CN108309251A - Quantitative acousto-optic imaging method based on deep neural network - Google Patents
Quantitative acousto-optic imaging method based on deep neural network Download PDFInfo
- Publication number
- CN108309251A CN108309251A CN201810228027.4A CN201810228027A CN108309251A CN 108309251 A CN108309251 A CN 108309251A CN 201810228027 A CN201810228027 A CN 201810228027A CN 108309251 A CN108309251 A CN 108309251A
- Authority
- CN
- China
- Prior art keywords
- layer
- minor structure
- image
- output
- minor
- 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.)
- Granted
Links
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/0093—Detecting, measuring or recording by applying one single type of energy and measuring its conversion into another type of energy
- A61B5/0095—Detecting, measuring or recording by applying one single type of energy and measuring its conversion into another type of energy by applying light and detecting acoustic waves, i.e. photoacoustic measurements
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Acoustics & Sound (AREA)
- Pathology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Public Health (AREA)
- Veterinary Medicine (AREA)
- Image Processing (AREA)
- Ultra Sonic Daignosis Equipment (AREA)
Abstract
The present invention relates to a kind of quantitative acousto-optic imaging method based on deep neural network, it is characterised in that including the following contents：Build deep neural network frame, i.e. ResU net；Using the initial pressure pattern of input and corresponding quantitative image training ResU net under different wave length；Quantitative photoacoustic imaging is carried out using the ResU net after training to the initial pressure pattern of multi-wavelength, exports quantitative image.The present invention proposes the deep neural network for quantitative photoacoustic imaging, that is ResU net, ResU net have used residual error study mechanism, so that network is easy optimization, and comparable depth can be reached, to obtain higher accuracy rate, further, the constricted path and path expander of deep neural network of the present invention setting allow the environmental information that comprehensive each resolution ratio level is extracted in the initial pressure pattern of the multi-wavelength of ResU net from input, and the high-resolution quantitative image of final output.
Description
Technical field
The present invention relates to a kind of acousto-optic imaging method, more particularly to a kind of quantitative optoacoustic based on deep neural network at
Image space method.
Background technology
Preferable spatial resolution and good specificity may be implemented in optoacoustic (photoacoustic, PA) imaging.It is quantitative
Multispectral PA images can be converted into a series of precise images by photoacoustic imaging (quantitative PA imaging, QPAI)：
Such as the specific molecular marker for molecular imaging concentration image, for assessing tumour growth, metabolism and to each
The blood oxygen saturation figure of resistance etc. of kind treatment.
Influence of traditional QPAI methods without apparent wavelength to luminous flux, and estimate SO using linear fit2It is (i.e. linear
Solution is mixed, linear unmixing) gross error can be brought.Most of luminous flux antidote dependent on it is very strong it is assumed that than
As optical property have piecewise constant characteristic, to be known in advance scattering parameter, require background optical parameter uniformly (and known)
Deng, or it is too big to the reconstruction calculation amount of large-scale high-definition picture.Diffuse Optical Tomography imaging can help to estimate luminous flux,
Cost is to lose the ingredient of high spatial resolution.
Invention content
In view of the above-mentioned problems, the object of the present invention is to provide it is a kind of can export high-resolution quantitative image based on depth
The quantitative acousto-optic imaging method of neural network.
To achieve the above object, the present invention quantitative optoacoustic based on deep neural network of taking following technical scheme a kind of at
Image space method, it is characterised in that including the following contents：Build deep neural network frame, i.e. ResU-net；Using under different wave length
The initial pressure pattern of input and corresponding quantitative image train ResU-net；Training is used to the initial pressure pattern of multi-wavelength
ResU-net afterwards carries out quantitative photoacoustic imaging, exports quantitative image.
Further, the detailed process of structure deep neural network frame is：1) son of deep neural network frame is built
Structure, minor structure include input minor structure, straton structure, shrink minor structure, expansion minor structure and export minor structure；2) it is based on deep
The minor structure for spending neural network framework carries out the successively structure of hierarchical structure, obtains deep neural network frame, detailed process is：
The top layer hierarchical structure of hierarchical structure includes input minor structure, top layer straton structure, expansion minor structure and output minor structure successively；
The 2nd layer of hierarchical structure is identical to the 4th layer of structure, successively include shrink minor structure, shrinkage layer minor structure, expansion minor structure and
Expansion layer minor structure；The bottom hierarchical structure of hierarchical structure includes shrinking minor structure and bottom straton structure successively；Wherein, it inputs
Minor structure, top layer straton structure shrink minor structure and shrinkage layer minor structure composition constricted path, bottom straton structure, expansion
Minor structure, expansion layer minor structure and output minor structure constitute path expander；The output of minor structure is inputted as top layer straton knot
The input of structure, the output of top layer straton structure is as this layer of expansion minor structure and the next layer of input for shrinking minor structure, top layer
Input of the output of the expansion minor structure of hierarchical structure as output minor structure；Contraction minor structure in 2nd layer to the 4th layer it is defeated
Go out the input of corresponding each layer shrinkage layer minor structure, the output of shrinkage layer minor structure as the expansion minor structure of equivalent layer input with
And the input of next layer of contraction minor structure；The corresponding each layer expansion layer minor structure of output of minor structure is expanded in 2nd layer to the 4th layer
Input, the input for exporting the bottom straton structure as equivalent layer of the contraction minor structure of bottom, the output of bottom straton structure
The input of minor structure is expanded as last layer, and so on, until the expansion layer minor structure of the second layer exports the expansion as top layer
Open the input of minor structure.
Further, each minor structure carries out residual error study by corresponding residual error study minor structure to input picture, residual
Difference study minor structure is used to carry out residual error study to image, and residual error study minor structure includes that main connection is connected with shortcut.
Further, quantitative photoacoustic imaging is carried out using the ResU-net after training to the initial pressure pattern of multi-wavelength,
Quantitative image is exported, detailed process is：Minor structure is inputted by the initial pressure pattern of the multi-wavelength of input by inputting residual error
It practises minor structure and carries out residual error study, in inputting minor structure, first convolutional layer in main connection can change the port number of tensor,
The wavelength number for inputting initial acoustic pressure is changed to 32, and in shortcut connection, be added to the convolutional layer of a convolution kernel 1 × 1 into
The identical number of active lanes transformation of row, i.e., be set as 32, specially by inputting minor structure by the number of active lanes of initial pictures：By
It is the 32 big small image for being 128 × 128 that minor structure, which is inputted, by the image output that multiple sizes are 128 × 128, inputs minor structure
32 images are input to fine tuning extraction feature in top layer straton structure, top layer straton structure is by 32 pictures sizes are exported
128 × 128 image is output to the expansion minor structure of top layer and the contraction minor structure of the second layer；Constricted path is to image from top layer
It is shunk successively to bottom, each contraction minor structure respectively locates image by shrinking residual error study minor structure respectively
Reason, image can first pass through maximum pond layer in main connection, and picture size reduces half, and then first convolutional layer makes channel
Number is double, and in shortcut connection, the convolutional layer of a convolution kernel 1 × 1 is used for so that number of active lanes is double, picture size subtracts
Half, shrink shrinkage layer minor structure fine tuning extraction feature of the minor structure by treated image is sent to equivalent layer, the receipts of the second layer
The corresponding output channel number of contracting straton structure be 64, picture size be 64 × 64 image to equivalent layer expansion minor structure and
Next layer of contraction minor structure, and so on, the contraction minor structure output channel number of bottom is 512, and picture size is 8 × 8
Image；Path expander expands image successively from bottom to top layer, and each expansion minor structure is learnt by expanding residual error
Minor structure respectively expands image respectively, and in main connection, low one layer of output image can be adopted first by increasing in path expander
At twice, then first convolutional layer can allow number of active lanes to halve；Before second convolutional layer, image can elder generation and constricted path
It is cascaded after the image procossing of output, number of active lanes can double again；After second convolutional layer, number of active lanes can halve extensive
It is multiple；Shortcut connection in, low one layer of the output of path expander can be adopted by increasing, then by a convolution kernel be 1 × convolution
Layer, number of active lanes compress half；Bottom straton structure is 512 to the number of active lanes of the contraction minor structure output of bottom, size 8
× 8 image fine tuning extraction feature is output to the expansion minor structure of last layer, and image increasing is adopted into twice of size by expansion minor structure
And number of active lanes is halved to the image cascade exported afterwards with constricted path and handles to obtain size to be 16 × 16, number of active lanes 256
Image be output to expansion layer minor structure fine tuning reconstruction image and be output to the expansion minor structure of last layer, and so on, top layer
Expansion minor structure the image of the expansion layer minor structure output of the image of top layer straton structure output and the 2nd layer handle
It is 128 × 128 to picture size, the image that number of active lanes is 32；Minor structure is exported to be used to learn minor structure by exporting residual error
Residual error study is carried out to the image of top layer expansion minor structure output, the third convolutional layer in main connection changes number of active lanes from 32
It is 1, using 1 × 1 convolution kernel, a convolutional layer with said function is also added in shortcut connection, exports minor structure
Export the quantitative image that a pictures size is 128 × 128.
The invention adopts the above technical scheme, which has the following advantages：1, the present invention propose for quantitative optoacoustic at
The deep neural network of picture, i.e. ResU-net, ResU-net have used residual error study mechanism so that and network is easy optimization, and
Comparable depth can be reached, to obtain higher accuracy rate.2, the constricted path of deep neural network setting of the present invention and expansion
Path, which allows, extracts comprehensive each resolution ratio level in the initial pressure pattern of the multi-wavelength of ResU-net from input
Environmental information, and the high-resolution quantitative image of final output.3, deep neural network of the invention is passed through after training,
Quick image reconstruction may be implemented in ResU-net, and the reconstruction through the secondary quantitative image of experiment one takes 22.06ms.The present invention can be with
It is widely used in quantitative photoacoustic imaging.
Description of the drawings
Fig. 1 is the deep neural network block schematic illustration of the present invention, and D is the abbreviation of Dimension (image size)；
Fig. 2 is residual error study minor structure schematic diagram of the present invention by taking straton structure as an example.
Specific implementation mode
Come to carry out detailed description to the present invention below in conjunction with attached drawing.It should be appreciated, however, that attached drawing has been provided only more
Understand the present invention well, they should not be interpreted as limitation of the present invention.
Deep learning (deep learning, DL) causes to pay close attention in many fields, including medical imaging.Using special
Depth nerve (deep neural network, DNN) characterizes nonlinear mapping, and net is adjusted by big training data
Network weight, DL can the detection feature from measurement data automatically, and using the feature prediction target data for excavating out or
Carry out decision.
Convolutional neural networks (convolutional neural network, CNN) have superior process performance to image,
The nervous layer of CNN can filter input data and extract useful information.U-net is a kind of full convolutional neural networks.U-net by
One constricted path (capturing ambient information) and symmetrical path expander (can a be accurately positioned) composition, the contraction of U-net
Path and path expander allow it to extract the environmental information under comprehensive different resolution, and final output high-resolution mesh
Logo image.Depth network can integrate different level feature and target (grader or image) in a multilayer end to end
Structure.A large amount of crenel nervous layer makes the feature under different level be mined out.However, too deep network can bring gradient
It the problem of disappearance/explosion and hinders to restrain.Internal covariant offset can be reduced and accelerate to restrain by criticizing normalization.In addition,
It will appear the problem of training precision is degenerated when network is too deep, it is meant that precision is saturated, then rapid to degenerate.Residual error
Habit mechanism can solve degenerate problem so that network is easy optimization, and realizes quite deep network, obtains high-precision.
Based on above-mentioned principle, the quantitative acousto-optic imaging method of the invention based on deep neural network, by residual error learning machine
System is introduced into U-net structures, and structure deep neural network frame, that is, ResU-net solves the problems, such as that QPAI, detailed process are：
1, deep neural network frame, i.e. ResU-net are built, specific building process is：
1) as shown in Figure 1, the minor structure of structure deep neural network frame, each minor structure are learnt by corresponding residual error
Minor structure carries out residual error study to input picture, wherein the minor structure of deep neural network frame includes input minor structure, straton
Structure shrinks minor structure, expansion minor structure and output minor structure.
Straton structure as shown in Figure 2, residual error learn minor structure and are used to carry out residual error study, residual error study son knot to image
Structure includes that main connection is connected with shortcut.What the shortcut connection in straton structure play a part of is identical mapping, and shortcut connects defeated
Go out the direct output with main connection to be overlapped, main connection is by convolutional layer Conv (convolution kernel 3 × 3), crowd normalization layer BN and repaiies
Linear positive unit (Rectified Linear Unit, ReLU) is used as excitation function.Specific the executing of residual error study minor structure
Journey is the prior art, and details are not described herein.
Input minor structure is handled the initial pressure pattern of multi-wavelength by inputting residual error study minor structure；
Straton structure learns minor structure by straton residual error and is adjusted extraction feature to the image of input, in straton structure
In, picture size and number of channels remain unchanged.
It shrinks minor structure and the image of input is shunk by shrinking residual error study minor structure, for extracting different resolutions
Environmental information under rate.
Expansion minor structure expands the image of input by expanding residual error study minor structure, for different resolution
Under environmental information rebuild.
Output minor structure carries out processing output quantitative image by exporting residual error study minor structure to the image of input.
2) the successively structure that hierarchical structure is carried out based on deep neural network frame sub-structure, obtains deep neural network frame
Frame.
As shown in Figure 1, the deep neural network frame of the present invention includes that 5 grades of hierarchical structures are (real as example from top to bottom
The series of border hierarchical structure can be determined according to actual treatment dimension of picture and port number).
Top layer level (the 1st layer) structure of hierarchical structure includes input minor structure, top layer straton structure, expansion son knot successively
Structure and output minor structure.
The 2nd layer of hierarchical structure is identical to the 4th layer of structure, includes shrinking minor structure, shrinkage layer minor structure, expansion successively
Minor structure and expansion layer minor structure.
Bottom level (the 5th layer) structure of hierarchical structure includes shrinking minor structure and bottom straton structure successively.
Wherein, minor structure, top layer straton structure are inputted, shrinks minor structure and shrinkage layer minor structure composition constricted path,
Bottom straton structure, expansion minor structure, expansion layer minor structure and output minor structure constitute path expander.
Input of the output of minor structure as top layer straton structure is inputted, the output of top layer straton structure is expanded as this layer
The input of minor structure and next layer of contraction minor structure, the output of the expansion minor structure of top layer hierarchical structure is as output minor structure
Input.The input of the corresponding each layer shrinkage layer minor structure of output of contraction minor structure in 2nd layer to the 4th layer, shrinks straton knot
The output of structure is as the input of the expansion minor structure of equivalent layer and the input of next layer of contraction minor structure；In 2nd layer to the 4th layer
The input of the corresponding each layer expansion layer minor structure of output of minor structure is expanded, the output of the contraction minor structure of bottom is as equivalent layer
The input of minor structure is expanded in the input of expansion layer minor structure, the output of bottom straton structure as last layer, and so on, until
Input of the expansion layer minor structure output of the second layer as the expansion minor structure of top layer.
2, using under different wave length the initial acoustic pressure of input (formula 1) image and corresponding quantitative image be used for train ResU-
When training, using mean square error as loss function, network parameter, instruction are solved using the Adam algorithms in tensorflow by net
It is the prior art to practice process, and details are not described herein.Wherein, initial acoustic pressure po(r)：
po(r)=Γ H (r) (1)
Wherein, poIt is initial acoustic pressure, Γ is Gr ü neisen parameters (assuming that being constant), energy deposited density H (r)=μ α
(r) Φ (r), μ α (r) are the absorption coefficients of light, and Φ (r) is luminous flux, and r is spatial position.
By after training, quick image reconstruction may be implemented in ResU-net, each iteration of the embodiment of the present invention makes
With 16 samples, total iterations are 24000 times, and the training time is about 3h, use the GPU of tall and handsome Da Taitan video cards
(12GB), by after training, the reconstruction of a quantitative image only needs 22.06ms.
3, quantitative photoacoustic imaging is carried out using the ResU-net after training to the initial pressure pattern of multi-wavelength, output is quantitative
Image, detailed process are：
As shown in Figure 1, input minor structure ties the initial pressure pattern of the multi-wavelength of input by inputting residual error study
Structure carries out residual error study.In inputting minor structure, first convolutional layer in main connection can change the port number of tensor, will input
The wavelength number of initial acoustic pressure is changed to 32, and in shortcut connection, it is identical to be added to a convolutional layer (convolution kernel 1 × 1) progress
Number of active lanes transformation, i.e., by input minor structure the number of active lanes of initial pictures is set as 32, specially：By input
The image output that multiple sizes are 128 × 128 is the 32 big small image for being 128 × 128 by structure, and input minor structure is by 32
Image is input to fine tuning extraction feature in top layer straton structure, and it is 128 × 128 that top layer straton structure, which will export 32 pictures sizes,
Image be output to the expansion minor structure of top layer and the contraction minor structure of the second layer.
Constricted path shrinks image successively from top to bottom, and each contraction minor structure is learnt by shrinking residual error
Minor structure is respectively respectively processed image, and image can first pass through maximum pond layer in main connection, and picture size reduces one
Half, then first convolutional layer so that number of active lanes is double, and in shortcut connection, a convolutional layer (convolution kernel 1 × 1) is used for making
Number of active lanes is double, picture size halves (convolution step-length 2), shrinking minor structure, image is sent to equivalent layer by treated
Shrinkage layer minor structure fine tuning extraction feature, the corresponding output channel number of shrinkage layer minor structure of the second layer is 64, and picture size is
64 × 64 image to equivalent layer expansion minor structure and next layer of contraction minor structure, and so on, the contraction of bottom
Structure output number of active lanes is 512, the image that picture size is 8 × 8.
Path expander expands image successively from bottom to top layer, and each expansion minor structure is learnt by expanding residual error
Minor structure respectively expands image respectively, and in main connection, low one layer of output image can be adopted first by increasing in path expander
At twice, then first convolutional layer can allow number of active lanes to halve.Before second convolutional layer, image can elder generation and constricted path
It is cascaded after the image procossing of output, number of active lanes can double again.After second convolutional layer, number of active lanes can halve extensive
It is multiple.In shortcut connection, low one layer of the output of path expander can be adopted by increasing, the convolution for being then 1 × 1 by a convolution kernel
Layer, number of active lanes compress half.Bottom straton structure is 512 to the number of active lanes of the contraction minor structure output of bottom, size 8
× 8 image fine tuning extraction feature is output to the expansion minor structure of last layer, and image increasing is adopted into twice of size by expansion minor structure
And number of active lanes is halved to the image cascade exported afterwards with constricted path and handles to obtain size to be 16 × 16, number of active lanes 256
Image be output to expansion layer minor structure fine tuning reconstruction image and be output to the expansion minor structure of last layer, and so on, top layer
Expansion minor structure the image of the expansion layer minor structure output of the image of top layer straton structure output and the 2nd layer handle
It is 128 × 128 to picture size, the image that number of active lanes is 32.
Output minor structure is used to carry out the image of top layer expansion minor structure output by output residual error study minor structure residual
Number of active lanes is changed to 1 by difference study, the third convolutional layer in main connection from 32, and using 1 × 1 convolution kernel, one has together
The convolutional layer of sample function is also added in shortcut connection, and output minor structure exports the high score that a pictures size is 128 × 128
The quantitative image of resolution.
The various embodiments described above are merely to illustrate the present invention, and the implementation steps etc. of wherein method may be changed,
Every equivalents carried out based on the technical solution of the present invention and improvement, should not exclude in protection scope of the present invention
Except.
Claims (4)
1. a kind of quantitative acousto-optic imaging method based on deep neural network, it is characterised in that including the following contents：
Build deep neural network frame, i.e. ResU-net；
Using the initial pressure pattern of input and corresponding quantitative image training ResU-net under different wave length；
Quantitative photoacoustic imaging is carried out using the ResU-net after training to the initial pressure pattern of multi-wavelength, exports quantitative image.
2. the quantitative acousto-optic imaging method based on deep neural network as described in claim 1, which is characterized in that structure depth
The detailed process of neural network framework is：
1) build deep neural network frame minor structure, minor structure include input minor structure, straton structure, shrink minor structure,
Expand minor structure and output minor structure；
2) minor structure based on deep neural network frame carries out the successively structure of hierarchical structure, obtains deep neural network frame
Frame, detailed process are：
The top layer hierarchical structure of hierarchical structure includes input minor structure, top layer straton structure, expansion minor structure and output successively
Structure；The 2nd layer of hierarchical structure is identical to the 4th layer of structure, includes shrinking minor structure, shrinkage layer minor structure, expansion successively
Structure and expansion layer minor structure；The bottom hierarchical structure of hierarchical structure includes shrinking minor structure and bottom straton structure successively；Its
In, input minor structure, top layer straton structure shrink minor structure and shrinkage layer minor structure composition constricted path, bottom straton knot
Structure, expansion minor structure and expansion layer minor structure and output minor structure constitute path expander；
Input of the output of minor structure as top layer straton structure is inputted, the output of top layer straton structure expands son knot as this layer
The input of structure and next layer of contraction minor structure, the output of the expansion minor structure of top layer hierarchical structure is as the defeated of output minor structure
Enter；The input of the corresponding each layer shrinkage layer minor structure of output of contraction minor structure in 2nd layer to the 4th layer, shrinkage layer minor structure
Export the input of the expansion minor structure as equivalent layer and the input of next layer of contraction minor structure；It is expanded in 2nd layer to the 4th layer
The input of the corresponding each layer expansion layer minor structure of output of minor structure, the bottom of the contraction minor structure of bottom exported as equivalent layer
The input of minor structure is expanded in the input of straton structure, the output of bottom straton structure as last layer, and so on, until second
Input of the expansion layer minor structure output of layer as the expansion minor structure of top layer.
3. the quantitative acousto-optic imaging method based on deep neural network as claimed in claim 2, which is characterized in that each sub- knot
Structure learns minor structure by corresponding residual error and carries out residual error study to input picture, and residual error learns minor structure and is used to carry out image
Residual error learns, and residual error study minor structure includes that main connection is connected with shortcut.
4. the quantitative acousto-optic imaging method based on deep neural network as claimed in claim 3, which is characterized in that multi-wavelength
Initial pressure pattern quantitative photoacoustic imaging is carried out using the ResU-net after training, export quantitative image, detailed process is：
It inputs minor structure and the initial pressure pattern of the multi-wavelength of input is learnt into minor structure progress residual error study by inputting residual error,
In inputting minor structure, first convolutional layer in main connection can change the port number of tensor, will input the wavelength of initial acoustic pressure
Number is changed to 32, and in shortcut connection, the convolutional layer for being added to a convolution kernel 1 × 1 carries out identical number of active lanes transformation,
The number of active lanes of initial pictures is set as 32 by inputting minor structure, specially：By inputting minor structure by multiple sizes
It is the 32 big small image for being 128 × 128 for 128 × 128 image output, 32 images are input to top layer by input minor structure
Fine tuning extraction feature in straton structure, top layer straton structure are output to top by the image that 32 pictures sizes are 128 × 128 is exported
The contraction minor structure of the expansion minor structure and the second layer of layer；
Constricted path shrinks image successively from top to bottom, and each contraction minor structure is by shrinking residual error study son knot
Structure is respectively respectively processed image, and image can first pass through maximum pond layer in main connection, and picture size reduces half, so
First convolutional layer so that number of active lanes is double afterwards, and in shortcut connection, the convolutional layer of a convolution kernel 1 × 1 is used for so that logical
Road number is double, picture size halves, and shrinks minor structure the shrinkage layer minor structure of treated image is sent to equivalent layer is micro-
Extraction feature is adjusted, the corresponding output channel number of shrinkage layer minor structure of the second layer is 64, and the image that picture size is 64 × 64 arrives
The expansion minor structure of equivalent layer and next layer of contraction minor structure, and so on, the contraction minor structure output channel number of bottom
Mesh is 512, the image that picture size is 8 × 8；
Path expander expands image successively from bottom to top layer, and each expansion minor structure is by expanding residual error study son knot
Structure respectively expands image respectively, and in main connection, low one layer of output image first can adopt into two by increasing in path expander
Times, then first convolutional layer can allow number of active lanes to halve；Before second convolutional layer, image can be exported first with constricted path
Image procossing after cascade, number of active lanes can double again；After second convolutional layer, number of active lanes can halve recovery；
In shortcut connection, low one layer of the output of path expander can be adopted by increasing, and the convolutional layer for being then 1 × 1 by a convolution kernel leads to
Road number compresses half；Bottom straton structure is 512 to the number of active lanes of the contraction minor structure output of bottom, and size is 8 × 8
Image fine tuning extraction feature is output to the expansion minor structure of last layer, and image increasing is adopted into twice of size and will led to by expansion minor structure
Road number halves the image cascade exported afterwards with constricted path and handles to obtain size to be 16 × 16, the image that number of active lanes is 256
It is output to expansion layer minor structure fine tuning reconstruction image and is output to the expansion minor structure of last layer, and so on, the expansion of top layer
Minor structure is handled the image of the image of top layer straton structure output and the output of the 2nd layer of expansion layer minor structure to obtain picture
Size is 128 × 128, the image that number of active lanes is 32；
It exports minor structure and is used to pass through the image progress residual error that output residual error study minor structure expands top layer minor structure output
It practises, number of active lanes is changed to 1 by the third convolutional layer in main connection from 32, and using 1 × 1 convolution kernel, one has same work(
The convolutional layer of energy is also added in shortcut connection, and output minor structure exports the quantitative figure that a pictures size is 128 × 128
Picture.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810228027.4A CN108309251B (en) | 2018-03-20 | 2018-03-20 | Quantitative photoacoustic imaging method based on deep neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810228027.4A CN108309251B (en) | 2018-03-20 | 2018-03-20 | Quantitative photoacoustic imaging method based on deep neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108309251A true CN108309251A (en) | 2018-07-24 |
| CN108309251B CN108309251B (en) | 2020-09-25 |
Family
ID=62898962
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810228027.4A Active CN108309251B (en) | 2018-03-20 | 2018-03-20 | Quantitative photoacoustic imaging method based on deep neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108309251B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109509178A (en) * | 2018-10-24 | 2019-03-22 | 苏州大学 | A kind of OCT image choroid dividing method based on improved U-net network |
| CN111193916A (en) * | 2018-12-29 | 2020-05-22 | 中科寒武纪科技股份有限公司 | Operation method, device and related product |
| CN111724450A (en) * | 2019-03-20 | 2020-09-29 | 上海科技大学 | Medical image reconstruction system, method, terminal and medium based on deep learning |
| CN111915691A (en) * | 2019-05-07 | 2020-11-10 | 上海科技大学 | Image processing system, method, terminal and medium based on neural network |
| CN115024739A (en) * | 2022-08-11 | 2022-09-09 | 之江实验室 | Method for measuring distribution of Glehnson parameters in organism and application |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106473751A (en) * | 2016-11-25 | 2017-03-08 | 刘国栋 | Palm blood vessel imaging based on arrayed ultrasonic sensor and identifying device and its imaging method |
| CN106510635A (en) * | 2016-11-30 | 2017-03-22 | 江西科技师范大学 | Skin vessel photoacoustic imaging device |
| CN106779050A (en) * | 2016-11-24 | 2017-05-31 | 厦门中控生物识别信息技术有限公司 | The optimization method and device of a kind of convolutional neural networks |
| CN107705305A (en) * | 2017-10-20 | 2018-02-16 | 上海联影医疗科技有限公司 | A kind of method and device of Medical Image Processing |
-
2018
- 2018-03-20 CN CN201810228027.4A patent/CN108309251B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106779050A (en) * | 2016-11-24 | 2017-05-31 | 厦门中控生物识别信息技术有限公司 | The optimization method and device of a kind of convolutional neural networks |
| CN106473751A (en) * | 2016-11-25 | 2017-03-08 | 刘国栋 | Palm blood vessel imaging based on arrayed ultrasonic sensor and identifying device and its imaging method |
| CN106510635A (en) * | 2016-11-30 | 2017-03-22 | 江西科技师范大学 | Skin vessel photoacoustic imaging device |
| CN107705305A (en) * | 2017-10-20 | 2018-02-16 | 上海联影医疗科技有限公司 | A kind of method and device of Medical Image Processing |
Non-Patent Citations (2)
| Title |
|---|
| LVMIN ZHANG ETC: "《Style Transfer for Anime Sketches with Enhanced Residual U-net and Auxiliary Classifier GAN》", 《ARXIV》 * |
| MD ZAHANGIR ALOM ETC: "《Recurrent Residual Convolutional Neural Network based on U-Net（R2U-Net） for Medical Segmentation》", 《HTTP：// WWW.RESEARCHGATE.NET/PUBLICATION》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109509178A (en) * | 2018-10-24 | 2019-03-22 | 苏州大学 | A kind of OCT image choroid dividing method based on improved U-net network |
| CN109509178B (en) * | 2018-10-24 | 2021-09-10 | 苏州大学 | OCT image choroid segmentation method based on improved U-net network |
| CN111193916A (en) * | 2018-12-29 | 2020-05-22 | 中科寒武纪科技股份有限公司 | Operation method, device and related product |
| CN111193916B (en) * | 2018-12-29 | 2022-03-29 | 中科寒武纪科技股份有限公司 | Operation method |
| CN111724450A (en) * | 2019-03-20 | 2020-09-29 | 上海科技大学 | Medical image reconstruction system, method, terminal and medium based on deep learning |
| CN111915691A (en) * | 2019-05-07 | 2020-11-10 | 上海科技大学 | Image processing system, method, terminal and medium based on neural network |
| CN111915691B (en) * | 2019-05-07 | 2023-08-22 | 上海科技大学 | Image processing system, method, terminal and medium based on neural network |
| CN115024739A (en) * | 2022-08-11 | 2022-09-09 | 之江实验室 | Method for measuring distribution of Glehnson parameters in organism and application |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108309251B (en) | 2020-09-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108309251A (en) | Quantitative acousto-optic imaging method based on deep neural network | |
| CN106600577A (en) | Cell counting method based on depth deconvolution neural network | |
| CN106296699A (en) | Cerebral tumor dividing method based on deep neural network and multi-modal MRI image | |
| CN108197606A (en) | The recognition methods of abnormal cell in a kind of pathological section based on multiple dimensioned expansion convolution | |
| CN105469100A (en) | Deep learning-based skin biopsy image pathological characteristic recognition method | |
| CN109146899A (en) | CT image jeopardizes organ segmentation method and device | |
| CN109815785A (en) | A kind of face Emotion identification method based on double-current convolutional neural networks | |
| CN106250707A (en) | A kind of based on degree of depth learning algorithm process head construction as the method for data | |
| CN108446730A (en) | A kind of CT pulmonary nodule detection methods based on deep learning | |
| CN110222701A (en) | A kind of bridge defect automatic identifying method | |
| CN107016406A (en) | The pest and disease damage image generating method of network is resisted based on production | |
| CN103871041B (en) | The image super-resolution reconstructing method built based on cognitive regularization parameter | |
| US20080139966A1 (en) | Automatic tongue diagnosis based on chromatic and textural features classification using bayesian belief networks | |
| CN106650789A (en) | Image description generation method based on depth LSTM network | |
| CN108510502A (en) | Melanoma picture tissue segmentation methods based on deep neural network and system | |
| CN108090403A (en) | A kind of face dynamic identifying method and system based on 3D convolutional neural networks | |
| CN108363979A (en) | Neonatal pain expression recognition method based on binary channels Three dimensional convolution neural network | |
| CN107563434A (en) | A kind of brain MRI image sorting technique based on Three dimensional convolution neutral net, device | |
| CN108229576A (en) | Across the multiplying power pathological image feature learning method of one kind | |
| CN108053398A (en) | A kind of melanoma automatic testing method of semi-supervised feature learning | |
| CN108198625A (en) | A kind of deep learning method and apparatus for analyzing higher-dimension medical data | |
| CN107330393A (en) | A kind of neonatal pain expression recognition method based on video analysis | |
| CN109325915A (en) | A kind of super resolution ratio reconstruction method for low resolution monitor video | |
| CN113724206A (en) | Fundus image blood vessel segmentation method and system based on self-supervision learning | |
| CN110532914A (en) | Building analyte detection method based on fine-feature study |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |