CN114219719A - CNN medical CT image denoising method based on dual attention and multi-scale features - Google Patents
CNN medical CT image denoising method based on dual attention and multi-scale features Download PDFInfo
- Publication number
- CN114219719A CN114219719A CN202111254289.6A CN202111254289A CN114219719A CN 114219719 A CN114219719 A CN 114219719A CN 202111254289 A CN202111254289 A CN 202111254289A CN 114219719 A CN114219719 A CN 114219719A
- Authority
- CN
- China
- Prior art keywords
- noise
- network
- layer
- image
- channels
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 21
- 230000009977 dual effect Effects 0.000 title claims description 11
- 238000012549 training Methods 0.000 claims abstract description 22
- 238000000605 extraction Methods 0.000 claims abstract description 12
- 238000007781 pre-processing Methods 0.000 claims abstract description 11
- 238000005070 sampling Methods 0.000 claims description 21
- 238000009826 distribution Methods 0.000 claims description 12
- 230000004044 response Effects 0.000 claims description 12
- 230000004927 fusion Effects 0.000 claims description 10
- 239000011159 matrix material Substances 0.000 claims description 10
- 238000010586 diagram Methods 0.000 claims description 9
- 230000006870 function Effects 0.000 claims description 9
- 238000010606 normalization Methods 0.000 claims description 6
- 238000012795 verification Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 5
- 230000006835 compression Effects 0.000 claims description 4
- 238000007906 compression Methods 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 230000000903 blocking effect Effects 0.000 claims description 3
- 238000005520 cutting process Methods 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 238000012360 testing method Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 8
- 230000007246 mechanism Effects 0.000 abstract description 4
- 238000002591 computed tomography Methods 0.000 description 17
- 238000013527 convolutional neural network Methods 0.000 description 6
- 238000002059 diagnostic imaging Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 229940050561 matrix product Drugs 0.000 description 2
- 238000004904 shortening Methods 0.000 description 2
- 238000013519 translation Methods 0.000 description 2
- 238000005481 NMR spectroscopy Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000002560 therapeutic procedure Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H30/00—ICT specially adapted for the handling or processing of medical images
- G16H30/20—ICT specially adapted for the handling or processing of medical images for handling medical images, e.g. DICOM, HL7 or PACS
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Epidemiology (AREA)
- Medical Informatics (AREA)
- Primary Health Care (AREA)
- Public Health (AREA)
- Image Analysis (AREA)
Abstract
The CNN medical CT image denoising algorithm based on double attention and multi-scale features specifically comprises the following steps: step 1) creating a medical CT image model; step 2) constructing a noise prior information extraction network; step 3), constructing a denoising network; step 4), preprocessing a data set; step 5), training a denoising network and updating parameters; the invention has the following advantages: 1) the method uses a double attention mechanism to be fused into a noise prior network to denoise the medical CT image, and enhances the noise extraction effect. 2) The multi-scale context module is used in the denoising network, so that the detail texture retaining effect of the network on the image is enhanced.
Description
Technical Field
The invention relates to the field of medical image denoising, in particular to a CNN medical CT image denoising method.
Technical Field
In the medical field, medical image processing techniques are increasingly used in therapy planning and disease diagnosis. The medical imaging techniques widely used in clinical and diagnostic applications are mainly nuclear Magnetic Resonance (MRI), Computed Tomography (CT), and the like. Although the medical imaging equipment and the image acquisition equipment are very advanced, the medical imaging equipment and the image acquisition equipment are influenced by objective factors, noise is inevitably generated in the medical image, and the noise can directly influence the quality of the medical image so as to influence the judgment of medical staff. Therefore, it is necessary to analyze and study the denoising technique in the medical image preprocessing. Denoising is the basis and precondition of image processing and is also an indispensable step in image preprocessing. The denoising technology for natural image preprocessing achieves obvious effect at present, but the research on the denoising technology in the medical image preprocessing has some problems.
Many early model-based methods found prior information for images and then applied optimization algorithms to iteratively solve the models, which is time consuming and inefficient. With the rise of deep learning, CNN is widely applied to denoising and achieves good effects. Today, various methods have been used to achieve good denoising of AWGN images, but it is difficult to perform denoising of real world images because the noise sources in the camera system are various (such as dark current noise, short noise, thermal noise, etc.), and further go through ISP processes (including demosaicing, Gamma correction, compression, etc.), so that the noise of real world images is much more complicated than gaussian noise. In addition, the noise level of the real world noise image is unknown, and if the non-blind AWGN denoising method is directly used, details can be easily eliminated or the image is too smooth, so that the generalization capability is poor.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a CNN medical CT image denoising method based on double attention and multi-scale characteristics
The invention aims to improve the denoising effect of a medical CT image, in the traditional denoising method based on a noise prior medical image, a single noise level image or multiple noise level images are often directly mixed to serve as training data, and a conventional cnn network is used for extracting noise information, so that the obtained denoising model can only process noise images in a limited range, the neural network model cannot be fully generalized to the images with wider noise levels, and the neural network cannot fully acquire image information during model training. In order to improve the denoising capability of the network, the invention adds a space attention mechanism and a channel attention mechanism based on double attention, namely parallel connection, into a noise extraction network, thereby realizing a noise prior network with stronger noise extraction capability. And connecting the extracted data and the noise image in series as the input of the denoising network, and optimizing parameters in the neural network by using the extracted information and the noise image.
The innovation and the advantages of the invention are as follows: the invention integrates a new double attention unit on the basis of the original noise prior network, and achieves better effect in the aspect of prior noise extraction. The generalization capability of denoising the real world noise image is enhanced.
In order to make the objects, technical solutions and advantages of the present invention clearer, the following is a detailed description of the technical solution of the present invention, and a CNN medical CT image denoising method based on dual attention and multi-scale features includes the following specific steps:
step 1) creating a medical CT image model:
creating a medical CT image model:
adopting a Gaussian noise model, wherein the mathematical expression is as follows:
Y=X+V (1)
wherein X is a clean image without noise, Y is an actual image with noise, and V is noise; the noise distribution of V obeys Gaussian distribution, the Gaussian noise is a type of noise with a probability density function obeying Gaussian distribution (namely normal distribution), namely a Gaussian random variable z probability density function, and the mathematical expression of the Gaussian noise is as follows:
where μ is expressed as a mathematical expectation and σ is expressed as a standard deviation;
step 2) constructing a noise prior information extraction network:
this stage focuses on extracting features from noisy images, using a simple five-layer full convolution sub-network, without posing and batch normalization, followed by ReLU activation after each Conv layer. At each convolutional layer, the eigenchannel is set to 32 (the last layer is set to 1), and the convolutional kernel size is 3 × 3.
Before the last layer, a Dual attribute module is inserted to effectively capture the global dependency of the feature. The Dual attribute module consists of a space attention module and a channel attention module which are connected in parallel.
The spatial attention module aims to mutually enhance the expression of respective features by utilizing the association between any two features. Specifically, firstly, a correlation strength matrix between any two point features is calculated, namely, the original feature A is subjected to convolution dimensionality reduction to obtain a feature B and a feature C, then the feature dimensionality B is changed to ((HxW) xC ') and C is changed to (C' x (HxW)), and then the matrix product is carried out to obtain a correlation strength matrix ((HxW) x (HxW)) between any two point features. And then obtaining an attention graph S of each position to other positions through the normalization of a softmax operation, wherein the more similar two-point characteristics have larger response values. The response values in the attention map are then used as weights to perform weighted fusion on the feature D, so that for each position point, similar features are fused in the global space through the attention map.
The channel attention module aims to enhance specific semantic response capabilities under the channels by modeling associations between the channels. The specific process is similar to the position attention module, except that when the feature attention diagram X is obtained, dimension transformation and matrix multiplication are carried out on any two channel features to obtain the correlation strength of any two channels, and then the attribute diagram between the channels is obtained through the softmax operation. And finally, fusing by weighting the attention images among the channels, so that global association can be generated among the channels, and characteristics of stronger semantic response are obtained.
Step 3), constructing a denoising network:
the denoising network main body is of a UNet structure with hopping connection and is divided into an encoder-decoder left part and an encoder right part, and the two parts are connected by a context block module with 4 cavity convolutions with different expansion rates:
encoder part: the left half part is a down-sampling layer and consists of 4 convolution layers with the same structure. Each convolution layer consists of two conv + ReLU layers, the convolution kernel is set to be 3 multiplied by 3, the step length is 1, and padding is 1; the number of convolution channels of the first layer is 128, the number of convolution channels of the second layer is 256, the number of convolution channels of the third layer is 512, and the number of convolution channels of the fourth layer is 1024; the 2 x 2-sized maxpool layer connections are used for each downsampling, and the number of channels is doubled for each downsampling.
Context block module removal of BN layer, use of only the translation rates set to 1, 2, 3, 4, to further simplify the operation shortening time, first compress the feature channels using 1 × 1 Conv.
The compression ratio is set to 4, 1 × 1 convolution is used in the fusion (fusion) section, and similarly, local skip connection is used between the input and output features to prevent information blocking.
Decoder part: the right half part is an up-sampling layer which is symmetrical to the encoder part and consists of 4 up-sampling layers with the same structure. Each up-sampling convolution layer is the same as the left half part and consists of two conv + ReLU layers, the convolution kernel is set to be 3 multiplied by 3, the step length is 1, and the padding is 1; and performing characteristic up-sampling on each layer through 2 multiplied by 2 sized up-conv layers, wherein the number of channels is halved when each up-sampling is performed.
Jump connection is used between the encoder layers of the network, the feature map output by each encoder layer is added with the up-sampling result of the corresponding decoder layer, and the network is strengthened to restore the image detail texture to ensure the convergence of network training.
Step 4) data set preprocessing:
in the preprocessing stage of the data set, the data set is divided into a training set, a verification set and a test set. Cutting the pictures of the training set and the verification set to enable the sizes of the pictures to be uniform, replacing noise in the medical CT image with white Gaussian noise, and adding the white Gaussian noise with the mean value of 0 and the standard deviation of 5, 10, 15, 20, 25, 30, 35, 40, 45 and 50 to the original clean image to obtain the denoising network training data.
Step 5), training a denoising network and updating parameters:
the noise prior network firstly initializes the weight randomly, then carries out feature extraction on input noise image data x to obtain noise estimation data noise _ est, and finally splices the original input x and the noise _ est to be used as the input of the denoising network. The denoising network initializes the weight immediately and obtains output according to the input training network of the noise prior network. In the loss function, an original picture, a noisy picture, noise estimation and a real noise level are used as input, the mean variance is used for calculating the total loss, and finally an Adam optimizer is used for updating the network weight parameters.
Further, in the step (2), in order to further obtain the features of the global dependency relationship, the output results of the two modules are added and fused to obtain the final features for classifying the pixels.
The network model realizes blind denoising processing of the noise picture without training for a certain noise level picture.
The invention has the following advantages:
1. the method uses a double attention mechanism to be fused into a noise prior network to denoise the medical CT image, and enhances the noise extraction effect.
2. The multi-scale context module is used in the denoising network, so that the detail texture retaining effect of the network on the image is enhanced.
Drawings
FIG. 1 is a schematic illustration of a medical CT image containing Gaussian noise;
FIG. 2 is a schematic diagram of the overall network architecture of the present invention;
FIG. 3 is a schematic diagram of the Dual Attention module of the present invention;
FIG. 4 is a diagram of the Context Block module of the present invention;
FIG. 5 is a schematic diagram of denoising a medical CT image according to the present invention
The specific implementation mode is as follows:
the invention is explained in detail below with reference to the drawings
The CNN medical CT image denoising algorithm based on double attention and multi-scale features specifically comprises the following steps:
step 1) creating a medical CT image model:
creating a medical CT image model:
adopting a Gaussian noise model, wherein the mathematical expression is as follows:
Y=X+V (1)
wherein X is a clean image without noise, Y is an actual image with noise, and V is noise; the noise distribution of V obeys Gaussian distribution, the Gaussian noise is a type of noise with a probability density function obeying Gaussian distribution (namely normal distribution), namely a Gaussian random variable z probability density function, and the mathematical expression of the Gaussian noise is as follows:
where μ is expressed as a mathematical expectation and σ is expressed as a standard deviation;
step 2) constructing a noise prior information extraction network:
this stage focuses on extracting features from noisy images, using a simple five-layer full convolution sub-network, without posing and batch normalization, followed by ReLU activation after each Conv layer. At each convolutional layer, the eigenchannel is set to 32 (the last layer is set to 1), and the convolutional kernel size is 3 × 3.
Before the last layer, a Dual attribute module is inserted to effectively capture the global dependency of the feature. The Dual attribute module consists of a space attention module and a channel attention module which are connected in parallel.
The spatial attention module aims to mutually enhance the expression of respective features by utilizing the association between any two features. Specifically, firstly, a correlation strength matrix between any two point features is calculated, namely, the original feature A is subjected to convolution dimensionality reduction to obtain a feature B and a feature C, then the feature dimensionality B is changed to ((HxW) xC ') and C is changed to (C' x (HxW)), and then the matrix product is carried out to obtain a correlation strength matrix ((HxW) x (HxW)) between any two point features. And then obtaining an attention graph S of each position to other positions through the normalization of a softmax operation, wherein the more similar two-point characteristics have larger response values. The response values in the attention map are then used as weights to perform weighted fusion on the feature D, so that for each position point, similar features are fused in the global space through the attention map.
The channel attention module aims to enhance specific semantic response capabilities under the channels by modeling associations between the channels. The specific process is similar to the position attention module, except that when the feature attention diagram X is obtained, dimension transformation and matrix multiplication are carried out on any two channel features to obtain the correlation strength of any two channels, and then the attribute diagram between the channels is obtained through the softmax operation. And finally, fusing by weighting the attention images among the channels, so that global association can be generated among the channels, and characteristics of stronger semantic response are obtained.
In order to further obtain the characteristics of the global dependency relationship, the output results of the two modules are added and fused to obtain the final characteristics for classifying the pixel points.
Step 3), constructing a denoising network:
the denoising network main body is of a UNet structure with hopping connection and is divided into an encoder-decoder left part and an encoder right part, and the two parts are connected by a context block module with 4 cavity convolutions with different expansion rates:
encoder part: the left half part is a down-sampling layer and consists of 4 convolution layers with the same structure. Each convolution layer consists of two conv + ReLU layers, the convolution kernel is set to be 3 multiplied by 3, the step length is 1, and padding is 1; the number of convolution channels of the first layer is 128, the number of convolution channels of the second layer is 256, the number of convolution channels of the third layer is 512, and the number of convolution channels of the fourth layer is 1024; the 2 x 2-sized maxpool layer connections are used for each downsampling, and the number of channels is doubled for each downsampling.
Context block module removal of BN layer, use of only the translation rates set to 1, 2, 3, 4, to further simplify the operation shortening time, first compress the feature channels using 1 × 1 Conv.
The compression ratio is set to 4, 1 × 1 convolution is used in the fusion (fusion) section, and similarly, local skip connection is used between the input and output features to prevent information blocking.
Decoder part: the right half part is an up-sampling layer which is symmetrical to the encoder part and consists of 4 up-sampling layers with the same structure. Each up-sampling convolution layer is the same as the left half part and consists of two conv + ReLU layers, the convolution kernel is set to be 3 multiplied by 3, the step length is 1, and the padding is 1; and performing characteristic up-sampling on each layer through 2 multiplied by 2 sized up-conv layers, wherein the number of channels is halved when each up-sampling is performed.
Jump connection is used between the encoder layers of the network, the feature map output by each encoder layer is added with the up-sampling result of the corresponding decoder layer, and the network is strengthened to restore the image detail texture to ensure the convergence of network training.
Step 4) data set preprocessing:
in the preprocessing stage of the data set, the data set is divided into a training set, a verification set and a test set. Cutting the pictures of the training set and the verification set to enable the sizes of the pictures to be uniform, replacing noise in the medical CT image with white Gaussian noise, and adding the white Gaussian noise with the mean value of 0 and the standard deviation of 5, 10, 15, 20, 25, 30, 35, 40, 45 and 50 to the original clean image to obtain the denoising network training data.
Step 5), training a denoising network and updating parameters:
the noise prior network firstly initializes the weight randomly, then carries out feature extraction on input noise image data x to obtain noise estimation data noise _ est, and finally splices the original input x and the noise _ est to be used as the input of the denoising network. The denoising network initializes the weight immediately and obtains output according to the input training network of the noise prior network. In the loss function, an original picture, a noisy picture, noise estimation and a real noise level are used as input, the mean variance is used for calculating the total loss, and finally an Adam optimizer is used for updating the network weight parameters.
The embodiments described in this specification are merely illustrative of implementations of the inventive concept and the scope of the present invention should not be considered limited to the specific forms set forth in the embodiments but rather by the equivalents thereof as may occur to those skilled in the art upon consideration of the present inventive concept.
Claims (2)
1. The CNN medical CT image denoising method based on double attention and multi-scale features comprises the following specific steps:
step 1) creating a medical CT image model:
creating a medical CT image model:
adopting a Gaussian noise model, wherein the mathematical expression is as follows:
Y=X+V (1)
wherein X is a clean image without noise, Y is an actual image with noise, and V is noise; the noise distribution of V obeys Gaussian distribution, the Gaussian noise is a type of noise with a probability density function obeying Gaussian distribution (namely normal distribution), namely a Gaussian random variable z probability density function, and the mathematical expression of the Gaussian noise is as follows:
where μ is expressed as a mathematical expectation and σ is expressed as a standard deviation;
step 2) constructing a noise prior information extraction network:
this stage focuses on extracting features from noisy images, using a simple five-layer full convolution sub-network, without posing and batch normalization, with ReLU activation after each Conv layer; at each convolutional layer, the eigenchannel is set to 32 (the last layer is set to 1), the convolutional kernel size is 3 × 3;
before the last layer, inserting a Dual attribute module to effectively capture the global dependency of the features; the Dual attribute module consists of a space attention module and a channel attention module which are connected in parallel;
the spatial attention module aims to mutually enhance the expression of the respective features by utilizing the association between any two features; specifically, firstly, calculating a correlation strength matrix between any two point features, namely, obtaining a feature B and a feature C by convolution dimension reduction of an original feature A, changing the feature dimension B to ((HxW) xC ') and the feature dimension C to (C' x (HxW)), and then obtaining a correlation strength matrix ((HxW) x (HxW)) between any two point features by matrix multiplication; then obtaining an attention graph S of each position to other positions through softmax operation normalization, wherein the more similar two-point characteristics are, the larger the response value is; then, taking the response value in the attention image as a weight to perform weighted fusion on the feature D, so that for each point of each position, similar features are fused in the global space through the attention image;
the channel attention module aims to enhance the specific semantic response capability under the channels by modeling the association between the channels; the specific process is similar to the position attention module, except that when the feature attention diagram X is obtained, dimension transformation and matrix multiplication are carried out on any two channel features to obtain the correlation strength of any two channels, and then the attention graph among the channels is obtained through softmax operation; finally, fusion is carried out through attention graph weighting among the channels, so that global association can be generated among the channels, and characteristics of stronger semantic response are obtained;
step 3), constructing a denoising network:
the denoising network main body is of a UNet structure with hopping connection and is divided into an encoder-decoder left part and an encoder right part, and the two parts are connected by a context block module with 4 cavity convolutions with different expansion rates:
encoder part: the left half part is a down-sampling layer which consists of 4 convolution layers with the same structure; each convolution layer consists of two conv + ReLU layers, the convolution kernel is set to be 3 multiplied by 3, the step length is 1, and padding is 1; the number of convolution channels of the first layer is 128, the number of convolution channels of the second layer is 256, the number of convolution channels of the third layer is 512, and the number of convolution channels of the fourth layer is 1024; in each downsampling, 2 x 2-sized maxpool layer connection is used, and the number of channels is doubled after each downsampling;
removing the BN layer, only using the differences set to 1, 2, 3 and 4, and firstly compressing the feature channel by using 1 multiplied by 1Conv in order to further simplify the operation and shorten the time; compression ratio is set to 4, 1 × 1 convolution is used in the fusion (fusion) section, and similarly, local skip connection is used between the input and output features to prevent information blocking;
decoder part: the right half part is an upper sampling layer, is symmetrical to the encoder part and consists of 4 upper sampling layers with the same structure; each up-sampling convolution layer is the same as the left half part and consists of two conv + ReLU layers, the convolution kernel is set to be 3 multiplied by 3, the step length is 1, and the padding is 1; performing characteristic up-sampling on each layer through 2 multiplied by 2 sized up-conv layers, and reducing the number of channels by half once per up-sampling;
jump connection is used between the encoder layers of the network, the feature map output by each encoder layer is added with the up-sampling result of the corresponding decoder layer, and the network is enhanced to restore the image detail texture to ensure the convergence of network training;
step 4) data set preprocessing:
dividing a data set into a training set, a verification set and a test set in a data set preprocessing stage; cutting the pictures of the training set and the verification set to enable the sizes of the pictures to be uniform, replacing noise in the medical CT image with white Gaussian noise, and adding the white Gaussian noise with the mean value of 0 and the standard deviation of 5, 10, 15, 20, 25, 30, 35, 40, 45 and 50 to the original clean image to obtain denoising network training data;
step 5), training a denoising network and updating parameters:
the noise prior network firstly initializes the weight randomly, then carries out feature extraction on input noise image data x to obtain noise estimation data noise _ est, and finally splices the original input x and the noise _ est to be used as the input of the noise removing network; the denoising network initializes the weight immediately and obtains output according to the input training network of the noise prior network; in the loss function, an original picture, a noisy picture, noise estimation and a real noise level are used as input, the mean variance is used for calculating the total loss, and finally an Adam optimizer is used for updating the network weight parameters.
2. The CNN medical CT image denoising method based on dual attention and multi-scale features of claim 1, wherein: in the step (2), in order to further obtain the features of the global dependency relationship, the output results of the two modules are added and fused to obtain the final features for classifying the pixel points.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111254289.6A CN114219719A (en) | 2021-10-27 | 2021-10-27 | CNN medical CT image denoising method based on dual attention and multi-scale features |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111254289.6A CN114219719A (en) | 2021-10-27 | 2021-10-27 | CNN medical CT image denoising method based on dual attention and multi-scale features |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114219719A true CN114219719A (en) | 2022-03-22 |
Family
ID=80696247
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111254289.6A Pending CN114219719A (en) | 2021-10-27 | 2021-10-27 | CNN medical CT image denoising method based on dual attention and multi-scale features |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114219719A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114781445A (en) * | 2022-04-11 | 2022-07-22 | 山东省人工智能研究院 | Deep neural network electrocardiosignal noise reduction method based on interpretability |
CN115409733A (en) * | 2022-09-02 | 2022-11-29 | 山东财经大学 | Low-dose CT image noise reduction method based on image enhancement and diffusion model |
CN116051408A (en) * | 2023-01-06 | 2023-05-02 | 郑州轻工业大学 | Image depth denoising method based on residual error self-coding |
CN116757966A (en) * | 2023-08-17 | 2023-09-15 | 中科方寸知微(南京)科技有限公司 | Image enhancement method and system based on multi-level curvature supervision |
CN117095074A (en) * | 2023-08-23 | 2023-11-21 | 浙江大学 | Medical image double denoising method and device based on deep learning |
-
2021
- 2021-10-27 CN CN202111254289.6A patent/CN114219719A/en active Pending
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114781445A (en) * | 2022-04-11 | 2022-07-22 | 山东省人工智能研究院 | Deep neural network electrocardiosignal noise reduction method based on interpretability |
CN115409733A (en) * | 2022-09-02 | 2022-11-29 | 山东财经大学 | Low-dose CT image noise reduction method based on image enhancement and diffusion model |
CN116051408A (en) * | 2023-01-06 | 2023-05-02 | 郑州轻工业大学 | Image depth denoising method based on residual error self-coding |
CN116051408B (en) * | 2023-01-06 | 2023-10-27 | 郑州轻工业大学 | Image depth denoising method based on residual error self-coding |
CN116757966A (en) * | 2023-08-17 | 2023-09-15 | 中科方寸知微(南京)科技有限公司 | Image enhancement method and system based on multi-level curvature supervision |
CN117095074A (en) * | 2023-08-23 | 2023-11-21 | 浙江大学 | Medical image double denoising method and device based on deep learning |
CN117095074B (en) * | 2023-08-23 | 2024-04-09 | 浙江大学 | Medical image double denoising method and device based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110570353B (en) | Super-resolution reconstruction method for generating single image of countermeasure network by dense connection | |
CN114219719A (en) | CNN medical CT image denoising method based on dual attention and multi-scale features | |
CN110889852B (en) | Liver segmentation method based on residual error-attention deep neural network | |
CN114092330B (en) | Light-weight multi-scale infrared image super-resolution reconstruction method | |
Fang et al. | Blind visual quality assessment for image super-resolution by convolutional neural network | |
CN112819910B (en) | Hyperspectral image reconstruction method based on double-ghost attention machine mechanism network | |
CN115409733A (en) | Low-dose CT image noise reduction method based on image enhancement and diffusion model | |
CN111968058A (en) | Low-dose CT image noise reduction method | |
CN109214989A (en) | Single image super resolution ratio reconstruction method based on Orientation Features prediction priori | |
CN116309648A (en) | Medical image segmentation model construction method based on multi-attention fusion | |
CN112836602B (en) | Behavior recognition method, device, equipment and medium based on space-time feature fusion | |
CN115170915A (en) | Infrared and visible light image fusion method based on end-to-end attention network | |
Li et al. | Hdrnet: Single-image-based hdr reconstruction using channel attention cnn | |
CN116664397B (en) | TransSR-Net structured image super-resolution reconstruction method | |
CN112561799A (en) | Infrared image super-resolution reconstruction method | |
CN116757986A (en) | Infrared and visible light image fusion method and device | |
Krishnan et al. | SwiftSRGAN-Rethinking super-resolution for efficient and real-time inference | |
CN116012255A (en) | Low-light image enhancement method for generating countermeasure network based on cyclic consistency | |
CN117036162B (en) | Residual feature attention fusion method for super-resolution of lightweight chest CT image | |
Qiu et al. | End-to-end residual attention mechanism for cataractous retinal image dehazing | |
CN117217997A (en) | Remote sensing image super-resolution method based on context perception edge enhancement | |
CN116152060A (en) | Double-feature fusion guided depth image super-resolution reconstruction method | |
CN115346091A (en) | Method and device for generating Mura defect image data set | |
Tian et al. | Retinal fundus image superresolution generated by optical coherence tomography based on a realistic mixed attention GAN | |
Zhang et al. | Deep residual network based medical image reconstruction |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |