Low-dose ct image denoising network based on residual multi-scale feature extraction

Abstract

The present application relates to a low-dose CT image denoising network based on residual multi-scale feature extraction, adopts an encoder-decoder framework, and comprises the following steps: adding a multi-scale feature extraction module to a residual connection of an encoder and a decoder convolution layer; and using a mixed loss function composed of an MSE loss, an SSIM loss and a perception loss to guide the denoising network.The present application adds a multi-scale feature extraction module to a residual connection, thereby increasing the information utilization rate of an input LDCT image, improving the operation speed while obtaining higher performance indicators; the use of a BN layer alleviates the overfitting phenomenon that occurs with the increase of model complexity; in order to further generate a denoised image with high correlation to human perception, the denoising network model is guided by the MSE loss, the SSIM loss and the perception loss, thereby further improving the visual effect of the denoised image.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to a low-dose CT image denoising network based on residual multi-scale feature extraction. BACKGROUND

[0002] Computed Tomography (CT) has been widely used in clinical medical diagnosis (such as tumor screening, cancer staging, human multi-site puncture biopsy, etc.) due to its advantages of fast imaging speed, good clarity, low requirement for patients, non-invasive and painless, etc. However, its ionizing radiation may cause adverse effects on patients and increase the risk of lesions in the patient's lesion area. Therefore, the ALARA (as low as reasonably achievable) principle should be followed when the patient is examined. That is, under the premise of not affecting the clinical diagnosis result, the radiation dose brought by CT scanning should be reduced as much as possible. Therefore, low-dose CT (LDCT) technology has attracted more and more attention, however, the lower the dose, the greater the noise of the reconstructed CT image, which leads to difficulty in diagnosis.

[0003] In the past few decades, researchers have developed many algorithms to solve the image quality problem of LDCT. They can be divided into two categories: one is the traditional LDCT denoising algorithm; the other is the learning-based LDCT denoising algorithm. The traditional LDCT denoising algorithm mainly includes: projection domain filtering algorithm, iterative reconstruction algorithm and post-processing algorithm. The projection domain filtering algorithm is also known as the domain processing algorithm, which directly takes the projection image as the optimization variable, filters the random noise of the projection image, and finally applies the filtered back-projection algorithm to obtain the reconstructed image. Typical methods mainly include: structure adaptive filtering, bilateral filtering and penalized weighted least-square (PWLS) algorithm; the iterative reconstruction algorithm mainly uses the Bayesian and imaging physics theory, according to the prior characteristics of the projection data, uses the likelihood function to relate the projection and the reconstructed image, and integrates the statistical characteristics of the image to be reconstructed to construct the objective function; the image post-processing algorithm is mainly to denoise the reconstructed image again, which is different from the former two algorithms, and does not depend on the original data, and can be directly applied to the LDCT image to realize the end-to-end mapping of the LDCT image to the NDCT (Normal Dose Computed Tomography) image. However, this method cannot determine the accurate distribution of noise, so it is difficult to achieve a balance between feature preservation and noise removal.

[0004] In recent years, due to the powerful function of data-driven, deep learning has been widely used in LDCT image denoising and artifact suppression due to its strong feature learning ability, wide adaptability and good portability. Deep learning post-processing algorithm is a research hotspot. It mainly realizes the end-to-end mapping of LDCT image to NDCT image through training a convolutional neural network (CNN). Chen et al. proposed a low-dose CT residual encoder-decoder convolutional neural network (RED-CNN), which combined the shortcut connection of deconvolution network into the CNN model, realized the improvement of LDCT image quality, and made a new height in both quantitative index measurement and subjective visual effect, but due to the use of mean square error (MSE) as the loss function, the denoising image is oversmoothed, causing visual embarrassment for clinicians in diagnosis; Yang et al. proposed a LDCT image denoising algorithm based on generative adversarial network (GAN), the proposed algorithm has Wasserstein distance and perceptual similarity, which transfers visual perception knowledge to image denoising task, not only can reduce the noise level of image, but also can maintain key information. Li et al. proposed a residual attention module RAM, and incorporated the module into two neural networks for LDCT image denoising, developed REDCNN-RAM and WGAN-RAM, although good results were obtained, but the model is complex and the running speed is slow. SUMMARY

[0005] The purpose of the present application is to provide a low-dose CT image denoising network based on residual multi-scale feature extraction, which can improve the operation speed and has better denoising effect.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is:

[0007] The low-dose CT image denoising network based on residual multi-scale feature extraction adopts an encoder-decoder framework, which comprises:

[0008] The multi-scale feature extraction module is added to the residual connection of the encoder and decoder convolutional layer;

[0009] The hybrid loss function composed of MSE loss, SSIM loss and perceptual loss is used to guide the denoising network.

[0010] Preferably, the encoder is composed of 6 convolutional layers, which are 2 layers of 5x5 shallow feature extraction layers and 4 layers of 3x3 deep feature extraction layers; the decoder is composed of 6 deconvolutional layers, which are 4 layers of 3x3 deconvolutional layers and 2 layers of 5x5 deconvolutional layers.

[0011] Preferably, the convolutional layers are all padded with zeros to ensure the image size of the input and output is consistent.

[0012] Preferably, a BN layer is added after each of the convolutional layers; the convolutional layers of the encoder are activated by a ReLU function; the deconvolutional layers of the decoder are activated by a PReLU function.

[0013] Preferably, the residual connections include a first residual connection, a second residual connection, and a third residual connection.

[0014] The first residual connection is connected from the original input to the BN layer after the last layer of the decoder.

[0015] The second residual connection is connected from the BN after the second layer of the encoder to the BN after the second-to-last layer of the decoder.

[0016] The third residual connection is connected from the BN after the fourth layer of the encoder to the BN after the fourth-to-last layer of the decoder.

[0017] Preferably, the multi-scale feature extraction module is denoted as an MSFE module; the MSFE module includes an MSFEA module and an MSFEB module; the MSFEA module is added to the second residual connection; and the MSFEB module is added to the third residual connection.

[0018] Preferably, the MSFEA module divides the input into 4 paths for multi-scale feature extraction, including 3 convolutional branches and 1 direct connection branch; the convolutional kernel sizes of the 3 convolutional branches are (1x1), (1x1, 3x3), and (3x3, 3x3, 3x3), respectively; the channel numbers are 16, (16, 16), and (16, 24, 32), respectively; the output results of the 3 convolutional branches are concatenated by dimension, then passed through a convolutional layer with a convolutional kernel size of 1x1 and a channel number of 96, and the output result is element-wise added to the direct connection branch and sent to the PReLU layer of the denoising network; a BN layer is added after each convolutional layer of the MSFEA module, and the BN layer is activated by a ReLU non-linear function.

[0019] Preferably, the MSFEB module divides the input into 3 paths for multi-scale feature extraction, including 2 convolutional branches and 1 direct connection branch; the convolutional kernel sizes of the 2 convolutional branches are (1x1) and (1x1, 1x7, 7x1), respectively; the channel numbers are 96, (96, 80, 64), respectively; the output results of the 2 convolutional branches are concatenated by dimension, then passed through a convolutional layer with a convolutional kernel size of 1x1 and a channel number of 96, and the output result is element-wise added to the direct connection branch and sent to the PReLU layer of the denoising network; a BN layer is added after each convolutional layer of the MSFEB module, and the BN layer is activated by a ReLU non-linear function.

[0020] Preferably, the MSE loss is:

[0021]

[0022] Where, (1) T(·) represents the denoising network, X represents the LDCT image, and Y represents the NDCT image;

[0023] The perceptual loss is:

[0024]

[0025] Where, X in (2) denoised represents the denoised image generated by the denoising model, Φ i (·) represents the i-th layer features extracted from the feature extraction network;

[0026] The SSIM is:

[0027]

[0028] Among them, in formula (3) μ Y Represents X denoised and the mean of Y, Represents X denoised and the covariance of Y, and Represents X denoised and the variance of Y, c1, c2 are constants,

[0029] The SSIM loss is:

[0030] L SSIM =1-SSIM(X denoised ,Y) (4)

[0031] Preferably, the hybrid loss function is:

[0032] L total =λ1L MSE +λ2L per +λ3L SSIM (5)

[0033] Among them, λ1, λ2, and λ3 in formula (5) are the weight coefficients of the corresponding loss function.

[0034] The application adds the MSFE module to the residual connection through the multi-scale feature extraction module MSFE, increases the information utilization rate of the input LDCT image, uses zero padding for all convolution layers to ensure that the input and output image sizes are the same, alleviates the loss of image structure information caused by continuous down-sampling, uses the BN layer to accelerate the convergence speed and alleviate the network overfitting problem, and introduces a hybrid loss function: the MSE loss, the SSIM loss and the perception loss to further improve the visual effect of the denoised image. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 It is a whole framework diagram of the denoising network of the application.

[0036] Figure 2 It is a schematic diagram of the MSFEA module.

[0037] Figure 3 It is a schematic diagram of the MSFEB module.

[0038] Figure 4 It is a down-sampling process.

[0039] Figure 5 It is the denoising result of several methods on representative slices of the AAPM 1 / 10 chest dataset Figure 1 ;

[0040] Figure 6 It is the denoising result of several methods on representative slices of the AAPM 1 / 10 chest dataset Figure 2 ;

[0041] Figure 7 It is Figure 5 an enlarged view of the region of interest.

[0042] Figure 8 It is Figure 6 an enlarged view of the region of interest.

[0043] Figure 9 It is the denoising result of several methods on representative slices of the AAPM 1 / 4 abdominal dataset Figure 1 ;

[0044] Figure 10 It is the denoising result of several methods on representative slices of the AAPM 1 / 4 abdominal dataset Figure 2 ;

[0045] Figure 11 It is Figure 9 an enlarged view of the region of interest.

[0046] Figure 12 It is Figure 10 an enlarged view of the region of interest. DETAILED DESCRIPTION

[0047] The application is further described below in combination with the drawings and specific embodiments.

[0048] As Figure 1 shown, the application is based on a low-dose CT image denoising network based on residual multi-scale feature extraction, which adopts an encoder-decoder framework, including:

[0049] A multi-scale feature extraction module is added to the residual connection of the encoder and decoder convolutional layer;

[0050] Specifically, the multi-scale feature extraction module is a lightweight multi-scale feature extraction module, denoted as MSFE module; adding the MSFE module to the residual connection increases the information utilization rate of the input LDCT image, while obtaining higher performance indicators, it also improves the operation speed;

[0051] Specifically, the denoising network is implemented using a classic network architecture: an encoder-decoder network, the encoder is a compression process for the input image, and the decoder is a decompression process for the input image;

[0052] The encoder is composed of 6 convolutional layers, which are 2 layers of 5x5 shallow feature extraction layers and 4 layers of 3x3 deep feature extraction layers; the decoder is composed of 6 deconvolutional layers, which are 4 layers of 3x3 deconvolutional layers and 2 layers of 5x5 deconvolutional layers; a BN layer is added after each convolutional layer to alleviate the overfitting phenomenon that occurs as the model complexity increases, the encoder part is activated by a ReLU function, and the decoder part is activated by a PReLU function, so that the gradient of the negative part can also be updated with the training of the network; in this denoising network, the channel number of all convolutional kernels is set to 96 except for the MSFEA and MSFEB modules;

[0053] In this embodiment, zero padding is used for the convolutional layers to ensure the consistency of the input and output image sizes;

[0054] The encoder is composed of consecutive down-sampling blocks, and correspondingly, the decoder is composed of consecutive up-sampling blocks. Due to consecutive down-sampling, the loss of image structure information is inevitable, which causes distortion of the final denoised image. Zero padding is used to ensure that the input and output image sizes are the same, extract more detailed information, and alleviate the loss of image structure information caused by consecutive down-sampling;

[0055] As Figure 4 shown, the process of ordinary down-sampling and down-sampling using zero padding is shown. Let the input image be HxW, from Figure 4(a) As can be seen, after the ordinary downsampling operation, the output image size is (H-2) x (W-2), and the reduction of the image size will inevitably cause the loss of detail information, especially after the continuous downsampling operation, the loss of detail information is more obvious, from Figure 4 (b) As can be seen, the output image size and the input image are consistent, thereby reducing the loss of detail information.

[0056] In this embodiment, as shown in the figure, Figure 1 The residual connection includes a first residual connection, a second residual connection and a third residual connection; the first residual connection is connected from the original input to the BN layer after the last layer of the decoder; the second residual connection is connected from the BN after the second layer of the encoder to the BN after the second to last layer of the decoder; and the third residual connection is connected from the BN after the fourth layer of the encoder to the BN after the fourth to last layer of the decoder.

[0057] In this embodiment, the MSFE module includes an MSFEA module and an MSFEB module; the MSFEA module is added to the second residual connection; and the MSFEB module is added to the third residual connection.

[0058] As shown in the figure, Figure 2 The MSFEA module divides the input into 4 paths for multi-scale feature extraction, including 3 convolution branches and 1 direct connection branch, the convolution kernel sizes of the 3 convolution branches are (1x1), (1x1, 3x3) and (3x3, 3x3, 3x3), and the channel numbers are 16, (16, 16) and (16, 24, 32) respectively, then the output results of the 3 convolution branches are dimensionally spliced and passed through a convolution layer with a convolution kernel size of 1x1 and a channel number of 96, the output result is elementarily added to the direct connection branch, and finally, it is sent to the PReLU layer of the main denoising network; as shown in the figure, Figure 3 The MSFEB module divides the input into 3 paths for multi-scale feature extraction, including 2 convolution branches and 1 direct connection branch, the convolution kernel sizes of the 2 convolution branches are (1x1) and (1x1, 1x7, 7x1), and the channel numbers are 96, (96, 80, 64) respectively. Then the output results of the 2 convolution branches are dimensionally spliced and passed through a convolution layer with a convolution kernel size of 1x1 and a channel number of 96, the output result is elementarily added to the direct connection branch, and finally, it is sent to the PReLU layer of the main denoising network; a BN layer is added after each convolution layer of the MSFEA module and the MSFEB module to prevent network overfitting, and a ReLU nonlinear activation is used.

[0059] The traditional CNN structure is a single size convolution kernel for feature extraction, which is not diverse enough in the scale of feature extraction, which often leads to low utilization of input image information and poor denoising performance. The MSFEA module divides the input into four paths for multi-scale feature extraction, including three convolution branches and one direct branch. The MSFEB module divides the input into three paths for multi-scale feature extraction, including two convolution branches and one direct branch, which increases the width of the network and also increases the adaptability of the network to the scale. Different receptive fields of different branches can also combine more nonlinear features, and convolution can be performed on multiple scales to extract features of different scales, which are more abundant and have higher information utilization rate.

[0060] The hybrid loss function composed of the weighted sum of MSE loss, SSIM loss and perceptual loss is used to guide the denoising network, which alleviates the problem of over-smoothing of the denoised image caused by the single MSE loss guiding the denoising network.

[0061] The MSE loss in the context of LDCT image denoising is the mean of the square sum of the denoising difference between the denoised image and the NDCT image, and its mathematical expression is:

[0062]

[0063] In formula (1), T(·) represents the denoising network, X represents the LDCT image, and Y represents the NDCT image.

[0064] The perceptual loss in the context of LDCT image denoising represents the comparison of the features obtained by convolving the denoised image and the features obtained by convolving the NDCT image, so that the content and global context structure are similar, and the denoising network generates a denoised image that is closer and closer to the NDCT. Its process can be represented by formula 2:

[0065]

[0066] In formula (2), X denoised represents the denoised image generated by the denoising model, and Φ i (·) represents the i-th layer feature extracted from the feature extraction network. The feature extraction network used in this embodiment is a VGG19 network.

[0067] The SSIM is an index for measuring the similarity between two images. The higher the SSIM, the higher the similarity between the two images, and vice versa. It can be represented as:

[0068]

[0069] In formula (3), X μ Y denote the mean of X denoised and Y, denote the mean of X denoised and Y, and denote the variance of X denoised and Y, c1, c2 are constants,

[0070] The SSIM loss is:

[0071] L SSIM = 1-SSIM(X denoised , Y) (4)

[0072] Therefore, the hybrid loss function is:

[0073] L total = λ1L MSE + λ2L per + λ3L SSIM (5)

[0074] Wherein, λ1, λ2, λ3 in formula (5) are weight coefficients of the corresponding loss function.

[0075] The present application has the following test experiments,

[0076] The denoising network of the present application is named MSFREDCNN, and is compared with several related methods, including REDCNN-RAM, WGAN-RAM and REDCNN. The denoising network of the present application selects the Adam optimizer to optimize the denoising model, the Patch_size is set to 64x64, the sliding interval is set to 10, and the initial learning rate is set to 1e-4. The learning rate is reduced by half every 2000 iteration steps, the total iteration number is set to 200, and the hyperparameters of the hybrid loss function are finally set to: λ1=1, λ2=0.02, λ3=0.01,

[0077] In order to compare fairly, all experiments are trained in the Pytorch1.11 environment, using NIVID RTX3080-10GB graphics accelerator,

[0078] The performance of the present application method is evaluated by performing experiments on the 1 / 10 chest CT data set and the 1 / 4 abdominal CT data set in the "2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge" authorized by the Mayo Clinic, and the results are shown in Figures 5-8 . Figure 5 and Figure 6 are the denoising results of several methods on two representative slices of the AAPM 1 / 10 chest data set, from Figure 5 (a) and Figure 6As can be seen from (a), the unprocessed LDCT image does indeed have severe noise and artifacts, making it difficult to distinguish the structural details of the LDCT and causing difficulties for clinicians in diagnosis. When compared with the LDCT image, all the algorithms can effectively suppress the noise and artifacts, and when compared with the NDCT image, the MSFRED CNN has the best structural details and visual effect. Figure 5 As can be seen from (c) and Figure 6 As can be seen from (c), although the RED CNN-RAM suppresses the noise to some extent, it introduces new noise and the image is blurred to some extent; from Figure 5 As can be seen from (d) and Figure 6 As can be seen from (d), although the WGAN-RAM generates a denoised image with relatively high similarity to human perception, the detail information recovery is still not obvious; from Figure 5 As can be seen from (e) and Figure 6 As can be seen from (e), although the RED CNN achieves good denoising results, since the MSE loss function is used to guide the denoising network, the denoised image is over-smoothed, still causing difficulties for clinical diagnosis. From Figure 5 As can be seen from (f) and Figure 6 As can be seen from (f), the MSFRED CNN restores the most image details and has the best visual effect. In order to further improve the visual effect of the denoised image and generate a denoised image with high similarity to human perception, the present application introduces a perception loss and a structural similarity loss to jointly guide the entire denoising network with the MSE, from Figure 5 As can be seen from (g) and Figure 6 As can be seen from (g), the denoising network proposed in the present application achieves the denoised image closest to the NDCT image under the joint guidance of the MSE loss, the perception loss and the structural similarity loss. Figure 7 and Figure 8 are Figure 5 As can be seen from (b) and Figure 6 As can be seen from the white circle and the white arrow in the enlarged image of the denoised region of interest (ROI) marked by the white box in (b), the MSFRED CNN is closest to the NDCT image in terms of structural similarity and visual effect.

[0079] Figure 9 and Figure 10 are the denoising results of several methods on representative slices of two AAPM 1 / 4 abdominal data sets, from Figure 9 As can be seen from (a) and Figure 10 As can be seen from (a), although the 1 / 4 dose abdominal CT image has better visual effect than the 1 / 10 dose chest CT image, it still has certain noise and artifacts, affecting the diagnosis results of doctors. From Figure 9 and Figure 10Overall, these methods can suppress noise to some extent and improve visual effects. However, compared with the NDCT image with clear lesions, it can be seen that WGAN-RAM and REDCNN still retain some noise; since REDCNN-RAM and the proposed MSFREDCNN use the MSE objective function to guide the denoising network, there is a certain degree of blurring; when the proposed hybrid loss function is used to guide the MSFREDCNN, the recovered denoised image is visually closest to the NDCT. By observing the enlarged ROI region, such as the white circle marked in Figure 11 and Figure 12 , it is not difficult to find that the proposed MSFREDCNN is visually closer to the human perception level. In terms of recovering small structures, such as the region marked by the white arrow in Figure 11 , the recovered structure by MSFREDCNN and MSFREDCNN+hybrid loss is clearer.

[0080] Objective evaluation of image denoising effect is performed by using peak-signal-to-noise ratio (PSNR), structural similarity (SSIM) and root-mean square error (RMSE). Table 1 and Table 2 are the average values of quantitative indicators on the 1 / 10 dose chest CT data set and the 1 / 4 dose abdominal CT data set, respectively.

[0081] Table 1 Average values of quantitative indicators of different algorithms on 1 / 10 chest CT data set

[0082]

[0083] Table 2 Average values of quantitative indicators of different algorithms on 1 / 4 abdominal CT data set

[0084]

[0085] From Table 1 and Table 2, it can be seen that compared with the two algorithms of REDCNN-RAM and WGAN-RAM, the proposed denoising network MSFREDCNN improves the running time by nearly 3 times, and the hybrid loss proposed in the application is used to guide MSFREDCNN, which improves the running time by nearly 2 times. And from the quantitative indicators, the proposed MSFREDCNN obtains the best quantitative indicator value (indicated by bold in the table), indicating that the MSFREDCNN network model has better denoising performance in maintaining structure and suppressing noise. Although the quantitative indicator value does not exceed that without adding the hybrid loss function, due to the particularity of the LDCT image: the background with the same LDCT and NDCT occupies a large part of the CT image, and this part often does not work for the doctor's diagnosis. Therefore, the quantitative indicator value of the ROI region is given in the ablation experiment part, and it can be seen that the denoising network with the added hybrid loss achieves the best performance indicator. At the same time, this result further confirms the previous visual comparison.

[0086] Since the MSFREDCNN denoising network contains multiple modules, ablation research is needed to verify the effectiveness of each module, and ablation research is conducted on the proposed model from two aspects of network structure and objective function, assuming:

[0087] BL: baseline model (representing a rough network without any enhancement module);

[0088] PA+BN: add BN layer and zero padding to the baseline model;

[0089] MSFE: add MSFEA and MSFEB multi-scale feature extraction modules to the BL+PA+BN model.

[0090] MSE: use MSE to guide the proposed MSFREDCNN network

[0091] MSE+VGG: use MSE+VGG to guide the proposed MSFREDCNN network

[0092] MSE+SSIM: use MSE+SSIM to guide the proposed MSFREDCNN network

[0093] MSE+SSIM+VGG: use MSE+SSIM+VGG to guide the proposed MSFREDCNN denoising network.

[0094] Ablation research of network structure, each time the denoising network adds an enhancement module to the baseline model, and the quantitative results are shown in Table 3. It can be observed that with the addition of each module, the denoising performance of the model gradually improves.

[0095] Ablation study of objective function, the proposed de-noising network also verifies the influence of different objective functions on the proposed de-noising model. Due to the particularity of CT images, the background and bone regions with the same LDCT and NDCT occupy a large part of the CT image, and this part often does not play a role for the doctor's diagnosis. Therefore, the present application selects the lesion region of interest of the representative slice to calculate the quantitative index, and the quantitative results are shown in Table 4.

[0096] Table 3 Network structure ablation experiment

[0097]

[0098] Table 4 Objective function ablation experiment

[0099]

Claims

1. Low-dose CT image denoising network based on residual multi-scale feature extraction, characterized by: Adopting the encoder-decoder framework, including: Add a multi-scale feature extraction module to the residual connections of the encoder and decoder convolutional layers; A hybrid loss function consisting of MSE loss, SSIM loss, and perceptual loss is used to guide the denoising network; The encoder consists of 6 convolutional layers, which are 2 5×5 shallow feature extraction layers and 4 3×3 deep feature extraction layers; the decoder consists of 6 deconvolutional layers, which are 4 3×3 deconvolutional layers and 2 5×5 deconvolutional layers. A BN layer is added after each convolutional layer; the convolutional layer of the encoder is activated by the ReLU function; the deconvolution layer of the decoder is activated by the PReLU function; The residual connection includes a first residual connection, a second residual connection and a third residual connection; The first residual connection is connected from the original input to the BN layer of the last layer of the decoder; The second residual connection is connected from the second BN layer of the encoder to the BN layer of the penultimate layer of the decoder; The third residual connection is connected from the encoder after the fourth layer BN to the decoder after the fourth layer BN.

2. The low-dose CT image denoising network based on residual multi-scale feature extraction according to claim 1 is characterized in that: The convolutional layers all use zero padding to ensure that the input and output image sizes are consistent.

3. The low-dose CT image denoising network based on residual multi-scale feature extraction according to claim 1, characterized in that: The multi-scale feature extraction module is denoted as an MSFE module; the MSFE module includes an MSFEA module and an MSFEB module; the MSFEA module is added to the second residual connection; and the MSFEB module is added to the third residual connection.

4. The low-dose CT image denoising network based on residual multi-scale feature extraction according to claim 3, characterized in that: The MSFEA module divides the input into four paths for multi-scale feature extraction, namely three convolution branches and one direct connection branch. The convolution kernel sizes of the three convolution branches are (1×1), (1×1, 3×3), (3×3, 3×3, 3×3), and the number of channels are 16, (16, 16), (16, 24, 32), respectively. After the output results of the three convolution branches are spliced ​​by dimension, they are passed through a convolution layer with a convolution kernel size of 1×1 and 96 channels, and the output results are element-wise added with the direct connection branch and sent to the PReLU layer of the denoising network; a BN layer is added after each convolution layer of the MSFEA module, and is activated by the ReLU nonlinearity.

5. The low-dose CT image denoising network based on residual multi-scale feature extraction according to claim 3, characterized in that: The MSFEB module divides the input into three paths for multi-scale feature extraction, namely two convolution branches and one direct connection branch. The convolution kernel sizes of the two convolution branches are (1×1), (1×1, 1×7, 7×1), and the number of channels are 96 and (96, 80, 64), respectively. After the output results of the two convolution branches are spliced ​​by dimension, they are passed through a convolution layer with a convolution kernel size of 1×1 and 96 channels, and the output results are element-wise added with the direct connection branch and sent to the PReLU layer of the denoising network; a BN layer is added after each convolution layer of the MSFEB module, and is activated by the ReLU nonlinearity.

6. The low-dose CT image denoising network based on residual multi-scale feature extraction according to claim 1, characterized in that: The MSE loss is: （1） Among them, (1) represents the denoising network, X represents the LDCT image, and Y represents the NDCT image; The perceptual loss is: （2） Among them, (2) represents the denoised image generated by the denoising model, represents the i-th layer feature extracted from the feature extraction network; The SSIM is: （3） Among them, (3) , Respectively and The average value of express and The covariance of and Respectively and The variance of , is a constant, The SSIM loss is: （4）。 7. The low-dose CT image denoising network based on residual multi-scale feature extraction according to claim 6, characterized in that: The hybrid loss function is: （5） Among them, in formula (5), , , is the weight coefficient of the corresponding loss function.

Images