A Low-Dose CT Denoising Method Based on Symmetric-Asymmetric Collaborative Module

Abstract

The present invention provides a low-dose CT denoising method based on a symmetric-asymmetric collaborative module. The method includes: obtaining a high-dose CT image and a corresponding low-dose CT image; constructing a symmetric-asymmetric collaborative module SCM; and using the symmetric-asymmetric collaborative module SCM to denoise the low-dose CT image. In the present invention, in the intermediate feature map size of the network, it can be well nested in various networks, solve the problem of insufficient extraction of irregular tissue and lesion features, and achieve the effect of reducing network parameters.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and particularly to a low-dose CT denoising network method based on a symmetric-asymmetric collaborative module. Background Art

[0002] High-dose computed tomography (CT) images can clearly display the tissue structure and lesions in the human body. Radiologists diagnose diseases based on CT images, such as fractures and tumors. However, the radiation brought by multiple high-dose CT scans will have an irreversible impact on the human body, such as increasing the risk of cancer. People can reduce the radiation of X-rays by reducing the current or voltage, that is, low-dose CT imaging, but it will inevitably introduce noise into the reconstructed image, reducing the visualization effect of the image and being unfavorable for doctors' diagnosis. Therefore, when using low-dose CT imaging, it is particularly important to restore the structural information of the image as much as possible.

[0003] Currently, in low-dose CT denoising, the denoising methods mainly include projection domain denoising, iterative reconstruction denoising, and image domain denoising. Projection domain denoising through sinogram denoising and iterative reconstruction denoising that converts between the sinogram and the image domain both require access to the original sinogram data of the CT scanner. Due to its non-disclosure, it is difficult to obtain, which limits the development of these two types of methods. Image domain post-processing denoising only processes images that remove patient privacy and has been widely used. It can be further divided into traditional machine learning and deep learning. Traditional machine learning mainly uses statistical methods to calculate the similarity between the target pixel point and other pixel points in different domains, which is complex and time-consuming. With the development of computer technology, deep learning has also made good progress in low-dose CT denoising.

[0004] Since the development of deep learning, various network models have shone in low-dose CT denoising. The encoder-decoder network extracts the high-level semantic information of the image and restores the low-level detail information of the image through the encoder and decoder, and has strong feature extraction capabilities. Its variant REDCNN has achieved good quantitative results through convolution, deconvolution, and residual operations, even better than many deep learning methods proposed currently. The generative adversarial network uses the adversarial principle of the generator and the discriminator to make the denoised CT image continuously approach the high-dose CT image. The WGAN-VGG adversarial network adds a perceptual loss based on the pre-trained VGG in the discriminator, making the denoised image closer to the human eye's observation of the image and improving the qualitative index of the image. Summary of the Invention

[0005] The object of the present invention is to provide a low-dose CT denoising method based on a symmetric-asymmetric collaborative module, which can be well nested in various networks in the middle-stage feature map size of the network, solve the problem of insufficient extraction of irregular tissue and lesion features, and achieve the effect of reducing network parameters.

[0006] A low-dose CT denoising method based on a symmetric-asymmetric collaborative module, comprising:

[0007] Obtaining a high-dose CT image and a low-dose CT image corresponding to the high-dose CT image;

[0008] Constructing a symmetric-asymmetric collaborative module SCM;

[0009] Using the symmetric-asymmetric collaborative module SCM to denoise the low-dose CT image.

[0010] The high-dose CT image and the low-dose CT image corresponding to the high-dose CT image include:

[0011] The high-dose CT image is a CT image obtained by a CT scanner at a normal X-ray dose;

[0012] The low-dose CT image is obtained by adding Poisson noise to the high-dose CT image.

[0013] Constructing the symmetric-asymmetric collaborative module SCM includes:

[0014] Improving the ordinary convolution in the convolutional layer after the second downsampling of the encoder-decoder network, and replacing the second and fourth convolution operations with asymmetric convolutions.

[0015] After using the symmetric-asymmetric collaborative module SCM to denoise the low-dose CT image, it further includes evaluating the performance of SCM through a loss function, specifically:

[0016] The loss function includes the mean square error MSE loss based on pixel points, the multi-scale perception loss based on the high-level feature space, and the gradient loss based on gradients;

[0017] The hybrid loss formula is:

[0018] L loss = w mse ·L mse + w multi-p ·L multi-p + w grd ·L grd

[0019] w mse , w multi-p , w grd respectively correspond to MSE, L multi-pMulti-scale perception loss, L grd The weight of the gradient loss, adjust the parameter weight to balance the loss function, L loss is the hybrid loss.

[0020] The mean square error MSE loss includes:

[0021]

[0022] In the formula, x i represents the low-dose CT image, y i represents the high-dose CT image, F is the noise reduction model of parameter θ. Therefore, F(x i , θ) represents the noise-reduced CT image, and N is the number of feature maps.

[0023] The multi-scale perception loss based on the high-level feature space includes:

[0024]

[0025] In the formula, φ is the trained parameter of the Resnet-50 network model, s is the number of features of different sizes, and φ s represents the s-th layer feature extracted from the Resnet-50 network;

[0026] Assume that the dataset to be operated on is A, and the sobel operators in the horizontal and vertical directions are S x and S y , G x and G y are the results after convolution operations. The formulas are as follows:

[0027] G x = S x * A G y = S y * A

[0028] The pixel value of the action point in the original image after convolution in the horizontal and vertical directions is G:

[0029]

[0030] The gradient loss based on the gradient includes:

[0031] L grd = ||G den - G ndct ||

[0032] G den and G ndct are the results after convolution of the noise-reduced CT with the corresponding high-dose and sobel respectively. Then, the L1 norm is used to calculate the difference between the two to obtain the gradient loss Lgrd where * represents the convolution operation.

[0033] A low-dose CT denoising system based on a symmetric-asymmetric collaborative module, comprising:

[0034] An image acquisition module for acquiring a high-dose CT image and a low-dose CT image corresponding to the high-dose CT image;

[0035] A data processing module for constructing a symmetric-asymmetric collaborative module SCM;

[0036] An image processing module for denoising the low-dose CT image using the symmetric-asymmetric collaborative module SCM.

[0037] When the processor executes the computer program, it implements the low-dose CT denoising method according to any one of the above.

[0038] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the low-dose CT denoising method according to any one of the above.

[0039] By fusing asymmetric convolution in the convolution layer after the second-layer downsampling in the present invention, while not reducing the network performance, the parameters of the network are reduced, which has good scalability, and compared with the prior art, the image details are richer and the processing efficiency is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings herein are incorporated into the specification and form a part of the specification, indicating the embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0042] Figure 1 is a flowchart of the present invention;

[0043] Figure 2 is a diagram of the SCM model of the present invention;

[0044] Figure 3 is a comparison diagram of the experimental results of the present invention;

[0045] Figure 4 is an enlarged view of the rectangular frame of the present invention;

[0046] Figure 5It is a diagram showing different combinations of symmetric convolution and asymmetric convolution of the present invention;

[0047] Figure 6 It is a diagram of the network parameters of the symmetric-asymmetric collaborative module SCM of the present invention;

[0048] Figure 7 It is a diagram of the dataset in npy format of the present invention. Detailed implementation manners

[0049] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0050] It should be noted that all directional indications (such as up, down, left, right, front, back...) in the embodiments of the present invention are only used to explain the relative positional relationship and movement conditions between components in a specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications will also change accordingly.

[0051] In addition, the descriptions involving "first", "second", etc. in the present invention are only for descriptive purposes, and cannot be understood as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between various embodiments can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0052] The present invention discloses a method for predicting the secondary water supply volume based on big data. Refer to Figure 1-7 , including:

[0053] A low-dose CT denoising method based on a symmetric-asymmetric collaborative module, including:

[0054] S100, obtaining a high-dose CT image and a low-dose CT image corresponding to the high-dose CT image;

[0055] S200, constructing a symmetric-asymmetric collaborative module SCM;

[0056] S300, using the symmetric-asymmetric collaborative module SCM to denoise the low-dose CT image.

[0057] The present invention fuses asymmetric convolutions in the convolutional layer after the second-layer downsampling, reducing the parameters of the network without degrading its performance, having good scalability, and having richer image details and higher processing efficiency compared to the prior art.

[0058] The high-dose CT images and the corresponding low-dose CT images of S100 include:

[0059] The high-dose CT images are CT images obtained by a CT scanner at normal X-ray doses;

[0060] The low-dose CT images are obtained by adding Poisson noise to the high-dose CT images.

[0061] Download the dicom datasets of 10 patients in the low-dose CT challenge competition, including high-dose CT and the corresponding low-dose CT (the high-dose CT images are CT images obtained by a CT scanner at normal X-ray doses, and the low-dose CT is obtained by adding Poisson noise to the high-dose CT images), and divide the dataset into a training set and a test set according to a ratio of 9:1. And read the image information in the dicom and convert it to the npy format.

[0062] Read the information in the dicom dataset through the read_file of the third-party library pydicom in the python language, including the CT image data matrix read by pixel_array, and convert the data matrix to between -1024 and 3071 by the scaling slope represented by RescaleIntercept and the scaling intercept of RescaleSlope. After normalization, name it in the format of "patient name_sequence number_input.npy" and "patient name_sequence number_target.npy" (input represents low-dose CT, target represents high-dose CT, and the sequence number represents the number of the image of this patient), as Figure 7 shown.

[0063] S200 constructs a symmetric-asymmetric collaborative module SCM, including:

[0064] Improve the ordinary convolution in the convolutional layer after the second-layer downsampling of the improved encoder-decoder network, and replace the second and fourth convolution operations with asymmetric convolutions.

[0065] Since the asymmetric convolution has different effects on feature maps of different sizes, in order to make the asymmetric convolution play its maximum role, improve the ordinary convolution in the convolutional layer after the second-layer downsampling of the improved encoder-decoder network, replace the second and fourth convolution operations with asymmetric convolutions, form a symmetric-asymmetric-symmetric-asymmetric structure, and name it a symmetric-asymmetric collaborative module to solve the problem that the network extracts irregular information insufficiently.

[0066] A hybrid loss function combining pixel points, high-level feature space, and gradient is used to guide the network to learn the information within the image and predict the performance of the network.

[0067] The original codec network can extract the feature information of the image well through the encoder and decoder, but the square convolution kernel of the ordinary convolution in the encoder has limited ability to extract irregular information. We replace the ordinary convolution with asymmetric convolution in the convolution layer after the second downsampling of the encoder. The convolution kernel with inconsistent height and width can effectively process and aggregate the information extracted by the previous ordinary convolution, thereby improving the network's extraction of irregular information.

[0068] The original codec network can extract the feature information of the image well through convolution, up-sampling and down-sampling and splicing channel operations, but the square convolution kernel of ordinary convolution has limited ability to extract the features of irregular structures, and the convolution kernel size in the original network is 5*5. Large convolution kernels are not conducive to the extraction of small structural features. Therefore, the present invention proposes to improve the convolution layer by combining asymmetric convolution to solve the problem of insufficient feature extraction of CT images. The overall network structure is as follows: Figure 2 .

[0069] From VGGNet proposing to stack small receptive field convolution layers to achieve a large receptive field, to GoogleNet reducing the number of channels through the Inception operation to reduce network parameters, to the proposal of asymmetric convolution, all of them are reducing network parameters without reducing network performance. The original codec network has a total of 18 convolution operations, which has good scalability. In the experimental process of asymmetric convolution, it was found that asymmetric convolution has certain requirements on the size of the feature map, and the effect is not good in the early convolution process. In order to better utilize the advantages brought by asymmetric convolution, the present invention integrates asymmetric convolution in the convolution layer after the second layer of downsampling. Asymmetric convolution is essentially a convolution operation, but the convolution process is different from ordinary convolution. In order to explore the impact of the combination of the two convolutions, the following is proposed. Figure 5 The three combinations, based on the output of the network’s intermediate feature map, use a symmetric-asymmetric-symmetric-asymmetric combination to form a symmetric-asymmetric collaborative module. The detailed network parameters are as follows: Figure 6 shown.

[0070] To verify the performance of the encoding and decoding network integrating SCM, we used the dataset in npy format processed from the challenge competition dataset and made comparisons with traditional learning method KSVD and classical deep learning methods. In the quantitative metrics, the peak signal-to-noise ratio (PSNR) based on the pixel gray-level difference and the structural similarity (SSIM) based on the structural information difference were used as the measurement criteria; in the qualitative metrics, a randomly selected abdominal CT image was taken as an example and multiple details of the texture were magnified and displayed. It can be found that the quantitative metrics of the encoding and decoding network integrating SCM achieved the best results in both PSNR and SSIM, and the details of the image were richer, such as Figure 4 indicated by the circles and arrows.

[0071] Table 1 Comparative test results of different algorithms

[0072]

[0073] After S300 uses the symmetric-asymmetric collaborative module SCM to denoise low-dose CT images, S400 also includes evaluating the performance of SCM through a loss function, specifically:

[0074] The loss function includes the mean square error (MSE) loss based on pixel points, the multi-scale perception loss based on the high-level feature space, and the gradient loss based on gradients;

[0075] The hybrid loss formula is:

[0076] L loss = w mse ·L mse + w multi-p ·L multi-p + w grd ·L grd

[0077] w mse , w multi-p , w grd correspond to MSE, L multi-p multi-scale perception loss, and L grd gradient loss weights respectively. Adjust the parameter weights to balance the loss function, and L loss is the hybrid loss.

[0078] When training the denoising network, the loss function is used to calculate the difference between the denoised CT image and the high-dose CT, guiding the network to further correctly learn the image features; after determination, the loss function is used to measure the performance of the model. The hybrid loss function consists of three parts: the mean square error (MSE) loss based on pixel points, the multi-scale perception loss based on the high-level feature space, and the gradient loss based on the gradient.

[0079] The mean square error MSE loss includes:

[0080]

[0081] In the formula, x i represents the low-dose CT image, y i represents the high-dose CT image, F is the denoising model with parameter θ, so F(x i , θ) represents the denoised CT image, and N is the number of feature maps.

[0082] The multi-scale perception loss based on the high-level feature space includes:

[0083]

[0084] In the formula, φ is the Resnet-50 network model with trained parameters , s is the number of features of different sizes, and φ s represents the feature of the s-th layer extracted from the Resnet-50 network;

[0085] Assume that the dataset to be operated on is A, and the sobel operators in the horizontal and vertical directions are S x and S y , G x and G y are the results after convolution operations, and the formula is as follows:

[0086] G x = S x *A G y = S y *A

[0087] The pixel value of the action point in the original image after convolution in the horizontal and vertical directions is G:

[0088]

[0089] The gradient loss based on the gradient includes:

[0090] L grd = ||G den - G ndct ||

[0091] G denAnd G ndct are the results of the denoising CT and the corresponding high-dose and Sobel convolution respectively. Then, the L1 norm is used to calculate the difference between the two to obtain the gradient loss L grd , and * represents the convolution operation.

[0092] A low-dose CT denoising system based on a symmetric-asymmetric collaborative module, comprising:

[0093] An image acquisition module for acquiring a high-dose CT image and a low-dose CT image corresponding to the high-dose CT image;

[0094] A data processing module for constructing a symmetric-asymmetric collaborative module SCM;

[0095] An image processing module for denoising the low-dose CT image using the symmetric-asymmetric collaborative module SCM.

[0096] When the processor executes the computer program, it implements any one of the low-dose CT denoising methods based on a symmetric-asymmetric collaborative module.

[0097] A computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements any one of the low-dose CT denoising methods based on a symmetric-asymmetric collaborative module.

[0098] In the present invention, asymmetric convolution is fused in the convolution layer after the second-layer downsampling. Without reducing the network performance, the parameters of the network are reduced, and it has good scalability. Moreover, compared with the prior art, the image details are richer and the processing efficiency is higher.

[0099] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A low-dose CT denoising method based on a symmetric-asymmetric collaborative module, characterized in that, It includes: Obtain a high-dose CT image and a low-dose CT image corresponding to the high-dose CT image; Construct a Symmetric-Asymmetric Collaborative Module (SCM); Use the Symmetric-Asymmetric Collaborative Module (SCM) to denoise the low-dose CT image; The construction of the Symmetric-Asymmetric Collaborative Module (SCM) includes: Improve the ordinary convolution in the convolutional layer after the second downsampling of the encoder-decoder network, and replace the second and fourth convolution operations with asymmetric convolutions; After using the Symmetric-Asymmetric Collaborative Module (SCM) to denoise the low-dose CT image, it also includes evaluating the performance of the SCM through a loss function, specifically: The loss function includes the Mean Squared Error (MSE) loss based on pixel points, the multi-scale perception loss based on the high-level feature space, and the gradient loss based on gradients; The hybrid loss formula is: ； corresponding to MSE respectively, multi-scale perception loss, weights of gradient loss, adjusting parameter weights to balance the loss function, is the hybrid loss; The multi-scale perception loss based on the high-level feature space includes: ； In the formula, is the trained parameter of the Resnet-50 network model, is the number of feature maps of different sizes, represents the -th layer feature extracted from the Resnet-50 network; Let the dataset to be operated on be A, and the Sobel operators in the horizontal and vertical directions be and , and be the results after convolution operation, represents the convolution operation, and the formula is as follows: ； ； The result after convolution of the acting point pixel value in the original image in the horizontal and vertical directions is as follows: ； The gradient loss based on gradients includes: ； and are the results after Sobel convolution of the noise-reduced CT and the corresponding high dose respectively, and then use the normal form to calculate the difference between the two to obtain the gradient loss ; The Mean Squared Error (MSE) loss includes: ； In the formula, represents a low-dose CT image, represents a high-dose CT image, is a parameter of the noise reduction model. Therefore, represents the noise-reduced CT image, is the number of feature maps.

2. The low-dose CT noise reduction method based on a symmetric-asymmetric collaborative module according to claim 1, characterized in that The high-dose CT image and the low-dose CT image corresponding to the high-dose CT image include: The high-dose CT image is a CT image obtained by a CT scanner at normal X-ray doses; The low-dose CT image is obtained by adding Poisson noise to the high-dose CT image.

3. A low-dose CT noise reduction system based on a symmetric-asymmetric collaborative module, which is applied to a low-dose CT noise reduction method based on a symmetric-asymmetric collaborative module according to any one of claims 1-2, and is characterized in that, It includes: An image acquisition module for obtaining a high-dose CT image and a low-dose CT image corresponding to the high-dose CT image; A data processing module for constructing a Symmetric-Asymmetric Collaborative Module (SCM); An image processing module for using the Symmetric-Asymmetric Collaborative Module (SCM) to denoise the low-dose CT image.

4. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements a low-dose CT denoising method based on a Symmetric-Asymmetric Collaborative Module as described in any one of claims 1-2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a low-dose CT denoising method based on a Symmetric-Asymmetric Collaborative Module as described in any one of claims 1-2.

Images