CN109924949A - A kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks - Google Patents
A kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks Download PDFInfo
- Publication number
- CN109924949A CN109924949A CN201910161938.4A CN201910161938A CN109924949A CN 109924949 A CN109924949 A CN 109924949A CN 201910161938 A CN201910161938 A CN 201910161938A CN 109924949 A CN109924949 A CN 109924949A
- Authority
- CN
- China
- Prior art keywords
- boundary
- neural networks
- convolutional neural
- near infrared
- infrared spectrum
- 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
Landscapes
- Image Analysis (AREA)
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
The invention discloses a kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks, belongs to field of medical image processing.Using the boundary survey value of convolutional neural networks expression near infrared spectrum tomographic imaging target and the Nonlinear Mapping relationship of internal optics parameter distribution, it can be absorbed from measured value index profile picture by network training, realize sensor domain directly reconstructing to image area.This method can in the case where guaranteeing optical parameter distribution accurate reconstruction computational efficiency with higher.
Description
Technical field
The invention belongs to field of medical image processing, be related to a kind of near infrared spectrum tomography based on convolutional neural networks at
As method for reconstructing.
Background technique
Near infrared spectrum (Near-infrared spectroscopy, NIRS) tomographic imaging is a kind of emerging functionality
Medical imaging mode, imaging device is placed on around biological tissue or surface, will not be to tissue without exogenous probes
Body causes to damage, and near-infrared light waves used by being in addition imaged are to tissue fanout free region, therefore the technology has non-invasi;
Near infrared light imaging can be carried out effectively mechanics of biological tissue and functional imaging, can study the morphosis of biological tissue,
Physiological characteristic, pathological characters, metabolic function etc., particularly suitable for the soft tissue lesions such as characterization breast cancer and the morning of progress cancer
Phase detection, in addition to this, which also has the advantages that portability and low cost, therefore receives the extensive concern of educational circles simultaneously
Various types of blood oxygen detections are successfully applied to, such as cerebral function imaging, neonatal cerebral monitoring, Bones and joints imaging and breast cancer
Detection etc..
The basic principle of the technology are as follows: irradiate organizer using the near-infrared light waves of 600nm to 950nm, and utilize photosensitive
Element measures the light intensity or the signals such as temporal resolution on organizer surface, with this come rebuild in tissue with structure or physiological activity
The distribution situation of relevant optical parameter.The theoretical foundation of the technology is that different material has different scatterings to light in tissue
And absorption characteristic, therefore organize optical parameter (such as absorption coefficient, scattering coefficient) can also change therewith, optical parameter with
Biological tissue's difference physiological status is related, using above-mentioned characteristic, can use the optical parameter of near infrared spectrum tomography rebuilding
In medical diagnosis project.
In actual reconstruction, light is understood some in transmission process and is absorbed by tissue.In the wave band of near infrared light imaging
Interior, the scattering of light is inevitably mixed with noise in boundary survey data much larger than absorption, and tomographic imaging is caused to have
Serious pathosis.It is limited additionally, due to the number of measurement, cause to rebuild organization internal optical parameter problem to be that a discomfort is fixed
The problem of.Spread light within the organization may after Multiple Scattering and absorption, and in addition the penetration depth of light in biological tissues is not
Greatly, therefore light-sensitive element is not also high to the detectivity of emergent light, this causes the spatial resolution of reconstructed results much smaller than other
Image mode.
In order to solve the ill-posedness and pathosis rebuild, it is often used the method for solving of regularization at present by near infrared light figure
As Problems of Reconstruction is transformed into a nonlinear optimization problem.However traditional regularization method is easily affected by noise production
Raw artifact, the imaging reconstruction time is also long, therefore the present invention considers to carry out atlas of near infrared spectra using convolutional neural networks
As rebuilding.
Convolutional neural networks (Convolutional Neural Networks, CNN) be deep learning representative algorithm it
One, with the development of deep learning theory and calculating equipment, convolutional neural networks also gradually rise.Convolutional neural networks are a kind of
Feedforward neural network comprising convolutional calculation is set forth in 1987 earliest, is a kind of improvement of traditional artificial neural network.It compares
For traditional fully-connected network, it has local receptor field and shared convolution kernel, greatly reduces trainable parameter, improves
Training effectiveness;High dimensional data can be handled and automatically extract global and local feature and form multi-characteristic, therefore be rolled up
Product neural network is widely used in the every field such as computer vision, natural language processing with its unique advantage.
Summary of the invention
The purpose of near infrared spectrum tomographic imaging is to utilize near infrared light biological tissue, and use several light sensors
The scattering light for measuring organizational boundary, by the optical parameter inside boundary scattering optical oomputing biological tissue, (absorption coefficient, scattering are
Number etc.) distribution.
To achieve the above object, the technical solution adopted by the present invention is a kind of near infrared spectrum based on convolutional neural networks
Tomography rebuilding method utilizes the measured value and imaging region internal optics parameter point of convolutional neural networks expression object boundary
The Nonlinear Mapping relationship of cloth saves the process of iterative calculation, realizes the reconstruction from sensor domain to image area.
In order to which the relationship between accurate description measured value and optical parameter is to train network, using the expansion of light in the tissue
Scattered approximate equation describes the propagation of light, form are as follows:
Wherein, κ is scattering coefficient;μaFor absorption coefficient;Φ (r, t) indicates photon density distribution;q0(r, t) indicates light source;
T indicates the time;R indicates coordinate position vector;C indicates light transmission in tissue speed.
Since near infrared spectrum tomographic imaging assumes that light source is isotropic light source, i.e. light source does not change over, therefore
Ignore influence of the time to diffusion equation, using the diffusion approximation equation under continuous wave mode, form are as follows:
Wherein, q0It (r) is isotropic light source;Φ (r) is the photon density distribution at the r of position;κ (r) is scattering system
Number;μaIt (r) is absorption coefficient.
The boundary condition that this method uses be Robin boundary condition to indicate that medium refraction index is inconsistent inside and outside boundary when,
For photon the phenomenon that boundary reflects, Robin boundary condition is also referred to as the type III condition or mixed boundary condition of index mismatch.
The boundary condition expresses the light radiation overall strength that photon is equal in medium in the part that boundary is reflected back toward medium, in equation form
Description are as follows:
Wherein, ξ is the point on tissue outer boundary;It is normal outwardly;AnDepending on the mistake between tissue and air
With relative index of refraction (refractive index, RI).Here, AnExpression formula are as follows:
Wherein, RnRefractive indices deviation is related inside and outside expression diffusion transport reflection coefficient, n and boundary.
The forward process of imaging is known optical parameter distribution, is gone out in conjunction with FInite Element and according to diffusion approximation equation solution
Boundary survey value Φ, using Φ as the input of convolutional neural networks, then optical parameter distribution map is the output of network.
The structure of convolutional neural networks of the invention is as shown in Figure 1.The network is divided into two parts, and first half is full connection
Layer FC, it is therefore an objective to which one-dimensional boundary survey value is extended to the data for forming two-dimensional matrix (i.e. image) element enough;Afterwards
Half portion is divided into convolutional layer C and warp lamination DC, to extract the advanced features in data.
If input data isTwo full articulamentum FC1 and FC2 separately include q2And n2A node, then two it is complete
The output y of articulamentumFC1And yFC2It is respectively as follows:
WhereinWithFor full articulamentum weight, f1() and f2() is the activation primitive of full articulamentum, bFC1With
bFC2For biasing, k and j respectively indicate which node of two full articulamentums.
To realize the conversion from sensor domain to image area, in the junction of full articulamentum and convolutional layer, by full articulamentum
One-dimensional vector outputIt is arranged as two-dimensional matrix and forms preliminary imageCarry out convolution again, then two
A convolutional layer output are as follows:
yC1=f3(Y*wC1+bC1) (7)
yC2=f4(yC1*wC2+bC2) (8)
Wherein wC1And wC2For convolutional layer weight, f3() and f4() is convolutional layer activation primitive, bC1And bC2For biasing, *
Indicate convolutional calculation.
Finally final output is obtained by transposition convolution
WhereinIndicate transposition convolution, wDC、f5() and bDCRespectively indicate the weight of transposition convolutional layer, activation primitive and
Biasing.
Network training is to constantly update network weight and biasing by optimization algorithm of RMSprop algorithm, until model error
Deconditioning when meeting the requirements or reaching other training stop conditions.The output of model is optical absorption coefficient distribution.
Detailed description of the invention
Fig. 1 is convolutional neural networks structural schematic diagram of the present invention, wherein input is boundary survey data, is exported as optics point
Cloth image;FC1 and FC2 respectively indicates the first and second full articulamentums;C1 and C2 indicates two convolutional layers;DC indicates transposition convolution
Layer.
Fig. 2 is the finite element fission grid of imitative body, shares 2001 nodes.
Fig. 3 is the position of light source and detector.
Fig. 4 is the initial absorption coefficient distribution of imitative body, and the biggish gray circular region of area is background area in figure, is absorbed
Coefficient is set as 0.01, and the lesser white area of area is the exceptions area added in background, and absorption coefficient is set as 0.039,
The black region at edge is non-imaged areas, and numerical value zero can be ignored in actual imaging.
Fig. 5 is the absorption coefficient distribution results rebuild by this method.
Specific embodiment
Below according to specific implementation example, the present invention will be described with attached drawing.
Firstly, establishing diameter by the tool box nirfast of Matlab is that 80mm circle imitates body and the limited of body is imitated in completion
First mesh generation, finite element mesh result are as shown in Figure 2.In experiment, optical light source and detector has 16, as shown in Figure 3
It is uniformly placed in imitative body surface face, each imitative body sample can measure 240 boundary survey values, and imaging pixel is circle upper 2001 uniform
Finite element node.To train network, the absorption coefficient value on 2001 nodes is mapped in the matrix of 98*98, is formed and is inhaled
Index profile picture, therefore input of the measured value as convolutional neural networks are received, absorption coefficient distributed image is as convolutional Neural
Network is output and input.Initial absorption coefficient distributed image is as shown in Figure 4.Data set for network training passes through
The forward direction solution procedure in the tool box nirfast is calculated, and altogether includes 3916 samples, wherein comprising 3500 training samples and
416 test samples, each sample include a round exceptions area, and radius and the absorption coefficient of each exceptions area randomly select,
The background absorption coefficient of each sample is 0.01, and exceptions area center location does not repeat and include 2% Gaussian noise.
The reconstruction of absorption coefficient of light distribution is carried out using the method for reconstructing based on convolutional neural networks, setting changes in an experiment
Generation number epoch=200, it is 9604 (98*98) that learning rate η=0.00002, FC1 number of nodes, which is 1520, FC2 number of nodes, C1,
The convolution kernel size of C2, DC are respectively 5*5,5*5,7*7, and momentum term parameter is 0.9.
The present invention is distributed method for reconstructing using the absorption coefficient of light based on convolutional neural networks.It is calculated by this method, it can
Directly to go out absorption coefficient distributed image from reconstructed, reconstructed results as shown in figure 5, net training time be 30 minutes,
Reconstruction time is 0.0503s.The experimental results showed that this method, which can not only be distributed absorption coefficient, carries out more accurate reconstruction,
There is higher computational efficiency simultaneously.
Claims (2)
1. a kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks, it is characterised in that: utilize convolution mind
The Nonlinear Mapping relationship of measured value and imaging region internal optics parameter distribution through network expression object boundary, saves iteration
The process of calculating realizes the reconstruction from sensor domain to image area;
The propagation of light, form are described using the diffusion approximation equation of light in the tissue are as follows:
κ is scattering coefficient;μaFor absorption coefficient;Φ (r, t) indicates photon density distribution;q0(r, t) indicates light source;When t is indicated
Between;R indicates coordinate position vector;C indicates light transmission in tissue speed;
Since near infrared spectrum tomographic imaging assumes that light source is isotropic light source, i.e. light source does not change over, therefore ignores
Influence of the time to diffusion equation, using the diffusion approximation equation under continuous wave mode, form are as follows:
Wherein, q0It (r) is isotropic light source;Φ (r) is the photon density distribution at the r of position;κ (r) is scattering coefficient;μa
It (r) is absorption coefficient;
The boundary condition that this method uses be Robin boundary condition to indicate that medium refraction index is inconsistent inside and outside boundary when, photon
The phenomenon that boundary reflects, Robin boundary condition is also referred to as the type III condition or mixed boundary condition of index mismatch;The side
Boundary's condition expresses the light radiation overall strength that photon is equal in medium in the part that boundary is reflected back toward medium, describes in equation form
Are as follows:
Wherein, ξ is the point on tissue outer boundary;It is normal outwardly;AnDepending on the mismatch phase between tissue and air
Refractive index (refractive index, RI);Here, AnExpression formula are as follows:
Wherein, refractive indices deviation is related inside and outside Rn expression diffusion transport reflection coefficient, n and boundary;
The forward process of imaging is known optical parameter distribution, goes out boundary in conjunction with FInite Element and according to diffusion approximation equation solution
Measured value Φ, using Φ as the input of convolutional neural networks, then optical parameter distribution map is the output of convolutional neural networks.
2. a kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks according to claim 1,
Be characterized in that: convolutional neural networks are divided into two parts, and first half is full articulamentum FC, it is therefore an objective to by one-dimensional boundary survey value
It is extended to have and forms the two-dimensional matrix i.e. data of pictorial element enough;Latter half is convolutional layer C and warp lamination DC, to
Extract the advanced features in data;
If input data isTwo full articulamentum FC1 and FC2 separately include q2And n2A node, then two full articulamentums
Output yFC1And yFC2It is respectively as follows:
WhereinWithFor full articulamentum weight, the activation primitive of f1 () and f2 () for full articulamentum, bFC1And bFC2
For biasing, k and j respectively indicate which node of two full articulamentums;
To realize the conversion from sensor domain to image area, in the junction of full articulamentum and convolutional layer, by the one of full articulamentum
Dimensional vector outputIt is arranged as two-dimensional matrix and forms preliminary imageConvolution is carried out again, then two volumes
Lamination output are as follows:
yC1=f3(Y*wC1+bC1) (7)
yC2=f4(yC1*wC2+bC2) (8)
Wherein wC1And wC2For convolutional layer weight, f3() and f4() is convolutional layer activation primitive, bC1And bC2For biasing, * is indicated
Convolutional calculation;
Finally final output is obtained by transposition convolution
WhereinIndicate transposition convolution, wDC、f5() and bDCRespectively indicate the weight of transposition convolutional layer, activation primitive and partially
It sets;
Network training is to constantly update network weight and biasing by optimization algorithm of RMSprop algorithm, until model error meets
It is required that or reach other training stop conditions when deconditioning;The output of model is optical absorption coefficient distribution.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910161938.4A CN109924949A (en) | 2019-03-05 | 2019-03-05 | A kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910161938.4A CN109924949A (en) | 2019-03-05 | 2019-03-05 | A kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109924949A true CN109924949A (en) | 2019-06-25 |
Family
ID=66986292
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910161938.4A Pending CN109924949A (en) | 2019-03-05 | 2019-03-05 | A kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109924949A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110243765A (en) * | 2019-07-02 | 2019-09-17 | 南京农业大学 | The fruit EO-1 hyperion quality detecting method of photon transmission simulation based on fruit double-layer plate model |
CN111104876A (en) * | 2019-12-04 | 2020-05-05 | 南京邮电大学 | Infrared spectrum deconvolution method based on neural network |
CN113821947A (en) * | 2021-07-29 | 2021-12-21 | 清华大学 | Time domain finite difference fast solving method based on cyclic convolution neural network |
CN115272590A (en) * | 2022-09-30 | 2022-11-01 | 慧创科仪(北京)科技有限公司 | Method, apparatus, system and medium for reconstructing spatial distribution of optical transmission parameters |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106846463A (en) * | 2017-01-13 | 2017-06-13 | 清华大学 | Micro-image three-dimensional rebuilding method and system based on deep learning neutral net |
CN106997581A (en) * | 2017-03-01 | 2017-08-01 | 杭州电子科技大学 | A kind of method that utilization deep learning rebuilds high spectrum image |
CN107123091A (en) * | 2017-04-26 | 2017-09-01 | 福建帝视信息科技有限公司 | A kind of near-infrared face image super-resolution reconstruction method based on deep learning |
CN107137072A (en) * | 2017-04-28 | 2017-09-08 | 北京科技大学 | A kind of ventricular ectopic beating detection method based on 1D convolutional neural networks |
CN107292950A (en) * | 2017-06-08 | 2017-10-24 | 深圳市唯特视科技有限公司 | A kind of end-to-end three-dimensional facial reconstruction method based on deep neural network |
WO2018045274A1 (en) * | 2016-09-01 | 2018-03-08 | The General Hospital Corporation | System and method for automated transform by manifold approximation |
CN108805977A (en) * | 2018-06-06 | 2018-11-13 | 浙江大学 | A kind of face three-dimensional rebuilding method based on end-to-end convolutional neural networks |
CN108814550A (en) * | 2018-04-16 | 2018-11-16 | 北京工业大学 | A kind of near infrared spectrum tomography rebuilding method neural network based |
CN109035169A (en) * | 2018-07-19 | 2018-12-18 | 西安交通大学 | A kind of unsupervised/semi-supervised CT image reconstruction depth network training method |
CN109389671A (en) * | 2018-09-25 | 2019-02-26 | 南京大学 | A kind of single image three-dimensional rebuilding method based on multistage neural network |
-
2019
- 2019-03-05 CN CN201910161938.4A patent/CN109924949A/en active Pending
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018045274A1 (en) * | 2016-09-01 | 2018-03-08 | The General Hospital Corporation | System and method for automated transform by manifold approximation |
CN106846463A (en) * | 2017-01-13 | 2017-06-13 | 清华大学 | Micro-image three-dimensional rebuilding method and system based on deep learning neutral net |
CN106997581A (en) * | 2017-03-01 | 2017-08-01 | 杭州电子科技大学 | A kind of method that utilization deep learning rebuilds high spectrum image |
CN107123091A (en) * | 2017-04-26 | 2017-09-01 | 福建帝视信息科技有限公司 | A kind of near-infrared face image super-resolution reconstruction method based on deep learning |
CN107137072A (en) * | 2017-04-28 | 2017-09-08 | 北京科技大学 | A kind of ventricular ectopic beating detection method based on 1D convolutional neural networks |
CN107292950A (en) * | 2017-06-08 | 2017-10-24 | 深圳市唯特视科技有限公司 | A kind of end-to-end three-dimensional facial reconstruction method based on deep neural network |
CN108814550A (en) * | 2018-04-16 | 2018-11-16 | 北京工业大学 | A kind of near infrared spectrum tomography rebuilding method neural network based |
CN108805977A (en) * | 2018-06-06 | 2018-11-13 | 浙江大学 | A kind of face three-dimensional rebuilding method based on end-to-end convolutional neural networks |
CN109035169A (en) * | 2018-07-19 | 2018-12-18 | 西安交通大学 | A kind of unsupervised/semi-supervised CT image reconstruction depth network training method |
CN109389671A (en) * | 2018-09-25 | 2019-02-26 | 南京大学 | A kind of single image three-dimensional rebuilding method based on multistage neural network |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110243765A (en) * | 2019-07-02 | 2019-09-17 | 南京农业大学 | The fruit EO-1 hyperion quality detecting method of photon transmission simulation based on fruit double-layer plate model |
CN111104876A (en) * | 2019-12-04 | 2020-05-05 | 南京邮电大学 | Infrared spectrum deconvolution method based on neural network |
CN113821947A (en) * | 2021-07-29 | 2021-12-21 | 清华大学 | Time domain finite difference fast solving method based on cyclic convolution neural network |
CN113821947B (en) * | 2021-07-29 | 2022-09-30 | 清华大学 | Time domain finite difference fast solving method based on cyclic convolution neural network |
CN115272590A (en) * | 2022-09-30 | 2022-11-01 | 慧创科仪(北京)科技有限公司 | Method, apparatus, system and medium for reconstructing spatial distribution of optical transmission parameters |
CN115272590B (en) * | 2022-09-30 | 2023-01-24 | 慧创科仪(北京)科技有限公司 | Method, apparatus, system and medium for reconstructing spatial distribution of optical transmission parameters |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Yoo et al. | Deep learning diffuse optical tomography | |
Gröhl et al. | Deep learning for biomedical photoacoustic imaging: A review | |
CN108814550A (en) | A kind of near infrared spectrum tomography rebuilding method neural network based | |
Hauptmann et al. | Deep learning in photoacoustic tomography: current approaches and future directions | |
CN109924949A (en) | A kind of near infrared spectrum tomography rebuilding method based on convolutional neural networks | |
Deng et al. | Deep learning in photoacoustic imaging: a review | |
Periyasamy et al. | Advances in Monte Carlo simulation for light propagation in tissue | |
Ren et al. | Frequency domain optical tomography based on the equation of radiative transfer | |
Jose et al. | Speed‐of‐sound compensated photoacoustic tomography for accurate imaging | |
Johnstonbaugh et al. | A deep learning approach to photoacoustic wavefront localization in deep-tissue medium | |
US6205353B1 (en) | Time-resolved optical backscattering tomographic image reconstruction in scattering turbid media | |
US20040030255A1 (en) | Hybrid-dual-Fourier tomographic algorithm for a fast three-dimensional optical image reconstruction in turbid media | |
Na et al. | Transcranial photoacoustic computed tomography based on a layered back-projection method | |
CN111915733B (en) | LeNet network-based three-dimensional cone-beam X-ray luminescence tomography method | |
CN110151133B (en) | Mammary gland optical imaging device and method based on image segmentation and time-frequency information fusion | |
WO2017004851A1 (en) | Bioluminescence tomography reconstruction algorithm based on multi-task bayesian compressed sensing method | |
CN115024739B (en) | Method for measuring distribution of Getsiram parameter in organism and application | |
Qi et al. | Cross-sectional photoacoustic tomography image reconstruction with a multi-curve integration model | |
Zuo et al. | Spectral crosstalk in photoacoustic computed tomography | |
Shahid et al. | Feasibility of a generative adversarial network for artifact removal in experimental photoacoustic imaging | |
Lou et al. | Impact of nonstationary optical illumination on image reconstruction in optoacoustic tomography | |
An et al. | Application of machine learning method in optical molecular imaging: a review | |
Zhao et al. | Optical properties reconstruction in nonhomogeneous participating medium based on an improved sequential quadratic programming | |
Zalev et al. | Opto-acoustic image reconstruction and motion tracking using convex optimization | |
Hakimnejad et al. | Unsupervised photoacoustic tomography image reconstruction from limited-view unpaired data using an improved CycleGAN |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190625 |
|
RJ01 | Rejection of invention patent application after publication |