CN108846790A - A method of accelerating image reconstruction - Google Patents
A method of accelerating image reconstruction Download PDFInfo
- Publication number
- CN108846790A CN108846790A CN201810616651.1A CN201810616651A CN108846790A CN 108846790 A CN108846790 A CN 108846790A CN 201810616651 A CN201810616651 A CN 201810616651A CN 108846790 A CN108846790 A CN 108846790A
- Authority
- CN
- China
- Prior art keywords
- node
- component
- computer
- computer node
- jacobian matrix
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T1/00—General purpose image data processing
- G06T1/20—Processor architectures; Processor configuration, e.g. pipelining
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T1/00—General purpose image data processing
- G06T1/60—Memory management
Abstract
The invention discloses a kind of methods for accelerating image reconstruction.The method depending on the application classifies to data, the data of different purposes take different storage modes to be stored, to improve the reading and copying speed of data in image reconstruction, in image reconstruction process, Jacobian matrix is split, and all sub- Jacobian matrixs are assigned to each computer node, each computer node calculates the one-component of final result according to the sub- Jacobian matrix received, and then the component being calculated is transmitted to master computer node and merged by each node.To which the transmitting of extensive matrix to be converted to the transmitting of vector, the transmission efficiency of data is improved, and then improve the speed of image reconstruction.
Description
Technical field
The present invention relates to tissue detection field, in particular to a kind of method for accelerating image reconstruction.
Background technique
Fluorescent molecular tomography technology is play an important role in field of biomedical research, is widely used in cancer
Diagnosis, medicament research and development and gene expression are visually studied.
Most of biological tissues are the three-dimensional turbid media of high scattering, establish a high-precision and efficient calculating side
Method has great significance for the quantitative accuracy of fluorescent molecular tomography.Monte Carlo is a kind of based on random sampling procedure
Discrete statistical methods.Be compared to other methods, Monte Carlo method analog random geometry, boundary condition and
Photon transport process under optical parameter.Due to its wide applicability, Monte Carlo method is considered as that simulated photons transport
Most direct, the most effective and most believable method of actual physics process.Thus, it becomes the other specific application methods of evaluation
Goldstandard, such as diffusion approximation method.However, high-precision Monte Carlo simulation is time-consuming.It is strong when light transmits in the medium
Degree can increase with transmission range and exponentially property decays, and media size is bigger, in order to reach same high computational accuracy, covers special
Time required for Caro emulates is also the growth of exponentially property.Huge calculation amount often causes computer biggish negative
Load.
The essential characteristic of Monte Carlo method is emulated to stochastic problems, can effectively solve stochastic problems, very
It can more easily be solved to many insoluble stochastic problems of Deterministic Methods.Photon transmission in biological tissues
In the research of problem, Monte Carlo also illustrates that its advantage relative to other methods (such as diffusion approximation)；It is tight and flexible, no
It is limited by the optical characteristics organized, easy to use, simulation precision is high, and by the area research, person regards as goldstandard.
Several Monte-Carlo Simulation Methods based on historical path have been suggested.Such as：One kind is glimmering based on time resolution
The efficient Monte Carlo method of light, this method establish transmission photons weight and group by saving a large amount of photon path information
Knit the parsing relationship of optical parameter；A kind of efficient Monte Carlo side of the fluorescence Jacobian matrix method time-gated based on calculating
Method, this method for being the fluorogen that one kind can be different to two service life while imaging, in this approach, from excitaton source to spy
The background weight matrix for surveying device can be calculated by photon path and fluorescent absorption coefficient；In addition, there are also it is a kind of directly calculate it is glimmering
The decoupling fluorescence monte-Carlo model of light, by decoupling excitation and transmitting conversion process, by photon path and respective organization
Optical parameter links together with remitted fluorescence photon, and the method is suitable for any fluorescence distribution of any turbid media.More than
Several method all remain a large amount of photon path information, in these methods, the routing information of photon is Monte Carlo mould
Fit very important a part in image reconstruction process.
Image reconstruction process is to find optimal solution to keep target function value minimum, and image reconstruction process can be related to matrix
Calculating and some iterative algorithms, such as conjugate gradient method.As the photon number of simulation increases, the number of simulation source is continuous
Increase, the time of computer simulation is also being multiplied, and the data volume of generation is increasing, the calculating of data, transmission etc.
The required time is also increasingly longer, and the speed of the existing image rebuilding method based on monte-Carlo model, which has been unable to meet, to be needed
It asks.
Summary of the invention
The object of the present invention is to provide a kind of methods for accelerating image reconstruction, to improve the speed of image reconstruction.
To achieve the above object, the present invention provides following schemes：
A method of accelerating image reconstruction, described method includes following steps：
Original dimension, the initial optical parameter of default biological tissue, and determine the initialization information of exciting light sources, and will
The initialization information of the original dimension, the initial optical parameter and the exciting light sources is stored to computer host node
Literal register in video card；
Biology is established according to the initialization information of the original dimension, the initial optical parameter and the exciting light sources
Tissue model, and determine the voxel index of the Animal tissue model, and the voxel index is stored to computer host node
Texture register；
Computer host node carries out Monte Carlo simulation to the Animal tissue model, and it is corresponding to calculate each exciting light sources
Animal tissue model outgoing fluorescence intensity, and fluorescence intensity and the voxel are calculated according to each fluorescence intensity
The Jacobian matrix of the fluorescence coefficient of index；
The Jacobian matrix is divided into multiple sub- Jacobian matrixs by computer host node, and by multiple refined gram of sons
Than matrix allocation to multiple computer nodes；
Each computer node calculates the component of its corresponding ▽ f (x) according to its corresponding sub- Jacobian matrix
And HskComponent, and by the component and Hs of ▽ f (x)kComponent pass to computer host node；The computer host node is to more
The component of a ▽ f (x) carries out synthesis and to multiple HskComponent synthesized, the image after being rebuild, wherein f (x) is
Function to be solved, H are Hesse matrix, skIt is the direction of search of kth wheel iteration.
Optionally, described to store the voxel index to the texture register of computer host node, it specifically includes：
Three-D grain binding is carried out to the array of voxel index composition, the voxel index after being bound；
Voxel index after the binding is stored to texture register.
Optionally, each computer node calculates its corresponding ▽ f according to its corresponding sub- Jacobian matrix
(x) component and HskComponent, and by the component and Hs of ▽ f (x)kComponent pass to computer host node；The computer
Host node carries out synthesis to the component of multiple ▽ f (x) and to multiple HskComponent synthesized, the image after being rebuild, tool
Body includes：
Each computer node is according to its corresponding Jacobian matrix submatrix and formula ▽ f (x)=Hx+b=(JTJ+λ)
x-JTΔ D calculates ▽ f (x) in the one-component of the computer nodeWherein, n is computer node
Number, f (x) is function to be solved,X represents the vector that voxel index is formed by indexed sequential, and H is
Hesse matrix, H=JTJ+ λ, λ are regularization parameters, and b is vector composed by the probe value of detector；xkIt is conjugate gradient method
Kth wheel iteration result, it represents the vector that parameter value is formed by indexed sequential in each voxel in tissue model；Jn, (n=0,
1, i) it is assigned to the sub- Jacobian matrix of n computer node,Corresponding Jn, (n=0,1's, i) turns
Set matrix；Δ D is the difference of analog detection value and actual detection value；
All computer nodes are by the one-component of ▽ f (x)It passes to master computer node and sums,
Obtain ▽ f (x)；
Master computer node judges whether ▽ f (x) restrains, if so, stopping algorithm, after being rebuild according to ▽ f (x)
Image；If it is not, then calculating the direction of search s of conjugate gradient methodk, and result is broadcast to each computer node；
Each computer node is respectively according to formula Hsk=JTJskCalculate HskComponent；Wherein, skIt is searching for kth wheel iteration
Suo Fangxiang is a vector；
Each computer node is HskComponent be sent to master computer node and master computer node sum, obtain Hsk；
Master computer node is according to Hsk, using conjugate gradient method, acquire mk,gk+1,βk；gkFor optimal discriminant parameter, gk=
▽f(xk)；mkFor optimum search step-length,βkFor direction of search undated parameter,
Master computer node is according to mkAnd sk, utilize conjugate gradient method formula xk+1=xk+mkskTo calculate xk+1, and by xk+1
Each computer node is passed to, returns to each computer node according to its corresponding Jacobian matrix submatrix and formula ▽ f (x)
=Hx+b=(JTJ+λ)x-JTΔ D calculates the one-component of ▽ f (x)
Optionally, the Jacobian matrix is divided into multiple sub- Jacobian matrixs, and by multiple sub- Jacobi squares
Battle array is assigned to multiple computer nodes, further includes before：
A core of the CPU of computer host node opens up the first process, and the video card of computer host node is calculated
Fluorescence intensity copies to memory from video memory, and after duplication, another core of CPU opens up the second process, fluorescence intensity is written hard
Disk, the video card of computer host node continue to calculate the fluorescence of the corresponding Animal tissue model outgoing of next exciting light sources
Intensity.
The specific embodiment provided according to the present invention, the invention discloses following technical effects：
The invention discloses it is a kind of accelerate image reconstruction method, the method includes：Firstly, default biological tissue is first
Beginning size, initial optical parameter, and determine the initialization information of exciting light sources, and by the original dimension, the initial light
The initialization information for learning parameter and the exciting light sources stores the literal register into GPU；Animal tissue model is established, and
Its voxel index stores the voxel index to texture register；Then, computer host node is to the Animal tissue model
Monte Carlo simulation is carried out, the fluorescence intensity of the corresponding Animal tissue model outgoing of each exciting light sources is calculated, and is counted
Calculate Jacobian matrix；Finally, Jacobian matrix is divided into multiple sub- Jacobian matrixs, and multiple sub- Jacobian matrixs are distributed
To multiple computer nodes；By each computer node according to its corresponding sub- Jacobian matrix, it is corresponding to calculate its
The component and Hs of ▽ f (x)kComponent, and by the component and Hs of ▽ f (x)kComponent pass to computer host node；By computer
Host node carries out synthesis to the component of multiple ▽ f (x) and to multiple HskComponent synthesized, the image after being rebuild.This
Invention classifies to data according to purposes, and the data of different purposes take different storage modes to be stored, to improve figure
Jacobian matrix is split by the reading and copying speed of data in image reconstruction process in picture reconstruction, and all sons
Jacobian matrix is assigned to each computer node, and each computer node calculates most according to the sub- Jacobian matrix received
The one-component of termination fruit, then the component being calculated is transmitted to master computer node and merged by each node.To advise greatly
The transmitting of modular matrix is converted to the transmitting of vector, improves the transmission efficiency of data, and then improve the speed of image reconstruction.
The present invention allows a core of the CPU of master computer node to open up process during calculating Jacobian matrix, adjusts
Copy and the calculating of data are completed with video memory, meanwhile, another core of CPU also opens up process, having answered in last round of circulation
Make the routing information write-in hard disk in the fluorescence intensity and tissue of memory, after the completion of waiting until two cpu process, a simulation
End cycle.The optimization of this process-time stream realize data write-in and data copy it is parallel, improve and entirely simulated
The degree of parallelism of journey, so as to shorten simulated time.
Detailed description of the invention
It in order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, below will be to institute in embodiment
Attached drawing to be used is needed to be briefly described, it should be apparent that, the accompanying drawings in the following description is only some implementations of the invention
Example, for those of ordinary skill in the art, without any creative labor, can also be according to these attached drawings
Obtain other attached drawings.
Fig. 1 is a kind of flow chart of method for accelerating image reconstruction provided by the invention；
Fig. 2 is the flow chart of the image after a kind of being rebuild of method for accelerating image reconstruction provided by the invention；
Fig. 3 is that the process of the image after a kind of being rebuild of method for accelerating image reconstruction provided by the invention optimizes
Acceleration principle figure.
Fig. 4 is a kind of monte-Carlo model (a) of the specific embodiment of method for accelerating image reconstruction provided by the invention
And reconstruction result map (b).
Specific embodiment
The object of the present invention is to provide a kind of methods for accelerating image reconstruction, to improve the speed of image reconstruction.
In order to make the foregoing objectives, features and advantages of the present invention clearer and more comprehensible, with reference to the accompanying drawing and specific real
Mode is applied to be described in further detail invention.
As shown in Figure 1, a kind of method for accelerating image reconstruction, described method includes following steps：
Step 101, original dimension, the initial optical parameter of biological tissue are preset, and determines the initialization of exciting light sources
Information, and the initialization information of the original dimension, the initial optical parameter and the exciting light sources is stored to analytic accounting
Literal register in the GPU (Graphics Processing Unit, display chip or video card) of calculation machine node.
Step 102, according to the initialization information of the original dimension, the initial optical parameter and the exciting light sources
Animal tissue model is established, and determines the voxel index of the Animal tissue model, and the voxel index is stored to calculating
The texture register of owner's node；Specifically, carrying out three-D grain binding to the array of voxel index composition, bound
Voxel index afterwards；Voxel index after the binding is stored to texture register.
The present invention is when calling GPU to carry out the calculating of Jacobian matrix, the data that need to be only read, i.e., described
The initialization information of original dimension, the initial optical parameter and the exciting light sources is stored with literal register, thus
Accelerate the reading speed of read-only data；The array that the parameter of voxel index each in Animal tissue model forms is carried out texture to tie up
It is fixed, it is stored using texture register, relative to the index speed that data deposit global storage is accelerated to tissue model.Using
Page locking page in memory saves the routing information that simulation obtains, and page locking page in memory is located at CPU (CentralProcessing Unit/
Processor, central processing unit), with the data in mapped memory (mappedmemory) processing page locking page in memory, make this block
The existing memory address of memory, and have video memory address, GPU and CPU can directly access this block storage, accelerate routing information
Access speed.
Step 103, computer host node carries out Monte Carlo simulation to the Animal tissue model, calculates each exciting light
The fluorescence intensity of the corresponding Animal tissue model outgoing of light source, and according to each fluorescence intensity calculate fluorescence intensity with
The Jacobian matrix of the fluorescence coefficient of the voxel index；Specifically, a core of the CPU of computer host node open up first into
The fluorescence intensity that the video card of computer host node is calculated is copied to memory from video memory by journey, and after duplication, CPU is another
A core opens up the second process, hard disk is written in fluorescence intensity, the video card of computer host node continues to calculate next excitation light
The fluorescence intensity of the corresponding Animal tissue model outgoing in source.
The present invention allows a core of the CPU of master computer node to open up process during calculating Jacobian matrix, adjusts
Copy and the calculating of data are completed with video memory, meanwhile, another core of CPU also opens up process, having answered in last round of circulation
Make the routing information write-in hard disk in the fluorescence intensity and tissue of memory, after the completion of waiting until two cpu process, a simulation
End cycle.The optimization of this process-time stream realize data write-in and data copy it is parallel, improve and entirely simulated
The degree of parallelism of journey, so as to shorten simulated time.
Step 104, the Jacobian matrix is divided into multiple sub- Jacobian matrixs by computer host node, and by multiple institutes
It states sub- Jacobian matrix and is assigned to multiple computer nodes；
Step 105, each computer node calculates its corresponding ▽ f according to its corresponding sub- Jacobian matrix
(x) component and HskComponent, and by the component and Hs of ▽ f (x)kComponent pass to computer host node；The computer
Host node carries out synthesis to the component of multiple ▽ f (x) and to multiple HskComponent synthesized, the image after being rebuild,
In, f (x) is function to be solved, and H is Hesse matrix, skIt is the direction of search of kth wheel iteration.It specifically includes：
Each computer node is according to its corresponding Jacobian matrix submatrix and formula ▽ f (x)=Hx+b=(JTJ+λ)
x-JTΔ D calculates ▽ f (x) in the one-component of the computer nodeWherein, n is computer node
Number, f (x) is function to be solved,X represents the vector that voxel index is formed by indexed sequential, and H is
Hesse matrix, H=JTJ+ λ, λ are regularization parameters, and b is vector composed by the probe value of detector；xkIt is conjugate gradient method
Kth wheel iteration result, it represents the vector that parameter value is formed by indexed sequential in each voxel in tissue model；Jn, (n=0,
1, i) it is assigned to the sub- Jacobian matrix of n computer node,Corresponding Jn, (n=0,1's, i) turns
Set matrix；Δ D is the difference of analog detection value and actual detection value；
All computer nodes are by the one-component of ▽ f (x)It passes to master computer node and sums,
Obtain ▽ f (x)；
Master computer node judges whether ▽ f (x) restrains, if so, stopping algorithm, after being rebuild according to ▽ f (x)
Image；If it is not, then calculating the direction of search s of conjugate gradient methodk, and result is broadcast to each computer node；
Each computer node is respectively according to formula Hsk=JTJskCalculate HskComponent；Wherein, skIt is searching for kth wheel iteration
Suo Fangxiang is a vector；
Each computer node is HskComponent be sent to master computer node and master computer node sum, obtain Hsk；
Master computer node is according to Hsk, using conjugate gradient method, acquire mk,gk+1,βk；gkFor optimal discriminant parameter, gk=
▽f(xk)；mkFor optimum search step-length,βkFor direction of search undated parameter
Master computer node is according to mkAnd sk, utilize conjugate gradient method formula xk+1=xk+mkskTo calculate xk+1, and by xk+1
Each computer node is passed to, returns to each computer node according to its corresponding Jacobian matrix submatrix and formula ▽ f (x)
=Hx+b=(JTJ+ λ) x-JT Δ D calculates the one-component of ▽ f (x)
Specifically, as shown in Fig. 2, in figure, Noden (n=0,1, i) and represent n node, wherein 0 generation of Node
Table host node, other nodes are all child nodes；
2.1, each node is split and is assigned to Jacobian matrix according to the difference in source, each node is according to refined gram
Than the one-component that the submatrix and formula of matrix calculate ▽ f (x), i.e.,N is node serial number；Wherein,
▽ f (x)=Hx+b=(JTJ+λ)x-JTΔ D,F (x) is function to be solved, and x represents tissue model
In in each voxel parameter value press the vector of indexed sequential composition, H is Hesse matrix, H=JTJ+ λ, λ are regularization parameters, and b is
Vector composed by the probe value of detector；xkIt is the kth wheel iteration result of conjugate gradient method, it represents each body in tissue model
Parameter value presses the vector of indexed sequential composition in element；Jn, (n=0,1, i) it is assigned to n node Jacobian matrix
Submatrix,Corresponding Jn, the transposed matrix of (n=0,1, i)；Δ D is analog detection value and actual detection value
Difference；
2.2, calculated result is passed to host node and summed by all child nodes；
2.3, host node judges whether to reach the condition of convergence, stops algorithm if reaching；Continue (2.4) if not up to；
2.4, host node calculates the direction of search of conjugate gradient method, and result is broadcast to child node；
2.5, all nodes are according to Hsk=JTJskCalculate HskOne-component；skIt is the direction of search of kth wheel iteration, is
One vector；
2.6, all child nodes transmit the calculation to host node and sum in host node, obtain Hsk；
2.7, according to HskResult and conjugate gradient method formula, acquire m in host nodek,gk+1,βk；gk=▽ f
(xk)；Represent optimum search step-length；
2.8, host node is according to conjugate gradient method formula xk+1=xk+mkskTo calculate xk+1, return (2.3)；
(3) during image reconstruction, the present invention allows a core of CPU to open up process, and GPU is called to complete data
Duplication and calculating, meanwhile, another core of CPU also opens up process, the photon for having been copied to memory in last round of circulation
Hard disk is written in routing information in status information and tissue.Until two cpu process after the completion of, a simulation cycle terminates,
As shown in Figure 3.
The present invention also provides a specific embodiments, to illustrate method of the invention：
It is tested by emulation experiment and realizes method of the invention.Shown in designed monte-Carlo model such as Fig. 4 (a),
The monte-Carlo model is homogeneous model, wherein being respectively 0.2cm comprising three fluorescence coefficients-1, 0.4cm-1, 0.6cm-1It is glimmering
Optical wand.Optical parameter under exciting light and fluorescence bands is consistent, and quantum efficiency is set as 1, in simulations used in its
His optical parameter is as shown in table 1.The size of the model is 2.1cm × 2.1cm × 3.0cm, and the size of single voxel is 0.1cm,
Model contains 13230 voxels altogether.The region of green is reconstruction regions, is located among cylinder in the range of 14mm, includes altogether
4270 voxels.In reconstruction process, the present embodiment simulates 120 sources, covers 360 degree of angular regions, among cylinder
In the range of 7.5mm, 6 layers are evenly distributed on, every layer of 20 source；We choose 240 detectors, cover 360 degree of angular regions,
In the range of 20mm among cylinder, 10 layers are evenly distributed on, every layer of 24 detector.The number of photons that each source is simulated is equal
It is 2000000.
The optical parameter of the monte-Carlo model for simulating of table 1
Wherein, n is refractive index, μsFor scattering coefficient, μaFor absorption coefficient, g is anisotropy factor, μafFor fluorogen suction
Receive coefficient.
According to the disk read-write performance of the number of each node GPU equipment and performance and each node, distributed on three nodes
The number of light sources of required simulation is respectively 25,50,45.Fig. 4 (b) is three-dimensional reconstruction result, and table 2 is shown in the upper figure of parallel architecture
As reconstruction time.
The time of 2 image reconstruction of table
The specific embodiment provided according to the present invention, the invention discloses following technical effects：
Since image reconstruction process needs default many read-only data.And GPU reads the data on different memories
Reading speed be different, initialization information (position and direction) of the present invention photon, the optical parameter of tissue, model
The lesser information of size these information content be stored in literal register.The texture register of GPU can be accelerated to index speed, so
Texture binding is carried out using the parameter of each voxel, is stored using texture register, GPU reads literal register and texture deposit
The reading speed of device is faster than global register.Because the data that CPU is calculated are located at calculator memory, the data that GPU is calculated are located at
Video memory, it is not public for being located at the data of memory (memory bar) and being located at the data of video memory (video card), if CPU needs to read
The data of video memory then need that data are first transferred to memory from video memory and read again, and the data that GPU reads memory are same.And it adopts
Page locking page in memory is handled with mapped memory, the existing memory address of this block storage can be allowed, and have video memory address, i.e., this block is deposited
Data in reservoir both can be used directly by GPU, can also directly be used by CPU, in this way, can both will be used by GPU, also can
It is stored by the way of page locking page in memory by the data that CPU is used, accelerates reading speed.
Method of the invention, according to the actual weight and the Jacobian matrix, rebuilds the biological tissue described
Also Jacobian matrix is split according to source during image, the corresponding submatrix in each source, and all submatrixs
It is assigned to each node, each node calculates the one-component of final result according to the submatrix received, then each node
The component being calculated is transmitted to host node to merge.First by Jacobian matrix piecemeal, parallel computation is realized, this obviously can add
Fast calculating speed.But result is needed to be transmitted to host node after the completion of each child node calculates, it is refined in each round iteration
Gram can be updated than matrix, if the matrix-block that a certain node calculates is m × n, after calculating, can obtain one it is new
M * n matrix block, and obtain a dimension be m vector, if be passed back to host node is updated matrix, this need
The data volume to be transmitted is m × n, if transmitting is the vector being calculated, the data volume for needing to transmit is m, is realized
Accelerated by way of reducing data transmission capacity.
As shown in figure 3, the present invention is also optimized process-time, since this algorithm needs to be calculated using GPU,
Video memory is written into from memory in data so needing CPU, after the completion of GPU calculating, calculated result can be transmitted to memory, then by
CPU opens up process, and hard disk is written in the result in memory.If with the time flow in Fig. 3 (a), GPU will appear waiting when
Between, if being opened up process with the time flow of Fig. 3 (b) using another core of CPU and hard disk being written in the data of memory, then GPU
The waiting time is not had, GPU can spend in all time in calculating and output calculated result.(due to calculating of the invention
As a result generally there are tens G, so data can be taken a long time from memory write-in hard disk, this time rebuild entirely
Sizable ratio is accounted in journey).
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other
The difference of embodiment, the same or similar parts in each embodiment may refer to each other.
Specific examples are used herein to describe the principles and implementation manners of the present invention, the explanation of above embodiments
Method and its core concept of the invention are merely used to help understand, described embodiment is only that a part of the invention is real
Example is applied, instead of all the embodiments, based on the embodiments of the present invention, those of ordinary skill in the art are not making creation
Property labour under the premise of every other embodiment obtained, shall fall within the protection scope of the present invention.
Claims (4)
1. a kind of method for accelerating image reconstruction, which is characterized in that described method includes following steps：
Original dimension, the initial optical parameter of default biological tissue, and determine the initialization information of exciting light sources, and will be described
The initialization information of original dimension, the initial optical parameter and the exciting light sources is stored to the video card of master computer node
In literal register；
Biological tissue is established according to the initialization information of the original dimension, the initial optical parameter and the exciting light sources
Model, and determine the voxel index of the Animal tissue model, and the voxel index is stored to the line of master computer node
Manage register；
Computer host node carries out Monte Carlo simulation to the Animal tissue model, calculates the corresponding institute of each exciting light sources
The fluorescence intensity of Animal tissue model outgoing is stated, and fluorescence intensity and the voxel index are calculated according to each fluorescence intensity
Fluorescence coefficient Jacobian matrix；
The Jacobian matrix is divided into multiple sub- Jacobian matrixs by computer host node, and by multiple sub- Jacobi squares
Battle array is assigned to multiple computer nodes；
Each computer node calculates the component and Hs of its corresponding ▽ f (x) according to its corresponding sub- Jacobian matrixk
Component, and by the component and Hs of ▽ f (x)kComponent pass to computer host node；The computer host node is to multiple ▽
The component of f (x) carries out synthesis and to multiple HskComponent synthesized, the image after being rebuild, wherein f (x) is wait ask
Function is solved, H is Hesse matrix, skIt is the direction of search of kth wheel iteration.
2. a kind of method for accelerating image reconstruction according to claim 1, which is characterized in that described by the voxel index
It stores to the texture register of master computer node, specifically includes：
Three-D grain binding is carried out to the array of voxel index composition, the voxel index after being bound；
Voxel index after the binding is stored to texture register.
3. a kind of method for accelerating image reconstruction according to claim 1, which is characterized in that each computer node
According to its corresponding sub- Jacobian matrix, the component and Hs of its corresponding ▽ f (x) are calculatedkComponent, and by ▽ f (x) point
Amount and HskComponent pass to computer host node；The computer host node synthesizes simultaneously the component of multiple ▽ f (x)
To multiple HskComponent synthesized, the image after being rebuild specifically includes：
Each computer node is according to its corresponding Jacobian matrix submatrix and formula ▽ f (x)=Hx+b=(JTJ+λ)x-JT
Δ D calculates ▽ f (x) in the one-component of the computer nodeWherein, n is the volume of computer node
Number, f (x) is function to be solved,X represents the vector that voxel index is formed by indexed sequential, and H is
Hesse matrix, H=JTJ+ λ, λ are regularization parameters, and b is vector composed by the probe value of detector；xkIt is conjugate gradient method
Kth wheel iteration result, it represents the vector that parameter value is formed by indexed sequential in each voxel in tissue model；Jn, (n=0,
1 ..., i) it is assigned to the sub- Jacobian matrix of n computer node,Corresponding Jn, the transposed matrix of (n=0,1 ..., i)；
Δ D is the difference of analog detection value and actual detection value；
All computer nodes are by the one-component of ▽ f (x)It passes to master computer node and sums, obtain
▽f(x)；
Master computer node judges whether ▽ f (x) restrains, if so, stopping algorithm, the figure after being rebuild according to ▽ f (x)
Picture；If it is not, then calculating the direction of search s of conjugate gradient methodk, and result is broadcast to each computer node；
Each computer node is respectively according to formula Hsk=JTJskCalculate HskComponent；Wherein, skIt is the searcher of kth wheel iteration
To being a vector；
Each computer node is HskComponent be sent to master computer node and master computer node sum, obtain Hsk；
Master computer node is according to Hsk, using conjugate gradient method, acquire mk,gk+1,βk；gkFor optimal discriminant parameter, gk=▽ f
(xk)；mkFor optimum search step-length,βkFor direction of search undated parameter
Master computer node is according to mkAnd sk, utilize conjugate gradient method formula xk+1=xk+mkskTo calculate xk+1, and by xk+1Transmitting
To each computer node, each computer node is returned according to its corresponding Jacobian matrix submatrix and formula ▽ f (x)=Hx
+ b=(JTJ+λ)x-JTΔ D calculates the one-component of ▽ f (x)
4. a kind of method for accelerating image reconstruction according to claim 1, which is characterized in that by the Jacobian matrix point
Multiple sub- Jacobian matrixs are cut into, and multiple sub- Jacobian matrixs are assigned to multiple computer nodes, further include before：
A core of the CPU of computer host node opens up the first process, the fluorescence that the video card of computer host node is calculated
Intensity copies to memory from video card, and after duplication, another core of CPU opens up the second process, and hard disk is written in fluorescence intensity, meter
The video card for calculating owner's node continues to calculate the fluorescence intensity of the corresponding Animal tissue model outgoing of next exciting light sources.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810616651.1A CN108846790A (en) | 2018-06-15 | 2018-06-15 | A method of accelerating image reconstruction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810616651.1A CN108846790A (en) | 2018-06-15 | 2018-06-15 | A method of accelerating image reconstruction |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108846790A true CN108846790A (en) | 2018-11-20 |
Family
ID=64202949
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810616651.1A Pending CN108846790A (en) | 2018-06-15 | 2018-06-15 | A method of accelerating image reconstruction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108846790A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112528456A (en) * | 2019-09-18 | 2021-03-19 | 曙光信息产业(北京)有限公司 | Heterogeneous node computing system and method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20050013407A (en) * | 2003-07-28 | 2005-02-04 | 경 연 김 | Satic image reconstruction algorithm with adaptive mesh-grouping method in electrical impedance tomography |
| CN102663207A (en) * | 2012-04-28 | 2012-09-12 | 浪潮电子信息产业股份有限公司 | Method for quickening solving of mesoscopic scale physical problem by using graphic processing unit (GPU) |
| CN103356170A (en) * | 2013-05-24 | 2013-10-23 | 天津大学 | Quick Monte Carlo imaging method for reconstructing optical parameter of tissue with heteroplasmon |
| US20160055633A1 (en) * | 2013-04-17 | 2016-02-25 | DIXIT S.r.I. | Method for the automatic recognition of anatomical structures in images obtained by positron emission tomography, system and computer program for performing said method |
-
2018
- 2018-06-15 CN CN201810616651.1A patent/CN108846790A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20050013407A (en) * | 2003-07-28 | 2005-02-04 | 경 연 김 | Satic image reconstruction algorithm with adaptive mesh-grouping method in electrical impedance tomography |
| CN102663207A (en) * | 2012-04-28 | 2012-09-12 | 浪潮电子信息产业股份有限公司 | Method for quickening solving of mesoscopic scale physical problem by using graphic processing unit (GPU) |
| US20160055633A1 (en) * | 2013-04-17 | 2016-02-25 | DIXIT S.r.I. | Method for the automatic recognition of anatomical structures in images obtained by positron emission tomography, system and computer program for performing said method |
| CN103356170A (en) * | 2013-05-24 | 2013-10-23 | 天津大学 | Quick Monte Carlo imaging method for reconstructing optical parameter of tissue with heteroplasmon |
Non-Patent Citations (1)
| Title |
|---|
| XU JIANG 等: "《Accelerating fDOT image reconstruction based on path-history fluorescence Monte Carlo model by using three-level parallel architecture》", 《OPTICS EXPRESS》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112528456A (en) * | 2019-09-18 | 2021-03-19 | 曙光信息产业(北京)有限公司 | Heterogeneous node computing system and method |
| CN112528456B (en) * | 2019-09-18 | 2024-05-07 | 曙光信息产业(北京)有限公司 | Heterogeneous node computing system and method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Gibson et al. | NiftyNet: a deep-learning platform for medical imaging | |
| Konukoglu et al. | Extrapolating glioma invasion margin in brain magnetic resonance images: Suggesting new irradiation margins | |
| JP7438230B2 (en) | Determining model parameters using a predictive model | |
| Castaño-Díez et al. | Performance evaluation of image processing algorithms on the GPU | |
| US9355492B2 (en) | System, method, and computer program product for utilizing a wavefront path tracer | |
| Wang et al. | Modeling and rendering of heterogeneous translucent materials using the diffusion equation | |
| Wong et al. | Tumor growth prediction with reaction-diffusion and hyperelastic biomechanical model by physiological data fusion | |
| Jeong et al. | Interactive visualization of volumetric white matter connectivity in DT-MRI using a parallel-hardware Hamilton-Jacobi solver | |
| Wang et al. | Real‐time rendering of heterogeneous translucent objects with arbitrary shapes | |
| US9189870B2 (en) | Method, computer readable medium and system for tomographic reconstruction | |
| Maximo et al. | Hardware‐Assisted Projected Tetrahedra | |
| Du et al. | Interactive segmentation of clustered cells via geodesic commute distance and constrained density weighted Nyström method | |
| Asllanaj et al. | Light propagation in biological tissue | |
| Pérez-García et al. | A self-supervised learning strategy for postoperative brain cavity segmentation simulating resections | |
| CN108846790A (en) | A method of accelerating image reconstruction | |
| Balasubramanian et al. | Near-isometric flattening of brain surfaces | |
| Jungblut et al. | Inertia based filtering of high resolution images using a GPU cluster | |
| Du Bois d’Aische et al. | Efficient multi-modal dense field non-rigid registration: alignment of histological and section images | |
| Baxter et al. | Directed acyclic graph continuous max-flow image segmentation for unconstrained label orderings | |
| Schmitz et al. | Efficient and quality contouring algorithms on the GPU | |
| Amiraghdam et al. | LOCALIS: Locally‐adaptive Line Simplification for GPU‐based Geographic Vector Data Visualization | |
| López et al. | An efficient GPU approach for designing 3D cultural heritage information systems | |
| Lyu et al. | Adaptive channel-modulated personalized federated learning for magnetic resonance image reconstruction | |
| Lo et al. | GPU-accelerated Monte Carlo simulation for photodynamic therapy treatment planning | |
| Köhn et al. | A GPU-based fiber tracking framework using geometry shaders |
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: 20181120 |
|
| RJ01 | Rejection of invention patent application after publication |