CN107886519A - Multichannel chromatogram three-dimensional image fast partition method based on CUDA - Google Patents
Multichannel chromatogram three-dimensional image fast partition method based on CUDA Download PDFInfo
- Publication number
- CN107886519A CN107886519A CN201710977831.8A CN201710977831A CN107886519A CN 107886519 A CN107886519 A CN 107886519A CN 201710977831 A CN201710977831 A CN 201710977831A CN 107886519 A CN107886519 A CN 107886519A
- Authority
- CN
- China
- Prior art keywords
- block
- weight
- label
- gpu
- image
- 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
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/162—Segmentation; Edge detection involving graph-based methods
-
- 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
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20072—Graph-based image processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30016—Brain
Abstract
The invention discloses a kind of multichannel chromatogram three-dimensional image fast partition method based on CUDA.On the basis of the present invention carries out multichannel chromatogram tag fusion algorithm with the sparse representation method of block, the union of specific ROI region is extracted and carries out appropriate expansion, and establishes index.Weight and label are asked in GPU, processing in need point parallel operation obtain block vector, preselected, block the regularization of block, weight calculation and determination label.Simultaneously using NVIDIA unified calculation equipment framework, CUDA thread block parallel computation processes are converted into, make subprogram accelerate to perform in GPU, in the case where ensureing certain segmentation precision, considerably improve the arithmetic speed of algorithm.The inventive method parallel section is mainly in solution l1Norm optimization problem, as long as many be related to solving l1The problem of norm optimization, can more it be quickly coped with by the inventive method, so the present invention more has universality.
Description
Technical field
The invention belongs to Three Dimensional Medical Visualization field, is specifically the multichannel chromatogram segmentation side based on rarefaction representation
Method, and accelerated parallel using CUDA.
Background technology
Medical image segmentation is the important foundation of medical image analysis, and it refers to draw medical image according to certain criterion
Be divided into several disjoint regions so that the feature of pixel is similar in the same area and different zones in pixel feature not
Together, and to region (Region of Interest, ROI) interested extracted, the purpose is to will be some in medical image
Region, organ with particular meaning, anatomical structure are split, and carry out feature extraction and parameter measurement to target, to face
Bed and pathological research provide reliable foundation, and auxiliary doctor makes more accurate diagnosis.
The traditional method of medical image segmentation have region-growing method, watershed, active contour, level set, Markov with
Airport model etc., but MRI is split with these traditional dividing methods, it is difficult to meet present segmentation precision
Requirement.Being now based on the dividing method of multichannel chromatogram, Successful utilization is into Three Dimensional Medical Visualization, and segmentation effect
Preferably, label in this method propagates (fusion) process, all segmentations that will be obtained be combined in certain method and then
Final label is obtained, there is many combined methods, such as most ballots (Majority Voting, MV), local weighted method at present
Block method of weighting (Patch-based Method, PBM) of (Local Weighted Voting, LWV), non-local mean etc.,
In order to obtain largely good candidate point to carry out tag fusion, the method (MV, LWV) based on one-to-one correlation is due to registration
Situations such as error, possibly can not then obtain more preferable Candidate Set, but PBM methods are not converted using complicated non-rigid, therefore
Input picture also differs surely registering well.Later again had SPBM tag fusion algorithms, using in atlas image block it is dilute
The block in input picture is rebuild in thin linear combination, and then, the weight reconstructed coefficients of the block calculated are used to calculate simultaneously
The attachment structure of figure and the weight of figure.Obtained segmentation precision is higher, but calculation cost is too big, operation it is slower, consumed when
Between it is too many.
Many dividing method amounts of calculation are all very big, when especially handling mass data collection, before the surgery and hand
It must be fast and accurately, so as to useful in clinical diagnosis that the view data obtained during art, which carries out processing,.It is now specific
The data that patient can use increase steadily, when split etc. processing to medical image, when being often required to expend substantial amounts of
Between, therefore the rate request more and more higher of brain image segmentation algorithm.
The existing multichannel chromatogram dividing method based on rarefaction representation is described as follows：
(1) it is registering.Each collection of illustrative plates (atlas) image is mapped to target image with specific mapping mode, then by it
Corresponding label is broadcast to label of the target image as target image in the same way.
Given target image T, that is, image to be split, define I={ Is| s=1 ..., } and L={ Ls| s=
1 ..., } represent label corresponding to the collection of illustrative plates core of N number of registration respectively, each pixel in target image is designated as x (x ∈ T),
A specific search neighborhood block in collection of illustrative plates is designated as P (x), defines column vectorAnd row
Vector Block vector respectively in collection of illustrative plates and target image, thus calculates the similitude of two blocks.With
Tuple b=(s, y) is represented around the index s and block central point y of atlas image.Therefore, each atlas image blockCan be with
It is reduced to βb(b=1 ..., Q), wherein Q=N × | n (x) | it is the total number of atlas image block.Also will simultaneouslyIt is simplified to
(2) tag fusion.The purpose of tag fusion is in order to automatically by label LTIt is mapped to target image., all collection of illustrative plates
The weight vectors of block areThe present invention uses the graph structure weight calculation side based on rarefaction representation
Method：
Parameter lambda controls the active force of sparse constraint, and B is one by all column vectorsThe matrix combined,
Before sparse coefficient is optimized, block vector needs to be normalized into unit vector.Assuming that there are M possible labels in atlas image
{l1..., lm..., lM, the label of the tissue points x among target image can be determined by following formula：
Wherein LbRepresent atlas image block βbThe label of central point,
Dirac functions
With the rapid development of computer hardware technology, graphics processing unit GPU computing capability is in geometry level in recent years
Number increases, and its outstanding Floating-point Computation ability, streaming parallel architecture and flexible programmability are becoming increasingly popular.CUDA is
A kind of computing architecture designed exclusively for raising parallel program development efficiency, when building high performance application, CUDA framves
Structure can give full play to GPU powerful computing function.
The content of the invention
The technical problems to be solved by the invention are：When three-dimensional image based on multichannel chromatogram is split, with block
On the basis of sparse representation method carries out multichannel chromatogram tag fusion algorithm, in order to reduce the time of clinical diagnosis wait, ensureing
On the premise of segmentation precision, the speed of service of the algorithm is improved.
With the rapid development of computer hardware technology, graphics processing unit GPU computing capability is in geometry level in recent years
Number increases, and its outstanding Floating-point Computation ability, streaming parallel architecture and flexible programmability are becoming increasingly popular.Therefore,
For above-mentioned demand, the present invention proposes to accelerate in GPU hardware, is carried out simultaneously using an instrument CUDA of GPU concurrent operations
Row Accelerating running partitioning algorithm, so as to greatly improve brain image segmentation speed.
The present invention mainly reduces multichannel chromatogram sliced time, utilizes GPU powerful computing function so that program parallelization is transported
OK, reduce calculating and expend the time.
The technical solution adopted for the present invention to solve the technical problems comprises the following steps：
Specifically implement according to following steps：
Step 1. counts areal interested and records the label sequence number first, only considers an area-of-interest
(hippocampus), and other ROI and non-brain structural region separate, therefore split whole brain region and reform into hippocampus area and non-
Hippocampus area.
Step 2. calculates the block Duplication (Dice) of image to be split and each atlas, two area-of-interest O1And O2Weight
Folded rate is defined as Dice (O1, O2).
Wherein | | represent the volume of specific region of interest ROI.
N-1 best collection of illustrative plates (BestAtlas) are picked out according to block Duplication and are ranked up.BestAtlas is carried out
Morphologic expansion process, the union of collection of illustrative plates label control the scale of image as initialization mask with this.
Step 3. selects the multichannel chromatogram dividing method (SPBM) based on rarefaction representation, and OpenMP is carried out on CPU and is transported parallel
2500 points are calculated, obtain weight and label vector, it is specific as follows：
(a) number of process points required for calculating, and obtain the coordinate each put.Definition various parameters matrix is simultaneously initial
Change, with cudaMalloc methods on GPU storage allocation.
(b) the brightness normalization inside image is carried out, that is, by each block regularization, to ensure that what image was subject to comes from
Outside change minimizes.
(c) before weight calculation is carried out, similarity is directly set to 0 less than the weight of the block of given threshold, excluded not
Influence of the similar block to overall weight, although this influence very little, because the number of block is very more, so as to reduce its long-tail
Effect.
Block similarity calculating method：
OrderWithFor block PT(x) average and variance,WithFor blockAverage
And variance, definition：
S (x, y) is x and y similarity, from above formula, if average and variance are closer, s (x, y) value closer to 1,
Otherwise closer to 0.It is considered as invalid block when therefore working as s (x, y)≤θ, θ is a threshold value, takes θ=0.05. here
(d) weight uses the quadratic sum SSD of gray scale difference, and it is defined as follows：
Wherein, F (x, y) and R (x, y) represents that floating image (i.e. collection of illustrative plates) F and reference picture (i.e. target image) R exist respectively
The gray value at (x, y) place, N represent total pixel number.
By the weight of each candidate point by sorting from small to large and doing normalized, weight normalizes formula：
M therein is value minimum in the weight W of all available points, and ε is constant, and its value is 0.01.
The weight and label for obtaining point every time are just placed in the matrix of predefined, are circulated 2500 times.Then use
CudaMemcpy methods pass to data on GPU internal memories from CPU internal memories.
Graph structure weight calculation of the step 4. based on rarefaction representation
Combined to rebuild the block in input picture by using the sparse linear of block in atlas image, then, calculated
The weight reconstructed coefficients of block be used to the attachment structure of reckoning figure simultaneously and the weight of figure, that is, solution (formula 1), this is one
l1Norm optimization problem, there is the computing much on matrix to be all placed on GPU parallel operation in this process.
In this process, the copy on data between CPU and GPU is as described below：
With cudaMemcpy functions the weighted data after sparse from video card memory copying to CPU in, CPU parallel computations
Each label weight summation effectively corresponding to neighborhood of a point point.After having handled 2500*5 point, remaining point is placed on CPU
Parallel processing, the most thick maximum label of weight of finding out is final label.
Using CPU/GPU isomeries, OpenMP is opened at CPU ends and calculates the weight each put and label, and by these data
It is incorporated into a matrix, is processed into the form for being appropriate for thread distribution, these matrixes are then copied to computing in GPU.
This place mainly solves l1Normal form optimization problem, it is related to many mathematical operations, it is necessary to which many temporal calculations, ours is main
Work is exactly to be optimized in this place.
The inventive method has the advantage that and beneficial outcomes are：
The present invention is indexed by establishing, and is reduced and is repeated to take block vectorial
We only carry out two class segmentations in an experiment --- hippocampus and non-hippocampus, by the union of specific ROI region
Extract and carry out appropriate expansion, and establish index.
Weight and label are asked in GPU, processing in need point parallel operation obtain block vector, block it is preselected,
Regularization, weight calculation and the determination label of block.
Although establishing the index stage needs to expend some times, carrying for a large amount of repeatable block vectors is largely reduced
Take, and ask weight and label all in GPU parallel computations, it is thus a little quicker than in CPU operations.
The present invention utilizes NVIDIA unified calculation equipment framework, is converted into CUDA thread block parallel computation processes, makes portion
Branch accelerates to perform in GPU, in the case where ensureing certain segmentation precision, considerably improves the arithmetic speed of algorithm.
The inventive method parallel section is mainly in solution l1Norm optimization problem, as long as many be related to solving l1Norm optimization
The problem of, can more it be quickly coped with by the inventive method, so the present invention more has universality.
Brief description of the drawings
Fig. 1 is the multichannel chromatogram partitioning algorithm flow chart based on rarefaction representation；
Fig. 2 is the parallel operational flow diagram of SPBM tag fusion algorithms；
Fig. 3 is the key data internal memory schematic diagram on each thread；
Fig. 4 is data global storage schematic diagram
Embodiment
The invention will be further described with reference to the accompanying drawings and examples.
As Figure 1-4, step 1. counts areal interested and records the label sequence number first, only considers one
Individual area-of-interest (hippocampus), and other ROI and non-brain structural region separate, therefore split whole brain region and reform into
Hippocampus area and non-hippocampus area.
Step 2. calculates the block Duplication (Dice) of image to be split and each atlas, two area-of-interest O1And O2Weight
Folded rate is defined as Dice (O1, O2).
Wherein | | represent the volume of specific region of interest ROI.
N-1 best collection of illustrative plates (BestAtlas) are picked out according to block Duplication and are ranked up.BestAtlas is carried out
Morphologic expansion process, the union of collection of illustrative plates label control the scale of image as initialization mask with this.
Step 3. selects the multichannel chromatogram dividing method (SPBM) based on rarefaction representation, and OpenMP is carried out on CPU and is transported parallel
2500 points are calculated, obtain weight and label vector, it is specific as follows：
(a) number of process points required for calculating, and obtain the coordinate each put.Definition various parameters matrix is simultaneously initial
Change, with cudaMalloc methods on GPU storage allocation.
(b) the brightness normalization inside image is carried out, that is, by each block regularization, to ensure that what image was subject to comes from
Outside change minimizes.
(c) before weight calculation is carried out, similarity is directly set to 0 less than the weight of the block of given threshold, excluded not
Influence of the similar block to overall weight, although this influence very little, because the number of block is very more, so as to reduce its long-tail
Effect.
Block similarity calculating method：
OrderWithFor block PT(x) average and variance,WithFor blockAverage
And variance, definition：
S (x, y) is x and y similarity, from above formula, if average and variance are closer, s (x, y) value closer to 1,
Otherwise closer to 0.It is considered as invalid block when therefore working as s (x, y)≤θ, θ is a threshold value, takes θ=0.05. here
(d) weight uses the quadratic sum SSD of gray scale difference, and it is defined as follows：
Wherein, F (x, y) and R (x, y) represents that floating image (i.e. collection of illustrative plates) F and reference picture (i.e. target image) R exist respectively
The gray value at (x, y) place, N represent total pixel number.
By the weight of each candidate point by sorting from small to large and doing normalized, weight normalizes formula：
M therein is value minimum in the weight W of all available points, and ε is constant, and its value is 0.01.
The weight and label for obtaining point every time are just placed in the matrix of predefined, are circulated 2500 times.Then use
CudaMemcpy methods pass to data on GPU internal memories from CPU internal memories.
Graph structure weight calculation of the step 4. based on rarefaction representation
Combined to rebuild the block in input picture by using the sparse linear of block in atlas image, then, calculated
The weight reconstructed coefficients of block be used to the attachment structure of reckoning figure simultaneously and the weight of figure, that is, solution (formula 1), this is one
l1Norm optimization problem, there is the computing much on matrix to be all placed on GPU parallel operation in this process.
In this process, the copy on data between CPU and GPU is as described below：
With cudaMemcpy functions the weighted data after sparse from video card memory copying to CPU in, CPU parallel computations
Each label weight summation effectively corresponding to neighborhood of a point point.After having handled 2500*5 point, remaining point is placed on CPU
Parallel processing, the most thick maximum label of weight of finding out is final label.
Using CPU/GPU isomeries, OpenMP is opened at CPU ends and calculates the weight each put and label, and by these data
It is incorporated into a matrix, is processed into the form for being appropriate for thread distribution, these matrixes are then copied to computing in GPU.
This place mainly solves l1Normal form optimization problem, it is related to many mathematical operations, it is necessary to which many temporal calculations, ours is main
Work is exactly to be optimized in this place.
Embodiment 1：
1. the present invention realizes the acceleration of multichannel chromatogram tag fusion algorithm based on CUDA frameworks using C language, we use
COMPUTER PARAMETER have：
(1) processor：The cores of Intel Xeon (to strong) E5-1620v4@3.50GHz tetra-；
(2) video card：DVIDIA Quadro K620 (2GB/ Hewlett-Packards)；
(3) system installation internal memory：16GB (Hynix DDR42400MHz)；
(4) development environment：Visual Studio 2013；
(5) CUDA versions：8.0
2. the key step realized
As shown in Fig. 2 this invention is the multichannel chromatogram partitioning algorithm based on rarefaction representation, parallel operation part mainly exists
On SPBM tag fusion algorithms.It is the specific descriptions that SPBM tag fusion algorithms are run parallel below.
Label corresponding to each putting is initialized, and calculates total number 12982 a little, and the seat of point is calculated required for obtaining
Mark is stored as three-dimensional vector array；
Matrix is defined with VNL and is initialized, and is distributed from CPU copied in GPU data with cudaMalloc functions first
Deposit, and setting copies the matrix size used in cpu data to from GPU.
As shown in Figures 3 and 4, OpenMP is opened on CPU to carry out parallel, due to needing to handle 12982 points altogether here,
Each 2500 points of parallel processing, therefore outer loop 5 times, interior loop number=2500 time, every time calling
SPBMonOnePoint_hhl () function, the function can obtain the weight and label of each point, and we use block-based think of
Think, each point extracts a small cubes as central point, equivalent to block.Then by the Data Integration of this 2500 points to pair
In the matrix answered, with cudaMemcpy functions by data copy into GPU internal memories.For example matrix corresponding to block vector stores, if
It is sizeof (double) * PatchLength* (50*number), PatchLength=125 here to put size, thus
Run time can be reduced, reduces biography value is carried out between CPU and GPU as far as possible；
Next, combined by using the sparse linear of block in atlas image to rebuild the block in input picture, then, meter
The weight reconstructed coefficients of the block calculated are used to the attachment structure of reckoning figure simultaneously and the weight of figure, that is, solution (formula 1), this
It is a l1Norm optimization problem, there is the computing much on matrix in this process.When starting kernel function, each thread block
Thread_N=16 thread is opened, but the thread block opened every time is not all, such as, opened in matrix multiplication concurrent operation
Opening thread block number is：
Block_N=((number)/thread_N)+(((number) %thread_N)1:0), in matrix subtraction simultaneously
Thread block number is opened during row computing is：
Block_N=((number*PatchLength)/thread_N)+(((number*PatchLength) %
thread_N)1:0),
The if-else execution time is adapted to a large amount of parallel data meters equivalent to total execution time that each branch deploys, CUDA
Calculate, branch judges that its concurrency can be destroyed.If-else efficiency is very bad in CUDA, thus we if-else structures all
Make into if judgement, when seeking the maximum of two numbers in addition, using formula max=(| a+b |+| a-b |)/2, can so subtract significantly
Few branch.Check whether it restrains in renewal weight, we do not go to judge whether to be less than error amount, but are ensureing to restrain
On the premise of precision, fixed cycle-index is rule of thumb set, and will can thus avoid Break sentences, to improve
The concurrency of program.
Afterwards with cudaMemcpy functions the weighted data after sparse from video card memory copying to CPU in, CPU is simultaneously
Row calculates each label weight summation effectively corresponding to neighborhood of a point point.Then to remaining point parallel processing on CPU, most
It is final label to find out the maximum label of weight afterwards.
Claims (1)
1. the multichannel chromatogram three-dimensional image fast partition method based on CUDA, it is characterised in that comprise the following steps：
Step 1. counts areal interested and records the label sequence number first, due to only considering a region of interest
Domain, and other ROI and non-brain structural region separate, therefore split whole brain region and reform into hippocampus area and non-hippocampus
Area；
Step 2. calculates the block Duplication (Dice) of image to be split and each atlas, two area-of-interest O1And O2Duplication
It is defined as Dice (O1, O2).
Wherein | | represent the volume of specific region of interest ROI；
N-1 best collection of illustrative plates are picked out according to block Duplication and are ranked up；Morphologic expansion process is carried out to best collection of illustrative plates,
The union of collection of illustrative plates label controls the scale of image as initialization mask with this；
Step 3. selects the multichannel chromatogram dividing method based on rarefaction representation, and OpenMP 2500 points of concurrent operation are carried out on CPU,
Weight and label vector are obtained, it is specific as follows：
(a) number of process points required for calculating, and obtain the coordinate each put；Define various parameters matrix and initialize, use
CudaMalloc methods storage allocation on GPU；
(b) carry out the brightness normalization inside image, that is, by each block regularization, so that it is guaranteed that image be subject to from outer
The change in portion minimizes；
(c) before weight calculation is carried out, similarity is directly set to 0 less than the weight of the block of given threshold, excluded dissimilar
Influence of the block to overall weight；
Block similarity calculating method：
OrderWithFor block PT(x) average and variance,WithFor blockAverage and side
Difference, definition：
S (x, y) is x and y similarity, and from above formula, if average and variance are closer, s (x, y) value is closer to 1, otherwise
Closer to 0；It is considered as invalid block when therefore working as s (x, y)≤θ, θ is a threshold value, takes θ=0.05. here
(d) weight uses the quadratic sum SSD of gray scale difference, and it is defined as follows：
Wherein, F (x, y) and R (x, y) represents the gray value of floating image F and reference image R at (x, y) place respectively, and N represents total
Pixel number；
By the weight of each candidate point by sorting from small to large and doing normalized, weight normalizes formula：
M therein is value minimum in the weight W of all available points, and ε is constant, and its value is 0.01；
The weight and label for obtaining point every time are just placed in the matrix of predefined, are circulated 2500 times；
Graph structure weight calculation of the step 4. based on rarefaction representation
Combined by using the sparse linear of block in atlas image to rebuild the block in input picture, then, the block calculated
Weight reconstructed coefficients be used to the attachment structure of reckoning figure simultaneously and the weight of figure, that is, solution (formula 1), this is a l1Model
Number optimization problem, there is the computing much on matrix to be all placed on GPU parallel operation in this process.
In this process, the copy on data between CPU and GPU is as described below：
With cudaMemcpy functions the weighted data after sparse from video card memory copying to CPU in, CPU parallel computations are each
Label weight summation corresponding to effective neighborhood of a point point.After having handled 2500*5 point, remaining point is placed on parallel on CPU
Processing, the most thick maximum label of weight of finding out is final label.
Using CPU/GPU isomeries, OpenMP is opened at CPU ends and calculates the weight each put and label, and by these Data Integrations
Into a matrix, the form for being appropriate for thread distribution is processed into, these matrixes are then copied to computing in GPU.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710977831.8A CN107886519A (en) | 2017-10-17 | 2017-10-17 | Multichannel chromatogram three-dimensional image fast partition method based on CUDA |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710977831.8A CN107886519A (en) | 2017-10-17 | 2017-10-17 | Multichannel chromatogram three-dimensional image fast partition method based on CUDA |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107886519A true CN107886519A (en) | 2018-04-06 |
Family
ID=61781896
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710977831.8A Pending CN107886519A (en) | 2017-10-17 | 2017-10-17 | Multichannel chromatogram three-dimensional image fast partition method based on CUDA |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107886519A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110288603A (en) * | 2019-05-22 | 2019-09-27 | 杭州电子科技大学 | Semantic segmentation method based on efficient convolutional network and convolution condition random field |
| CN113222979A (en) * | 2021-06-01 | 2021-08-06 | 长春理工大学 | Multi-map-based automatic skull base foramen ovale segmentation method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102622749A (en) * | 2012-02-22 | 2012-08-01 | 中国科学院自动化研究所 | Automatic segmentation method of three-dimensional magnetic resonance image of brain structure |
| CN102646268A (en) * | 2012-02-22 | 2012-08-22 | 中国科学院自动化研究所 | Magnetic resonance image brain structure automatic dividing method based on statistics multi-map registration optimization |
| CN104063714A (en) * | 2014-07-20 | 2014-09-24 | 詹曙 | Fast human face recognition algorithm used for video monitoring and based on CUDA parallel computing and sparse representing |
| US8977029B2 (en) * | 2012-08-24 | 2015-03-10 | Siemens Aktiengesellschaft | Method and system for multi-atlas segmentation of brain computed tomography image data |
| CN105739951A (en) * | 2016-03-01 | 2016-07-06 | 浙江工业大学 | GPU-based L1 minimization problem fast solving method |
-
2017
- 2017-10-17 CN CN201710977831.8A patent/CN107886519A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102622749A (en) * | 2012-02-22 | 2012-08-01 | 中国科学院自动化研究所 | Automatic segmentation method of three-dimensional magnetic resonance image of brain structure |
| CN102646268A (en) * | 2012-02-22 | 2012-08-22 | 中国科学院自动化研究所 | Magnetic resonance image brain structure automatic dividing method based on statistics multi-map registration optimization |
| US8977029B2 (en) * | 2012-08-24 | 2015-03-10 | Siemens Aktiengesellschaft | Method and system for multi-atlas segmentation of brain computed tomography image data |
| CN104063714A (en) * | 2014-07-20 | 2014-09-24 | 詹曙 | Fast human face recognition algorithm used for video monitoring and based on CUDA parallel computing and sparse representing |
| CN105739951A (en) * | 2016-03-01 | 2016-07-06 | 浙江工业大学 | GPU-based L1 minimization problem fast solving method |
Non-Patent Citations (2)
| Title |
|---|
| 孟晓林等: "基于CUDA的快速三维医学图像分割", 《中国医学物理学杂志》 * |
| 郭其淼: "多图谱医学图像分割方法研究及应用", 《中国优秀硕士学位论文全文数据库 信息科技辑（月刊）计算机软件及计算机应用》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110288603A (en) * | 2019-05-22 | 2019-09-27 | 杭州电子科技大学 | Semantic segmentation method based on efficient convolutional network and convolution condition random field |
| CN113222979A (en) * | 2021-06-01 | 2021-08-06 | 长春理工大学 | Multi-map-based automatic skull base foramen ovale segmentation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11508076B2 (en) | Learning rigidity of dynamic scenes for three-dimensional scene flow estimation | |
| JP7143008B2 (en) | Medical image detection method and device based on deep learning, electronic device and computer program | |
| US20220044412A1 (en) | Identifying image segmentation quality using neural networks | |
| CN110574044A (en) | Method and apparatus for enhancing binary weighted neural networks using dependency trees | |
| CN113571203B (en) | Multi-center federal learning-based brain tumor prognosis survival prediction method and system | |
| CN110637307A (en) | Method and apparatus for enhancing neural networks using pairs of binary tensors and scaling factors | |
| CN110838125B (en) | Target detection method, device, equipment and storage medium for medical image | |
| CN112927217B (en) | Thyroid nodule invasiveness prediction method based on target detection | |
| US20090092299A1 (en) | System and Method for Joint Classification Using Feature Space Cluster Labels | |
| CN112085714B (en) | Pulmonary nodule detection method, model training method, device, equipment and medium | |
| WO2021073120A1 (en) | Method and device for marking lung area shadows in medical image, server, and storage medium | |
| US20060018549A1 (en) | System and method for object characterization of toboggan-based clusters | |
| Shi et al. | Automatic segmentation of cardiac magnetic resonance images based on multi-input fusion network | |
| Xiang et al. | Turbopixel segmentation using eigen-images | |
| CN111462060A (en) | Method and device for detecting standard section image in fetal ultrasonic image | |
| CN112132815B (en) | Pulmonary nodule detection model training method, detection method and device | |
| WO2022032824A1 (en) | Image segmentation method and apparatus, device, and storage medium | |
| CN107886519A (en) | Multichannel chromatogram three-dimensional image fast partition method based on CUDA | |
| CN110992310A (en) | Method and device for determining partition where mediastinal lymph node is located | |
| Ali et al. | GPU-based segmentation of dental x-ray images using active contours without edges | |
| CN112164447A (en) | Image processing method, device, equipment and storage medium | |
| Quesada-Barriuso et al. | Efficient 2D and 3D watershed on graphics processing unit: block-asynchronous approaches based on cellular automata | |
| CN116402987A (en) | Three-dimensional segmentation method, system, equipment and medium based on diffusion model and 3DTransformer | |
| Kato et al. | Impact of large-scale penetration of photovoltaic power generation systems on fluctuation property of electricity load | |
| CN115661170A (en) | Method, device and medium for automatically segmenting abdomen three-dimensional CT image |
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 | ||
| CB03 | Change of inventor or designer information | ||
| CB03 | Change of inventor or designer information |
Inventor after: Yan Chenggang Inventor after: Huang Hailiang Inventor after: Fu Zuxian Inventor after: Zhao Chongyu Inventor before: Huang Hailiang Inventor before: Yan Chenggang Inventor before: Fu Zuxian Inventor before: Zhao Chongyu |
|
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180406 |