CN107292346A - A kind of MR image hippocampus partitioning algorithms learnt based on Local Subspace - Google Patents
A kind of MR image hippocampus partitioning algorithms learnt based on Local Subspace Download PDFInfo
- Publication number
- CN107292346A CN107292346A CN201710541559.9A CN201710541559A CN107292346A CN 107292346 A CN107292346 A CN 107292346A CN 201710541559 A CN201710541559 A CN 201710541559A CN 107292346 A CN107292346 A CN 107292346A
- Authority
- CN
- China
- Prior art keywords
- training sample
- hippocampus
- sample
- label
- voxel
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 210000001320 hippocampus Anatomy 0.000 title claims abstract description 52
- 238000000638 solvent extraction Methods 0.000 title claims abstract description 11
- 238000012549 training Methods 0.000 claims abstract description 83
- 238000000034 method Methods 0.000 claims abstract description 38
- 230000011218 segmentation Effects 0.000 claims abstract description 26
- 230000013016 learning Effects 0.000 claims abstract description 17
- 230000000644 propagated effect Effects 0.000 claims abstract description 8
- 239000011159 matrix material Substances 0.000 claims description 35
- 238000012360 testing method Methods 0.000 claims description 12
- 238000013507 mapping Methods 0.000 claims description 7
- 238000003064 k means clustering Methods 0.000 claims description 3
- 238000001228 spectrum Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 2
- 210000005013 brain tissue Anatomy 0.000 abstract description 2
- 230000000694 effects Effects 0.000 abstract description 2
- 210000004556 brain Anatomy 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 230000008859 change Effects 0.000 description 4
- 230000006870 function Effects 0.000 description 3
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 2
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 description 2
- 208000024827 Alzheimer disease Diseases 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000015572 biosynthetic process Effects 0.000 description 2
- 239000012141 concentrate Substances 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 201000010099 disease Diseases 0.000 description 2
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000003709 image segmentation Methods 0.000 description 2
- 238000003786 synthesis reaction Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 101000831205 Danio rerio Dynein axonemal assembly factor 11 Proteins 0.000 description 1
- 102100024282 Dynein axonemal assembly factor 11 Human genes 0.000 description 1
- 241001559542 Hippocampus hippocampus Species 0.000 description 1
- 101000831210 Homo sapiens Dynein axonemal assembly factor 11 Proteins 0.000 description 1
- 208000028017 Psychotic disease Diseases 0.000 description 1
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 206010015037 epilepsy Diseases 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 208000024714 major depressive disease Diseases 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000001272 neurogenic effect Effects 0.000 description 1
- 210000001769 parahippocampal gyrus Anatomy 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 201000000980 schizophrenia Diseases 0.000 description 1
- 241000894007 species Species 0.000 description 1
- 230000006641 stabilisation Effects 0.000 description 1
- 238000011105 stabilization Methods 0.000 description 1
- 210000003478 temporal lobe Anatomy 0.000 description 1
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
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- 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
- G06T2207/20081—Training; Learning
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of MR image hippocampus partitioning algorithms learnt based on Local Subspace, registering collection of illustrative plates will be obtained on marked atlas registration to target image first;Neighborhood is selected centered on target voxel in N number of registering collection of illustrative plates, training sample set is obtained;Extract the characteristic information such as texture and gray scale and constitute original feature space, MFA sub-space learning is carried out to original feature space;Then propagated by label, obtain the segmentation result of target image.The hippocampus dividing method that the present invention is learnt based on Local Subspace, the efficiency propagated independent of registration accuracy and label can preferably distinguish hippocampus structure and other brain tissue structures, with more stable segmentation effect.
Description
Technical field
The present invention relates to medical science brain Image Segmentation field, more particularly to a kind of MR images sea learnt based on Local Subspace
Horse body partitioning algorithm.
Background technology
Hippocampus, also known as hippocampal gyrus, are the important components of human brain nervous system, main to be responsible for memory, study
And space orientation.Research shows, change and alzheimer syndrome, schizophrenia, the temporal lobe of hippocampus volume and form
A variety of neurogenic diseases such as epilepsy, major depressive disorder have close contact.In clinical practice, by appropriate algorithm to sea
Horse body is accurately split, so as to analyze the change of its volume, form, can provide diagnosis basis for above-mentioned disease.However,
There is in irregular shape, small volume due to hippocampus, easily have partial volume effect, edge big without obvious boundary, individual difference
The features such as so that carrying out Accurate Segmentation to hippocampus turns into an extremely challenging task.Further, since between different images
The otherness that gray scale, texture, the diversity of marginal information and complexity and different images segmentation are required, so far still without one kind
Unified method is all effective to all images to be split.Traditional hippocampus dividing method species is various, relatively conventional
Have threshold method, region-growing method, clustering procedure, the dividing method based on edge, Markov random field model, based on small echo become
Dividing method changed etc..
Mechanism from hippocampus, the influence to human brain memory and spatialization function and its with alzheimer syndrome etc.
Between psychotic disorder being closely connected for researcher find since, hippocampus cutting techniques experienced manual segmentation, semi-automatic point
Cut, split three phases automatically.Due to the backwardness of technological means, the researcher of early stage generally uses threshold method, region-growing method
Hippocampus is split.And with the development of science and technology the hippocampus dividing method based on atlas registration also result in recent years
Researchers greatly pay close attention to.Dividing method based on atlas registration uses image registration as an important step, it is established that pre-
Corresponding relation between the atlas image and target image that first mark.Then using obtained Deformation Field, by the label in collection of illustrative plates
Target image space is traveled to, the final segmentation result of a target image is obtained.Dividing method based on atlas registration can be with
It is divided into based on single collection of illustrative plates, based on average shape collection of illustrative plates and based on multiple three kinds of collection of illustrative plates.
Dividing method based on single collection of illustrative plates is carried out registering with single collection of illustrative plates and target image, and therefore, it is difficult to adapt to difference
Otherness between individual, so as to it is difficult to obtain preferable segmentation result.
Researchers propose the segmentation side based on average shape collection of illustrative plates on the basis of the dividing method based on single collection of illustrative plates
Method, this method uses multiple collection of illustrative plates and certain means, obtains the information of all collection of illustrative plates, and there are multiple frequency spectrums to put down for generation one
The synthesis collection of illustrative plates of equal shape, then the synthesis collection of illustrative plates is registering with target image to realize final segmentation.
Similar with based on average shape collection of illustrative plates dividing method, the dividing method based on multiple atlas registrations is equally using multiple
Collection of illustrative plates.This method is registering with target image progress respectively by multiple collection of illustrative plates, obtains multigroup deformation parameter, respectively to each collection of illustrative plates correspondence
Mark image carry out deformation, corresponding collection of illustrative plates label is traveled in target image space, then merge the collection of illustrative plates of all propagation
Label and the final segmentation result for producing a target image.
Compared with traditional hippocampus dividing method, the segmentation performance of the hippocampus dividing method based on atlas registration has
Larger lifting.By taking the optimal dividing method based on many spectrograms registration of wherein segmentation performance as an example, by merging target image
The propagation label of multichannel chromatogram in space, this method can obtain the higher segmentation result of robustness, the degree of accuracy.However, being based on
The efficiency that the performance of the dividing method of multichannel chromatogram registration is propagated dependent on the precision and label of registration, therefore stabilization can not be obtained
Segmentation result.
Therefore, the efficiency that the segmentation performance for how solving existing hippocampus body method depends on registration accuracy and label to propagate
The problem of problem is current medical science brain Image Segmentation field urgent need to resolve.
The content of the invention
For the deficiency of prior art, the present invention proposes a kind of MR images hippocampus segmentation learnt based on Local Subspace
Algorithm, comprises the following steps:
Step 1:N number of registering collection of illustrative plates will be obtained on marked atlas registration to target hippocampus MR images, it is described to have matched somebody with somebody
Positive sample in quasi- collection of illustrative plates is the hippocampus labeled as+1, and the negative sample in the registering collection of illustrative plates is the non-hippocampus labeled as -1
Body;
Step 2:Size is selected to be w × w × w neighborhood body centered on target voxel in N number of registering collection of illustrative plates respectively
Element, and N × w × w × w candidate's training sample is obtained, q is chosen respectively from the positive and negative samples of candidate's training sample1+q2
Individual sample builds positive and negative training sample set, wherein, Positive training sample collection and Negative training sample collection composing training sample set are described
Training sample set includes 2 (q1+q2) individual training sample;
Step 3:Concentrated from training sample and extract the original spy of characteristic information composition at least including texture and image intensity value
Levy space;
Step 4:The sub-space learning for the original feature space analyze based on marginal Fisher;
Step 4.1:The type label of training sample is expanded into the neighborhood for t × t centered on the training sample × t sizes
The structure type tag block of interior all training samples, the corresponding structure type mark of all training samples that the training sample is concentrated
Sign block and constitute structure type tally set;
Step 4.2:K means clustering methods are applied on the structure type tally set, obtained and the structure type
The corresponding subtype tally set of tally set, the structure type tally set is replaced and by the subtype with the subtype tally set
Tally set combines the training sample set for obtaining new tape label with corresponding training sample set;
Step 4.3:Build intrinsic figure GIScheme G with punishmentP;
Step 4.4:Define the intrinsic figure GISimilar matrix SIAnd the punishment figure GPSimilar matrix SP;
Step 4.5:Define the intrinsic figure GIDiagonal matrix DIWith Laplacian Matrix LIAnd the punishment figure GP's
Diagonal matrix DPWith Laplacian Matrix LP;
Step 4.6:Subspace mapping matrix is calculated, the voxel maps that training sample concentration is characterized with characteristic information are to
In Learning Subspaces;
Step 5:Label is propagated;
Step 6:Input target hippocampus MR images;
Step 7:Extract the characteristic information of each test sample voxel in the target hippocampus MR images;
Step 8:The test sample voxel characterized with characteristic information in the target hippocampus MR images is projected into institute
State in Learning Subspaces;
Step 9:Used in the Learning Subspaces based on k nearest neighbor classification device to the target hippocampus MR images
The classification of each test sample voxel is differentiated and obtains final segmentation result.
The beneficial effects of the present invention are:
The segmentation performance of existing dividing method excessively relies on the precision of registration and rich, the present invention of extracted feature
The dividing method of proposition, by carrying out the sub-space learning based on MFA to feature space, can effectively improve voxel feature
Resolution, makes segmentation performance obtain larger lifting, and segmentation performance is more stablized, and the scope of application is wider.
Brief description of the drawings
Fig. 1 is the method flow diagram of partitioning algorithm of the present invention;
Fig. 2 is the schematic diagram that a multiscale image block is built by Gaussian filter;With
Fig. 3 is the intrinsic figure of the present invention and the syntople schematic diagram of punishment figure.
Embodiment
It is described in detail below in conjunction with the accompanying drawings.
Voxel in the present invention refers to:Voxel is volume element, is the least unit that data are split with three dimensions.
Target voxel in the present invention refers to:The voxel of classification to be determined in feeling the pulse with the finger-tip mark hippocampus MR images.
Label in the present invention refers to:Label is the corresponding mark of sample voxel in the present invention, and wherein positive sample is mark
+ 1 hippocampus is designated as, negative sample is the non-hippocampus labeled as -1.
For prior art to hippocampus split exist deficiency, the present invention propose it is a kind of based on Local Subspace learn
MR image hippocampus partitioning algorithms.Fig. 1 is the method flow diagram of invention partitioning algorithm.In conjunction with accompanying drawing, the present invention is described in detail
A kind of MR image hippocampus partitioning algorithms learnt based on Local Subspace, hippocampus partitioning algorithm proposed by the present invention includes
Following steps:
Step 1:Marked collection of illustrative plates is inputted, spatial alternation, such as affine transformation, piecewise affine are carried out to these marked collection of illustrative plates
Conversion or nonlinear transformation, map that on target hippocampus MR images, obtain N number of Registration of MR collection of illustrative plates.
Step 2:Size is selected to be w × w × w neighboring voxels centered on target voxel N number of registering atlas image,
Obtain N × w × w × w candidate's training sample.Q is selected respectively from the positive and negative samples of candidate's training sample1+q2Individual sample structure
Training sample set is built, training sample set includes Positive training sample collection and Negative training sample collection.
A kind of specific embodiment, for the voxel x set the goal on hippocampus MR images, all registration figures
All voxels in spectrum in neighboring voxels V (x) corresponding with voxel x all be used to generate training sample set, neighboring voxels V (x)
Size is w × w × w.Therefore, it can obtain N × w × w × w training sample from N number of registering collection of illustrative platesCandidate's training sample set is constructed, whereinTo be derived from i-th of collection of illustrative plates
J-th of sample voxel, lI, j∈ {+1, -1 } is the type label of each candidate's training sample.
Candidate's training sample, which is concentrated, includes similarity degrees different from the voxel sample x of type to be discriminated in target image
Sample.In order that training sample concentrates positive and negative samples number to reach balance during sub-space learning, from the positive and negative of candidate's training sample
Q is chosen in sample respectively1+q2Individual sample builds positive and negative training sample set.Specifically, for Positive training sample, selecting q1Individual and mesh
Standard type element sample x similarity highest samples, and randomly select q in remaining Positive training sample2Individual positive sample constitutes positive instruction
Practice sample set.Similarly, for Negative training sample, also selection selects q1Individual and target voxel sample x similarity highest samples,
And randomly select q in remaining Negative training sample2Individual negative sample constitutes Negative training sample collection.Positive training sample collection and negative training
Sample set, which is combined, obtains training sample set.
Similitude is calculated by following structural similarity calculation formula (1):
The average and standard deviation for the neighborhood for being w × w × w in voxel x and j size centered on wherein μ and σ difference.
Step 3:The characteristic information at least including texture and image intensity value is extracted for constructing original feature space.The original
Beginning feature space it is follow-up by the training for being used for subspace to improve the area of the voxel characterized with characteristic information in the target image
Indexing.
Step 31:The texture feature information that this algorithm is extracted includes:
First-order difference wave filter (FODs) is exported, such as formula (2)
{ H (z+u)-H (z-u), u=(r cos θ sin φ, r sin θ sin φ, r cos φ) } (2)
Second differnce wave filter (SODs) is exported, such as formula (3)
{ H (z+u)+H (z-u) -2H (z), u=(r cos θ sin φ, r sin θ sin φ, r cos φ) } (3)
Three-dimensional hyperplane wave filter output, such as formula (4)
3D-Sobel wave filters output such as formula (5)
Laplace filter is exported, such as formula (6)
Tonal range wave filter is exported, such as formula (7)
Wherein, CA, b, c(z) central point is represented in the rectangular area that z volumes are a × b × c, and the u amounts of being biased towards, r is u
Length, θ and φ are u two anglecs of rotation, and Ψ is filter operator, and H is corresponding filter function, Op(z) z p neighborhoods are represented
In voxel, * represents convolution algorithm.In texture information, FODs and SODs are used for the change of intensity on one cut-off rule of detection
Situation.
Here, r ∈ { 1,2,3 }, θ ∈ { 0, π/4, pi/2,3 π/4 }, φ ∈ { 0, π/4, pi/2 } are made.Three-dimensional hyperplane filtering
Device and 3D-Sobel wave filters are the extensions to FODs and SODs in the plane.In addition, also using other two kinds of wave filters, i.e.,
Laplace filter and tonal range wave filter.Laplace filter is isotropic, for detecting that second-order intensity becomes
Change.Tonal range wave filter is then used to calculate the minimum gray scale difference between maximum of each voxel in given neighborhood.The present invention's
In technical scheme, p neighborhoods are dimensioned to p ∈ { 7,19,27 }.
Step 32:The step of extracting image intensity value characteristic information is as follows:
The image block not overlapped each other is divided an image into first.In the present invention, using different Gaussian filters come by
Green strength information replaces with the strength information after convolution.
Fig. 2 is the schematic diagram that a multiscale image block is built by Gaussian filter.As shown in Fig. 2 due to segmentation
Final purpose is to determine the type label of target image voxel, is caught in an appropriate scope (subgraph on the right in such as Fig. 2)
Obtain the central information (internal layer area in such as Fig. 2) of voxel block, and gradually expand this region, with image block centre distance
Have increased access to rough multiple dimensioned strength information.
Step 4:To preserve the anatomical information of MR images in segmentation, the type label of training sample is expanded as knot
Structure type label block, be subsequently based on k mean clusters obtain subtype tally set corresponding with the structure type tally set and with
This constructs new training sample set, then limit Fisher analyses (Margin Fisher are carried out to original feature space
Analysis, MFA) sub-space learning:
Step 4.1:The type label of training sample is expanded as the big small neighbourhoods of t × t × t centered on the training sample
The structure type tag block of interior all training samples, the corresponding structure type tag block of all training samples that training sample is concentrated
Constitute structure type tally set.
Specifically, with U=[u1, u2..., uM] training sample set is represented,Represent training sample, Y=[y1,
y2..., yM] represent voxel sample type tally set, yi∈ {+1, -1 } is represented and uiThe type label of corresponding voxel sample,
Positive sample namely the sample type label for belonging to hippocampus are+1, and the type label of negative sample namely non-hippocampus sample is -1.M
=2 (q1+q2) it is the quantity that training sample concentrates training sample, d is sample characteristics dimension.By the type mark of each training sample
Label Expand as the knot of all training samples in the big small neighbourhoods of t × t centered on training sample × t
Structure type label blockWherein t × t × t represents the volume of tag block.
Step 4.2:Structure type tally set corresponding with its for training sample set UEach training sample
This uiA structure type tag block is assignedBecause MFA cannot be directly used to train subspace, thus it is rightUsing k means clustering methods, subtype corresponding with structure type tally set is obtained under the conditions of unsupervised
Tally set { y '1..., y 'M, y 'i∈ {+1, -1 } represents subtype label, by subtype tally set replacing structure type label
Collection, and the new training sample set of generation is combined with training sample set, for follow-up MFA sub-space learnings.Specifically, described in
Subtype tally set replaces the structure type tally set and combines the subtype tally set with corresponding training sample set
To the training sample set of new tape label.
Step 4.3:Build intrinsic figure GIScheme G with punishmentP;
Two non-directed graphs, i.e., intrinsic figure G are built according to figure embedding theoryI={ U, SIAnd punishment figure GP={ U, SP, use
Manifold structure and manifold border in the different types of class of description, wherein, U represents candidate's training set,For
GIAnd GPCorresponding similarity matrix;
Step 4.4:Define intrinsic figure similarity matrix SIAnd punishment figure similarity matrix SP;
In order to describe the compactness of sample in class, intrinsic figure GICorresponding similar matrix SIIt can be defined by following formula (8):
WhereinRepresent u in same classjK1The indexed set of arest neighbors.
In addition, MFA defines punishment figure GPSimilarity matrix SPDiscrete case for portraying sample between class, it is defined
As shown in formula (9):
WhereinIt is set { (i, j), yi=y, yj≠ y } in k2A nearest data set.
Fig. 3 is the syntople figure of intrinsic figure and punishment.For MFA, the syntople of intrinsic figure and punishment figure can be by Fig. 3
Draw.
Step 4.5:Pass through intrinsic figure similarity matrix SIDefine the diagonal matrix D of intrinsic figureIWith Laplacian Matrix LI,
And by punishing figure similarity matrix SPThe diagonal matrix D of definition punishment figurePWith Laplacian Matrix LP。
Specifically, the corresponding diagonal matrix D of intrinsic figureIWith Laplacian Matrix LIIt can be drawn by following formula (10) definition:
Punishment figure GPCorresponding Laplacian Matrix LPIt can be defined with the method similar with formula (10).
Step 4.6:The L obtained by step 4.5IAnd LPSubspace mapping matrix is calculated, will by subspace mapping matrix
The voxel maps that training sample concentration is characterized with characteristic information are into Learning Subspaces.
Specifically, being embedded in framework, subspace mapping matrix according to figure(wherein p ties up for subspace
Degree) it can be calculated by solving object function (11):
Wherein Tr () is the mark of matrix, mapping matrix ΦMFAIncluding matrix The corresponding characteristic vector of preceding several larger characteristic values.
Obtain after subspace mapping matrix, the training sample u that training sample concentration is characterized with characteristic informationi∈ U, i=1,
2 ..., M is mapped in Learning Subspaces in the form of formula (12), and formula (12) is as follows:
In Learning Subspaces, distance is compressed between the sample with similar tags, and the sample with different labels
Between distance be extended, effectively improve the resolution of the image voxel characterized with characteristic information.
Step 5:Label is propagated.Label propagate basic ideas be:Prediction is gone not mark with the label information of marked sample
Remember the label information of sample, the label of each sample is broadcast to adjacent node by similarity.Label propagate have it is simple, efficiently,
Efficiently advantage.
Step 6:Input target hippocampus MR images.
Step 7:The characteristic information of each test sample voxel in target hippocampus MR images is extracted, to target hippocampus
Each test sample voxel in MR images is according to the method characteristic information extraction consistent with training sample, i.e., empty in primitive character
Between in characterized.
Step 8:The test sample voxel characterized with characteristic information in target hippocampus MR images is projected into step 4.6
In in the subspace that has learnt;
Step 9:Used in the Learning Subspaces that step 4.6 is obtained based on k nearest neighbor classification device to target hippocampus MR
The classification of the test sample voxel of image is differentiated, obtains final segmentation result.The basic ideas of k nearest neighbor sorting technique are:
If most of in K of the sample in feature space most like samples belong to some classification, the sample
Belong to this classification.It is that hippocampus is also non-hippocampus to judge test sample voxel.
By the present invention and two kinds of traditional partitioning algorithms based on multichannel chromatogram, i.e., non local piecemeal label propagation algorithm with it is dilute
Piecemeal label propagation algorithm is dredged compared to knowable to, algorithm proposed by the present invention is by using MFA sub-space learnings to marked collection of illustrative plates
Label propagated, and in the subspace of acquistion carry out k nearest neighbor classification method, can be by the hippocampus in MR brain images
Body structure is preferably distinguished with other brain tissue structures.It is substantial amounts of test result indicates that, the final segmentation arrived of the present invention
As a result in Dice coefficients, Jaccard coefficients, above two point is superior in the evaluation index such as PI, RI and Hausdroff distance
Cut algorithm.The efficiency that technical scheme is propagated independent of the precision and label of registration so that segmentation performance is more stable,
The algorithm scope of application is wider.
It should be noted that above-mentioned specific embodiment is exemplary, those skilled in the art can disclose in the present invention
Various solutions are found out under the inspiration of content, and these solutions also belong to disclosure of the invention scope and fall into this hair
Within bright protection domain.It will be understood by those skilled in the art that description of the invention and its accompanying drawing be illustrative and not
Constitute limitations on claims.Protection scope of the present invention is limited by claim and its equivalent.
Claims (1)
1. a kind of MR image hippocampus partitioning algorithms learnt based on Local Subspace, it is characterised in that
Step 1:N number of registering collection of illustrative plates, the registration figure will be obtained on marked atlas registration to target hippocampus MR images
Positive sample in spectrum is the hippocampus labeled as+1, and the negative sample in the registering collection of illustrative plates is the non-hippocampus labeled as -1;
Step 2:Size is selected to be w × w × w neighboring voxels centered on target voxel in N number of registering collection of illustrative plates respectively, and
N × w × w × w candidate's training sample is obtained, q is chosen respectively from the positive and negative samples of candidate's training sample1+q2Individual sample
The positive and negative training sample set of this structure, wherein, Positive training sample collection and Negative training sample collection composing training sample set, the training
Sample set includes 2 (q1+q2) individual training sample;
Step 3:Concentrated from training sample and extract the characteristic information composition primitive character sky at least including texture and image intensity value
Between;
Step 4:The sub-space learning for the original feature space analyze based on marginal Fisher;
Step 4.1:The type label of training sample is expanded as institute in the neighborhood of t × t centered on the training sample × t sizes
There are the structure type tag block of training sample, the corresponding structure type tag block of all training samples that the training sample is concentrated
Constitute structure type tally set;
Step 4.2:K means clustering methods are applied on the structure type tally set, obtained and the structure type label
Collect corresponding subtype tally set, the structure type tally set is replaced and by the subtype label with the subtype tally set
Collect the training sample set for being combined with corresponding training sample set and obtaining new tape label;
Step 4.3:Build intrinsic figure GIScheme G with punishmentP;
Step 4.4:Define the intrinsic figure GISimilar matrix SIAnd the punishment figure GPSimilar matrix SP;
Step 4.5:Define the intrinsic figure GIDiagonal matrix DIWith Laplacian Matrix LIAnd the punishment figure GPIt is diagonal
Matrix DPWith Laplacian Matrix LP;
Step 4.6:Subspace mapping matrix is calculated, the voxel maps that training sample concentration is characterized with characteristic information are to having learnt
In subspace;
Step 5:Label is propagated;
Step 6:Input target hippocampus MR images;
Step 7:Extract the characteristic information of each test sample voxel in the target hippocampus MR images;
Step 8:By the test sample voxel characterized with characteristic information in the target hippocampus MR images project to it is described
In Learning Subspaces;
Step 9:In the Learning Subspaces use based on k nearest neighbor classification device to the target hippocampus MR images each
The classification of test sample voxel is differentiated and obtains final segmentation result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710541559.9A CN107292346B (en) | 2017-07-05 | 2017-07-05 | A kind of MR image hippocampus partitioning algorithm based on Local Subspace study |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710541559.9A CN107292346B (en) | 2017-07-05 | 2017-07-05 | A kind of MR image hippocampus partitioning algorithm based on Local Subspace study |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107292346A true CN107292346A (en) | 2017-10-24 |
CN107292346B CN107292346B (en) | 2019-11-15 |
Family
ID=60100778
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710541559.9A Expired - Fee Related CN107292346B (en) | 2017-07-05 | 2017-07-05 | A kind of MR image hippocampus partitioning algorithm based on Local Subspace study |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107292346B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107895373A (en) * | 2017-11-14 | 2018-04-10 | 西安建筑科技大学 | Image partition method based on regional area uniformity manifold constraint MRF models |
CN109584244A (en) * | 2018-11-30 | 2019-04-05 | 安徽海浪智能技术有限公司 | A kind of hippocampus dividing method based on Sequence Learning |
CN112183639A (en) * | 2020-09-30 | 2021-01-05 | 四川大学 | Mineral image identification and classification method |
CN112674720A (en) * | 2020-12-24 | 2021-04-20 | 四川大学 | Alzheimer disease pre-diagnosis method based on 3D convolutional neural network |
CN117557821A (en) * | 2024-01-11 | 2024-02-13 | 兰州大学 | Semi-supervised subspace clustering method and device based on soft MFA |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103714536A (en) * | 2013-12-17 | 2014-04-09 | 深圳先进技术研究院 | Sparse-representation-based multi-mode magnetic resonance image segmentation method and device |
CN105205810A (en) * | 2015-08-30 | 2015-12-30 | 南方医科大学 | Distance-field-fusion-based hippocampus segmentation method of MR image |
WO2016020008A1 (en) * | 2014-08-07 | 2016-02-11 | Mcmri Ag | System and method for image processing and analysis |
US20160090598A1 (en) * | 2014-09-30 | 2016-03-31 | The Regents Of The University Of California | Antisense compounds and uses thereof |
CN106204562A (en) * | 2016-07-04 | 2016-12-07 | 西安交通大学 | A kind of method of the arched roof white matter segmentation merged based on fMRI Yu DTI |
-
2017
- 2017-07-05 CN CN201710541559.9A patent/CN107292346B/en not_active Expired - Fee Related
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103714536A (en) * | 2013-12-17 | 2014-04-09 | 深圳先进技术研究院 | Sparse-representation-based multi-mode magnetic resonance image segmentation method and device |
CN103714536B (en) * | 2013-12-17 | 2017-06-16 | 深圳先进技术研究院 | The dividing method and device of the multi-modal MRI based on rarefaction representation |
WO2016020008A1 (en) * | 2014-08-07 | 2016-02-11 | Mcmri Ag | System and method for image processing and analysis |
US20160090598A1 (en) * | 2014-09-30 | 2016-03-31 | The Regents Of The University Of California | Antisense compounds and uses thereof |
CN105205810A (en) * | 2015-08-30 | 2015-12-30 | 南方医科大学 | Distance-field-fusion-based hippocampus segmentation method of MR image |
CN106204562A (en) * | 2016-07-04 | 2016-12-07 | 西安交通大学 | A kind of method of the arched roof white matter segmentation merged based on fMRI Yu DTI |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107895373A (en) * | 2017-11-14 | 2018-04-10 | 西安建筑科技大学 | Image partition method based on regional area uniformity manifold constraint MRF models |
CN107895373B (en) * | 2017-11-14 | 2021-04-27 | 西安建筑科技大学 | Image segmentation method based on local region consistency manifold constraint MRF model |
CN109584244A (en) * | 2018-11-30 | 2019-04-05 | 安徽海浪智能技术有限公司 | A kind of hippocampus dividing method based on Sequence Learning |
CN112183639A (en) * | 2020-09-30 | 2021-01-05 | 四川大学 | Mineral image identification and classification method |
CN112674720A (en) * | 2020-12-24 | 2021-04-20 | 四川大学 | Alzheimer disease pre-diagnosis method based on 3D convolutional neural network |
CN112674720B (en) * | 2020-12-24 | 2022-03-22 | 四川大学 | Alzheimer disease pre-judgment method based on 3D convolutional neural network |
CN117557821A (en) * | 2024-01-11 | 2024-02-13 | 兰州大学 | Semi-supervised subspace clustering method and device based on soft MFA |
Also Published As
Publication number | Publication date |
---|---|
CN107292346B (en) | 2019-11-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107292346B (en) | A kind of MR image hippocampus partitioning algorithm based on Local Subspace study | |
Xie et al. | Beyond classification: structured regression for robust cell detection using convolutional neural network | |
CN107506761B (en) | Brain image segmentation method and system based on significance learning convolutional neural network | |
CN112017198B (en) | Right ventricle segmentation method and device based on self-attention mechanism multi-scale features | |
CN107016405B (en) | A kind of pest image classification method based on classification prediction convolutional neural networks | |
CN106682633B (en) | The classifying identification method of stool examination image visible component based on machine vision | |
CN108830326B (en) | Automatic segmentation method and device for MRI (magnetic resonance imaging) image | |
CN105869173B (en) | A kind of stereoscopic vision conspicuousness detection method | |
CN104933711B (en) | A kind of automatic fast partition method of cancer pathology image | |
Zhang et al. | Hybrid region merging method for segmentation of high-resolution remote sensing images | |
CN104766046B (en) | One kind is detected using traffic mark color and shape facility and recognition methods | |
Pan et al. | Cell detection in pathology and microscopy images with multi-scale fully convolutional neural networks | |
CN110889852A (en) | Liver segmentation method based on residual error-attention deep neural network | |
CN106097335A (en) | Digestive tract focus image identification system and recognition methods | |
Morris | A pyramid CNN for dense-leaves segmentation | |
CN106296638A (en) | Significance information acquisition device and significance information acquisition method | |
CN109409384A (en) | Image-recognizing method, device, medium and equipment based on fine granularity image | |
CN107133651A (en) | The functional magnetic resonance imaging data classification method of subgraph is differentiated based on super-network | |
Pan et al. | Mitosis detection techniques in H&E stained breast cancer pathological images: A comprehensive review | |
He et al. | Automatic magnetic resonance image prostate segmentation based on adaptive feature learning probability boosting tree initialization and CNN-ASM refinement | |
CN107944403A (en) | Pedestrian's attribute detection method and device in a kind of image | |
Ghesu et al. | Marginal space deep learning: Efficient architecture for detection in volumetric image data | |
Mazzini et al. | Spatial sampling network for fast scene understanding | |
CN111639697B (en) | Hyperspectral image classification method based on non-repeated sampling and prototype network | |
CN105512670A (en) | HRCT peripheral nerve cutting based on KECA feature dimension reduction and clustering |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20191115 |