided MDRAN model consists of four sub-network models: a multi-scale initial denoising network model, a multi-scale texture feature prediction network model, a multi-scale image and texture feature fusion network model, and a multi-scale main denoising network model. The specific working process of the constructed texture feature-guided MDRAN model is as follows: First, the multi-scale initial denoising network model is used to process the input low-dose CT image I... LD Perform preliminary denoising to generate a set of output images, named I. output The I output after the multi-scale initial denoising network model output It was used to generate a pair of datasets, namely the texture detail dataset, named I. output_Edge and the corresponding difference set I Diff Among them, difference set I Diff It is the real texture detail information that multi-scale initial denoising network models cannot learn. output_Edge It is I output Generated using the Sobel operator, then input into a multi-scale texture feature prediction network model to learn I output_Edge and I Diff The mapping relationship is used to predict the missing texture detail information I. Edge Texture detail information I predicted by the multi-scale texture feature prediction network model Edge As a texture detail information guide map, it undergoes multi-scale spatial pyramid attention, and further uses the Sigmoid function to obtain attention weights, before being compared with the input low-dose CT image I. LD Fusion was performed using a multi-scale image and texture feature fusion network model; low-dose CT images I LD and texture detail information I Edge After passing through the multi-scale image and texture feature fusion network model, the image is then input into the multi-scale master denoising network model, thereby outputting the denoised CT image. Step 3, using the data set established in step 1, the sub-networks in the texture feature guided multi-scale deep residual attention network model MDRAN are trained respectively, so as to obtain the trained network model parameters; Step 4, input the low-dose CT image to be processed into the texture feature guided multi-scale deep residual attention network model MDRAN with the obtained network model parameters, and output the CT image with removed noise and stripe artifacts.

2. The method for texture feature guided texture preserving low dose CT image denoising according to claim 1, wherein: The texture feature guided multi-scale deep residual attention network model MDRAN is constructed, a loss function between the denoised CT image and the label data is established in the form of absolute error loss, the network model parameters are iteratively updated by stochastic gradient descent, the loss value is reduced, and thus the network model parameters are determined.

3. The method for texture feature guided texture preserving low dose CT image denoising according to claim 2, characterized in that: The loss function between the denoised CT image and the label data is established in the form of absolute error loss, and is represented as: wherein f(x) represents a texture feature guided multi-scale deep residual attention network model MDRAN, I LD(i) represents the input i-th low-dose CT image, I ND(i) represents the corresponding i-th normal-dose CT image, N is the total number of training samples; specifically f(x) = F Mden (F Fusion (F Text (F PreDen (I LD )))), wherein F Mden , F Fusion , F Text and F PreDen respectively represent a multi-scale initial denoising network model, a multi-scale texture feature prediction network model, a multi-scale image and texture feature fusion network model and a multi-scale main denoising network model; an L1 norm loss is used as a loss function for training the whole network.

Images