CN104915950B - A kind of region based on energy constraint increases ultrasonoscopy automatic division method - Google Patents
A kind of region based on energy constraint increases ultrasonoscopy automatic division method Download PDFInfo
- Publication number
- CN104915950B CN104915950B CN201510212533.0A CN201510212533A CN104915950B CN 104915950 B CN104915950 B CN 104915950B CN 201510212533 A CN201510212533 A CN 201510212533A CN 104915950 B CN104915950 B CN 104915950B
- Authority
- CN
- China
- Prior art keywords
- super
- pixel
- energy
- image
- region
- 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.)
- Active
Links
Landscapes
- Image Analysis (AREA)
- Ultra Sonic Daignosis Equipment (AREA)
Abstract
Increase ultrasonoscopy automatic division method the invention discloses a kind of region based on energy constraint, comprise the following steps：(1) raw ultrasound image is pre-processed；(2) seed point is carried out to pretreated ultrasonoscopy to choose automatically；(3) starting point increased by the use of the seed point obtained in step (2) as region, row constraint is entered under the energy function of the present invention and is increased, reaches that constraints then stops growth and obtains final segmentation result.The segmentation result that the present invention is obtained is accurate, and cutting procedure is truly realized the segmentation of fully-automatic ultrasonic image without manually participating in, and is conducive to the processing such as follow-up focal area feature extraction.
Description
Technical field
The present invention relates to the technical field of ultrasonoscopy processing, more particularly to a kind of region based on energy constraint increases super
Acoustic image automatic division method.
Background technology
Extracted in ultrasonic computer-aided diagnosis system, it is necessary to be accomplished that first from the complicated background of ultrasonoscopy
Go out occupying lesion region, further focus could be analyzed, monitored and treated, image segmentation is exactly to realize this function
Committed step.And seldom, mostly study directly jump on the research that ultrasonoscopy occupying lesion region is split automatically at present
These steps are crossed, by edge is sketched the contours by hand, are taken time and effort, the requirement for the computer-aided diagnosis that is far from reaching.Therefore effectively
The automatic cutting techniques of ultrasonoscopy are not only for post-processing and diagnose significant, and ultrasonic calculating can be improved
Machine assistant diagnosis system automaticity.
The characteristics of there are a large amount of speckle noises and low contrast due to ultrasonoscopy, by the edge or contours segmentation of focus
It is still out a huge challenge.Movable contour model (ACM) is also snake methods by [the Kass M, Witkin such as Kass
A,Terzopoulos D,“Snakes:Active contour models”,International journal of
Computer vision, pp.321-331,1988] propose, it makes initial profile constantly receive using energy function model is minimized
Objective contour is held back, is widely used at present in the Ultrasound Image Segmentations such as breast cancer, heart, arteria carotis.But this method needs to carry
For an initial profile as close to target object edge, limited by ultrasonoscopy image-forming principle, it is difficult to automatically carry
For such a profile, generally require manually to draw, cause automaticity to be limited.
In addition, in terms of being also used in ultrasonoscopy based on the dividing method that region increases.Such as Huang et al. [Huang Q
H,Lee S Y,Liu L Z,et al.“Arobust graph-based segmentation method for breast
Tumors in ultrasound images, " Ultrasonics, 52 (2), pp.266-275,2012] it is based on to efficient
The partitioning algorithm (Effective Graph-based, EGB) of graph theory is improved, it is proposed that the sane method based on graph theory
Region is increased the fusion with region and is combined to enter ultrasonoscopy by (Robust-graph-based method, RGB), this method
Row over-segmentation, realizes the segmentation to breast cancer ultrasonoscopy.But, this method needs to set two crucial parameters, if ginseng
It is improper it is possible that the situation of over-segmentation or less divided that number is set.In addition, conventional cluster segmentation method, such as K- mean clusters
Algorithm (K-means), FCM Algorithms (FCM) etc. can also divide the image into several regions.But, these be all half from
Dynamic method is, it is necessary to which it is focal area which region, which is manually specified,.
Current China has in the field Patents：Initial wheel is obtained in Ultrasound Image Segmentation based on active contour model
Wide method (patent No. 201410141568.5).This method is sweared in the ROI region image comprising tumour using generalized gradient
Amount stream (generalized gradient vector flow, GGVF) Snake models carry out the automatic segmentation of liver neoplasm, point
The initial profile cut is obtained using linear discriminant analysis (Linear Discriminant Analysis, LDA) sorting algorithm, can
To obtain preferable segmentation effect.But this method needs just obtain the training that tumour and non-tumor region carry out having supervision in advance
Obtain initial profile.
In summary, existing Ultrasound Image Segmentation algorithm needs artificial participation mostly, and automaticity is not high.
The content of the invention
In order to overcome the disadvantages mentioned above and deficiency of prior art, energy constraint is based on it is an object of the invention to provide one kind
Region increase ultrasonoscopy automatic division method, obtained segmentation result is accurate, and cutting procedure is real real without manually participate in
Show the segmentation of fully-automatic ultrasonic image, be conducive to the processing such as follow-up focal area feature extraction.
The purpose of the present invention is achieved through the following technical solutions：
A kind of region based on energy constraint increases ultrasonoscopy automatic division method, comprises the following steps：
(1) raw ultrasound image is pre-processed, generates super-pixel；
(2) carry out seed point to pretreated ultrasonoscopy to choose automatically, be specially：
(2-1) with one 22 dimension characteristic vector yvRepresent super-pixel v, v ∈ V (V is the super-pixel set of image), yv=
{ T_a, T_d, T_h }, wherein, T_a is the mean flow rate of super-pixel, T_d is super-pixel brightness variance, T_h are super-pixel
Luminance Distribution；T_a is one-dimensional vector, and T_d is one-dimensional vector, and T_h is 20 dimensional vectors；
(2-2) tectonic setting dictionary：The super-pixel of image border surrounding is selected as background candidate's super-pixel, and by background
10% and 10% minimum super-pixel that brightness is maximum in candidate's super-pixel remove, and remaining background candidate super-pixel is used as background
Super-pixel, with the feature y of background super-pixelbBackground dictionary matrix D, i.e. D=[y are synthesized as Column vector groupsb1,yb2,…,ybm],
M is background super-pixel number；
(2-3) carries out sparse reconstruction using the background dictionary obtained in step (2-2), obtains notable figure；
(2-4) selects seed point：Binaryzation is carried out to the notable figure obtained in step (2-3) using Da-Jin algorithm (OTSU),
Using maximum continuum as area-of-interest, the center of gravity super-pixel of the area-of-interest is then seed point；
(3) starting point increased by the use of the seed point obtained in step (2) as region, enters row constraint increasing under energy function
It is long, reach that constraints then stops growth and obtains final segmentation result, be specially：
(3-1) represents the image as non-directed graph G=(V using each super-pixel in entire image as node；E), V is represented
Node set, i.e. super-pixel set, E represents line set, and side is present if two super-pixel are adjacent, otherwise in the absence of side, every
Side ei,j(ei,j∈ E) there is a non-negative weight w (vi,vj) represent the contact between super-pixel, it is shown below：
w(vi,vj) represent side right value, vi、vjAny super-pixel point respectively in image, I (vi)、I(vj) represent respectively
Super-pixel vi、vjMean flow rate, i, j ∈ [0, N), N represents the sum of super-pixel included in image；
The super-pixel point of image is divided into 3 parts, integration region R super-pixel vr, the proximity that is connected with integration region
Domain L super-pixel vlAnd background area B super-pixel vb；
The definition of (3-2) point diversity factor, internal diversity degree and contrast on border
Point diversity factor Node_diff (vl):Represent the node v in adjacent domain LlWith integration region R diversity factor, calculate
Formula is as follows
Wherein, elrFor connection super-pixel vl、vrSide, w (elr) it is side elrWeights；
Internal diversity degree Int_dif：The difference degree of all nodes in integration region R is represented, calculation formula is as follows,
MST(R；E integration region R minimum spanning tree) is represented, e represents to belong to a line of minimum spanning tree, and w (e) is
Side e weights；
Contrast on border Edge_dif：Represent being averaged for all nodes in adjacent domain L and integration region R diversity factor
Value,
Wherein, k represents the number of adjacent domain L interior joints, vlRepresent the arbitrary node in adjacent domain L；
The definition of (3-3) energy function
The energy function is defined as follows：
Energy=Eint+Eext
Wherein, EintFor internal energy, specific formula for calculation is：
Wherein, C is the node number of integration region, and pa1 is normal number；
EextFor external energy, specific formula for calculation is：
Eext=Int_dif-pa2*Edge_dif
Wherein, pa2 is normal number；
(3-4) region propagation process
Region propagation process is constantly by diversity factor Node_diff (v in adjacent domain Ll) less super-pixel addition
Into integration region, when the value of energy function reaches minimum, integration region stops growing, and comprises the following steps that：
(3-4-1) is initialized：With seed point super-pixel vsAs present fusion region R, with neighbouring all super of seed point
Pixel is adjacent domain L, to the node in L according to diversity factor Node_diff (vl) ascending sequence, and computation energy function
Energy value, sets vernier index=1, internal diversity degree Int_dif=0；
(3-4-2) is backed up with R ', L ', Energy' to R, L, Energy above, after Fusion failure, is recovered
State before fusion；
(3-4-3) is by L i-th ndex node vindexIt is added to integration region R, R=R+vindex, L is updated, and to L
In node according to Node_diff (vl) ascending sequence；Computation energy function Energy value；
(3-4-4) sets index=1, gone to (3-4-2) if Energy≤Energy'；Otherwise, Fusion failure, it is right
State after fusion is reduced, even R=R ', L=L ', Energy=Energy', sets index=index+1；If
Index is less than or equal to L interior joint number K, then goes to (3-4-3), if index is segmentation knot more than L interior joint numbers K, R
Really.
Step (2-1) super-pixel v characteristic vector yv={ T_a, T_d, T_h }, calculation formula is as follows,
Super-pixel v mean flow rate T_a：
Super-pixel v brightness variance T_d：
Wherein, n is the pixel sum included in super-pixel v, stRepresent t in super-pixel v (t ∈ (0, n]) individual picture
The brightness of vegetarian refreshments；
Luminance Distribution T_h is the vector of one 20 dimension, and specific extraction process is：By by the minimum brightness of image to maximum
Scope between brightness is evenly dividing 20 brightness sections, counts histogram conduct of each super-pixel in this 20 brightness sections
Luminance Distribution T_h.
The background dictionary obtained in step (2-3) (2-2) using step carries out sparse reconstruction, obtains notable figure, has
Body is：
(2-3-1) calculates sparse coefficient xs of each super-pixel v under background dictionary according to Its Sparse Decomposition formulav, it is as follows
Shown in formula：
Each super-pixel v is expressed as in background dictionary D sparse reconstructive residual error on image：
Wherein, lambda parameter is used for distributing error termWith sparse item ‖ xv‖1Weight, λ value is bigger, dilute
Dredge item ‖ xv‖1Constraint it is bigger, error termContribution is smaller；Conversely, sparse item ‖ xv‖1Constraint it is smaller, error
Contribution is bigger；
(2-3-2) will calculate obtained residual epsilonvAs super-pixel v significance, notable figure is obtained.
Step (1) is described to be pre-processed to raw ultrasound image, is specially：
Denoising is carried out to image first, over-segmentation then is carried out to the image after denoising, super-pixel is generated.
The denoising, be specially：
Algorithm is expanded using spot noise reduction anisotropy denoising is carried out to image.
The image to after denoising carries out over-segmentation, is specially：
Image is split using simple linear iteraction clustering algorithm.
Compared with prior art, the present invention has advantages below and beneficial effect：
(1) a kind of region based on energy constraint of the invention increases ultrasonoscopy automatic division method, dividing processing mistake
Journey full automation, without artificial participation, overcoming existing dividing method needs to provide initial profile or initial seed point not
Foot.
(2) uniformity and occupying lesion region and background of the invention by using occupying lesion region internal structure
Otherness, using an effective energy function from it is inside and outside and meanwhile constraint increase, can accurately be divided
Cut result.
(3) present invention replaces single pixel point as computing unit by the use of super-pixel, considerably reduces operand, simultaneously
Retain image edge information and local structural information well so that cutting procedure is more quick, effective.
(4) present invention realizes seed point and automatically selected, by using the method for sparse reconstruction, obtains notable figure, and select
Maximum continuous center of gravity super-pixel interested is selected as initial seed point.The existing dividing method increased based on region, often
By setting threshold value to obtain initial seed point, cause seed point location not accurate enough, method proposed by the invention can be accurate
Ground positions initial seed point, falls it and ensures that subsequent sections propagation process is correct in area-of-interest.
(5) present invention combines ultrasonoscopy imaging characteristicses in tectonic setting dictionary, to exclude the influence such as artifact, noise,
When setting up background dictionary using image border surrounding super-pixel, the 10% of wherein brightness maximum 10% and brightness minimum is removed,
So that the background dictionary correlation of construction is bigger, effect is rebuild more preferably, marking area is more protruded.
Brief description of the drawings
Fig. 1 increases the flow of ultrasonoscopy automatic division method for the region based on energy constraint of embodiments of the invention
Figure.
Fig. 2 is the region division schematic diagram of embodiments of the invention.
Fig. 3 (a) is the image to be split of the hepatic hemangioma ultrasonoscopy of embodiments of the invention.
Fig. 3 (b) is image after the denoising of the hepatic hemangioma ultrasonoscopy of embodiments of the invention.
Fig. 3 (c) is the result figure after the hepatic hemangioma ultrasonoscopy over-segmentation of embodiments of the invention.
Fig. 3 (d) is the notable figure of the hepatic hemangioma ultrasonoscopy of embodiments of the invention.
Fig. 3 (e) is the seed point selection result figure of the hepatic hemangioma ultrasonoscopy of embodiments of the invention.
Fig. 3 (f) is the segmentation result figure of the hepatic hemangioma ultrasonoscopy of embodiments of the invention.
Fig. 3 (g) is the manual segmentation figure of doctor of the hepatic hemangioma ultrasonoscopy of embodiments of the invention.
Fig. 4 (a) is the image to be split of the liver tumour ultrasonoscopy of embodiments of the invention.
Fig. 4 (b) is image after the denoising of the liver tumour ultrasonoscopy of embodiments of the invention.
Fig. 4 (c) is the result figure after the liver tumour ultrasonoscopy over-segmentation of embodiments of the invention.
Fig. 4 (d) is the notable figure of the liver tumour ultrasonoscopy of embodiments of the invention.
Fig. 4 (e) is the seed point selection result figure of the liver tumour ultrasonoscopy of embodiments of the invention.
Fig. 4 (f) is the segmentation result figure of the liver tumour ultrasonoscopy of embodiments of the invention.
Fig. 4 (g) is the manual segmentation figure of doctor of the liver tumour ultrasonoscopy of embodiments of the invention.
Embodiment
With reference to embodiment, the present invention is described in further detail, but the implementation of the present invention is not limited to this.
Embodiment
As shown in figure 1, the growth of the region based on the energy constraint ultrasonoscopy automatic division method of the present embodiment, including with
Lower step：
(1) to raw ultrasound image (such as Fig. 3 (a), Fig. 4 (a), wherein 3 (a) is hepatic hemangioma ultrasonoscopy, Fig. 4 (a)
Liver tumour ultrasonoscopy) pre-processed, generate super-pixel；The pretreatment is：Denoising is carried out to image first, so
Over-segmentation is carried out to the image after denoising afterwards, super-pixel is generated；
Larger interference can be caused to Ultrasound Image Segmentation because ultrasonoscopy contains a large amount of speckle noises, artifact etc., therefore
Need to carry out denoising to original image before dividing processing.The present embodiment selection spot noise reduction anisotropy parameter (Speckle
Reducing Anisotropic Diffusion, SRAD) algorithm [Yu Y, Acton S T.Speckle reducing
Anisotropic diffusion] carry out denoising.Fig. 3 (b), Fig. 4 (b) are Fig. 3 (a), the effect after Fig. 4 (a) denoisings
Figure.
Over-segmentation is carried out to ultrasonoscopy after the completion of denoising, super-pixel is generated, the present embodiment is using simple linear
Iterative Clustering (Simple Liner Iterative Clustering, SLIC) algorithm [Achanta R, Shaji A,
Smith K,et al.“SLIC superpixels compared to state-of-the-art superpixel
methods,”Pattern Analysis and Machine Intelligence,IEEE Transactions on,34
(11):2274-2282,2012] divide the image into that size one by one is similar and the super-pixel of image border is pressed close at edge.With
Super-pixel is basic operation unit, can not only greatly reduce operand, and can retain image edge information drawn game well
The structural information in portion, or even robustness of the algorithm to noise can be increased.Ultrasound Image Segmentation is 200 super-pixel by the present embodiment
(can be needed to choose super-pixel number according to real image), such as Fig. 3 (c), Fig. 4 (c) are respectively Fig. 3 (a), after Fig. 4 (a) segmentations
Ultrasonoscopy.
(2) carry out seed point to pretreated ultrasonoscopy to choose automatically, be specially：
(2-1) extracts super-pixel feature；With the characteristic vector y of one 22 dimensionvRepresent that (V is image to super-pixel v, v ∈ V
Super-pixel set), yv={ T_a, T_d, T_h }, wherein, the brightness side that T_a is the mean flow rate of super-pixel, T_d is super-pixel
Difference, T_h are the Luminance Distribution of super-pixel；T_a is one-dimensional vector, and T_d is one-dimensional vector, and T_h is 20 dimensional vectors；
The characteristic vector y of the super-pixel vv={ T_a, T_d, T_h }, calculation formula is as follows,
Super-pixel v mean flow rate T_a：
Super-pixel v brightness variance T_d：
Wherein, n is the pixel sum included in super-pixel v, stRepresent t in super-pixel v (t ∈ (0, n]) individual picture
The brightness of vegetarian refreshments；
Luminance Distribution T_h is the vector of one 20 dimension, and specific extraction process is：By by the minimum brightness of image to maximum
Scope between brightness is evenly dividing 20 brightness sections, counts histogram conduct of each super-pixel in this 20 brightness sections
Luminance Distribution T_h；
(2-2) tectonic setting dictionary：
In ultrasonoscopy, occupying lesion region and surrounding background area, often have on gray scale, Texture eigenvalue compared with
For obvious difference.This and inconsistent pixel aggregation zone of background, is considered area-of-interest in image procossing.And feel
Interest region is often positioned in image center location, therefore the super-pixel of selection image border surrounding is used as background candidate's super-pixel.
Situations such as influence and occupying lesion region in view of noise or artifact are likely located at image surrounding, the present embodiment selection figure
As the super-pixel of edge surrounding is as background candidate's super-pixel, and by brightness in background candidate's super-pixel it is maximum 10% and it is minimum
10% super-pixel remove, remaining background candidate super-pixel is as background super-pixel, with the feature y of background super-pixelbAs
Column vector groups synthesize background dictionary matrix D, i.e. D=[yb1,yb2,…,ybm], m is background super-pixel number；
(2-3) carries out sparse reconstruction using the background dictionary obtained in step (2-2), obtains notable figure, is specially：
(2-3-1) calculates sparse coefficient xs of each super-pixel v under background dictionary according to Its Sparse Decomposition formulav, it is as follows
Shown in formula：
Each super-pixel v is expressed as in background dictionary D sparse reconstructive residual error on image：
Wherein, lambda parameter is used for distributing error termWith sparse item ‖ xv‖1Weight, λ value is bigger, dilute
Dredge item ‖ xv‖1Constraint it is bigger, error termContribution is smaller；Conversely, sparse item ‖ xv‖1Constraint it is smaller, error
Contribution is bigger；
(2-3-2) will calculate obtained residual epsilonvAs super-pixel v significance, notable figure is obtained；Fig. 3 (d), Fig. 4 (d)
Respectively Fig. 3 (a), Fig. 4 (a) notable figure.
(2-4) selects seed point：
Because noise, artifact etc. also have difference with background area in ultrasonoscopy, therefore significance is also larger.But they
Often some discontinuous smaller areas, and occupying lesion region is one piece of larger continuum.
The present embodiment utilizes Da-Jin algorithm (OTSU) [N.Otsu, Athreshold selection method from
Gray-level histograms] binaryzation is carried out to the notable figure obtained in step (2-3), maximum continuum is made
For area-of-interest, the center of gravity super-pixel of the area-of-interest is then seed point；Fig. 3 (e), Fig. 4 (e) are respectively Fig. 3 (a), Fig. 4
(a) seed point selection result figure.
(3) starting point increased by the use of the seed point obtained in step (2) as region, enters row constraint increasing under energy function
It is long, reach that constraints then stops growth and obtains final segmentation result.Preferable integration region is that internal diversity is smaller, and side
The obvious region of edge.The partitioning algorithm of the present embodiment, the starting point increased using seed point as region, by minimizing energy letter
Number constantly merges the minimum super-pixel of diversity factor in adjacent domain.
Segmentation step is specially：
(3-1) represents the image as non-directed graph G=(V using each super-pixel in entire image as node；E), V is represented
Node set, i.e. super-pixel set, E represents line set, and side is present if two super-pixel are adjacent, otherwise in the absence of side, every
Side ei,j(ei,j∈ E) there is a non-negative weight w (vi,vj) represent the contact between super-pixel, it is shown below：
w(vi,vj) represent side right value, vi、vjAny super-pixel point respectively in image, I (vi)、I(vj) represent respectively
Super-pixel vi、vjMean flow rate, i, j ∈ [0, N), N represents the sum of super-pixel included in image；
The super-pixel of image is divided into 3 parts, integration region R super-pixel vr, the adjacent domain L that is connected with integration region
Super-pixel vlAnd background area B super-pixel vb, as shown in Fig. 2 it is v that the node of square markings, which is super-pixel,r, triangle
The node of shape mark is super-pixel vl, the node of circle markings is super-pixel vb；
The definition of (3-2) point diversity factor, internal diversity degree and contrast on border
(3-2-1) point diversity factor Node_diff (vl):Represent the node v in adjacent domain LlWith integration region R difference
Degree, calculation formula is as follows,
Wherein, elrFor connection super-pixel vl、vrSide, w (elr) it is side elrWeights；
(3-2-2) internal diversity degree Int_dif：The difference degree of all nodes in integration region R is represented, calculation formula is such as
Under,
MST(R；E integration region R minimum spanning tree) is represented, e represents to belong to a line of minimum spanning tree, and w (e) is
Side e weights.
Minimum spanning tree is used widely in partitioning algorithm of many based on figure, wherein efficient based on graph theory
Partitioning algorithm (Effective Graph-based, EGB) [K.B.Jayanthi, R.S.D.W.Banu, " Carotid artery
boundary extraction using segmentation techniques:a comparative study,”
Proceedings of ITNG, LasVegas, NV, USA, pp.1290-1295,2009] employ the most your pupil of integration region
Cheng Shu maximum side represents the internal diversity of integration region, achieves preferable segmentation effect.
(3-2-3) contrast on border Edge_dif：Represent the difference of all nodes and integration region R in adjacent domain L
The average value of degree,
Wherein, k represents the number of adjacent domain L interior joints, vlRepresent the arbitrary node in adjacent domain L；
The definition of (3-3) energy function
Energy function includes " internal energy " E for the purpose of specification curve shapeint, and specification curve and target object wheel
" external energy " E of profile degree of closenessext.In calculating process, minimizing internal energy can make curve continue internally to tighten
And keeping smooth so that internal diversity is small；And minimize external energy can then make curve persistently press close to objects' contour line and
Untill reaching unanimously.
The energy function is defined as follows：
Energy=Eint+Eext
Wherein, EintFor internal energy, specific formula for calculation is：
Wherein, C is the node number of integration region, and pa1 is normal number；The main function of internal energy is to force corresponding circle of sensation
Domain constantly outwards increases.Energy function is minimized, driving is diminished, when the super-pixel number of fusion is C smaller,
Drive item very big, force integration region constantly to become big to the i.e. C of external expansion, merge neighbouring node.But work as and be expanded to certain journey
When spending, driving item will diminish with C increase, and slowly, and at this moment energy function depends primarily on external energy Eext.As for parameter
Pa1 setting, will determine, pa1 is bigger, obtained integration region is bigger according to actual conditions.
EextFor external energy, specific formula for calculation is：
Eext=Int_dif-pa2*Edge_dif
Wherein, pa2 is normal number；The present embodiment is set to 7.The main function of external energy is that constraint increases, and is compeled
Make algorithmic statement to borderline tumor position.External energy diminishes, and seeks to Edge_dif and becomes big, Int_dif diminishes, it is ensured that fusion
Region internal diversity degree is small, edge significant degree is big.
(3-4) region propagation process
Region propagation process is constantly by diversity factor Node_diff (v in adjacent domain Ll) less super-pixel addition
Into integration region, when the value of energy function reaches minimum, integration region stops growing, and comprises the following steps that：
(3-4-1) is initialized：With seed point super-pixel vsAs present fusion region R, with neighbouring all super of seed point
Pixel is adjacent domain L, to the node in L according to diversity factor Node_diff (vl) ascending sequence, and computation energy function
Energy value, sets vernier index=1, internal diversity degree Int_dif=0；
(3-4-2) is backed up with R ', L ', Energy' to R, L, Energy above, after Fusion failure, is recovered
State before fusion；
(3-4-3) is by L i-th ndex node vindexIt is added to integration region R, R=R+vindex, L is updated, and to L
In node according to Node_diff (vl) ascending sequence；Computation energy function Energy value；
(3-4-4) sets index=1, gone to (3-4-2) if Energy≤Energy'；Otherwise, Fusion failure, it is right
State after fusion is reduced, even R=R ', L=L ', Energy=Energy', sets index=index+1；If
Index is less than or equal to L interior joint number K, then goes to (3-4-3), if index is segmentation knot more than L interior joint numbers K, R
Really.
Fig. 3 (a) of the present embodiment, Fig. 4 (a) final segmentation result respectively as shown in Fig. 3 (f), Fig. 4 (f), Fig. 3 (g),
Fig. 4 (g) is the occupying lesion edges of regions that doctor sketches the contours by hand.Compare Fig. 3 (f) and Fig. 3 (g), Fig. 4 (f) and Fig. 4 (g) is sent out
Existing, the result that the segmentation result and doctor of the present embodiment sketch the contours is coincide, and there is the present invention good ultrasonoscopy to split effect automatically
Really.
Above-described embodiment is preferably embodiment, but embodiments of the present invention are not by the embodiment of the invention
Limitation, other any Spirit Essences without departing from the present invention and the change made under principle, modification, replacement, combine, simplification,
Equivalent substitute mode is should be, is included within protection scope of the present invention.
Claims (6)
1. a kind of region based on energy constraint increases ultrasonoscopy automatic division method, it is characterised in that comprise the following steps：
(1) raw ultrasound image is pre-processed, generates super-pixel；
(2) carry out seed point to pretreated ultrasonoscopy to choose automatically, be specially：
(2-1) with one 22 dimension characteristic vector yvIt is the super-pixel set of image, y to represent super-pixel v, v ∈ V, Vv=T_a,
T_d, T_h }, wherein, T_a is the mean flow rate of super-pixel, T_d is super-pixel brightness variance, the brightness that T_h is super-pixel point
Cloth；T_a is one-dimensional vector, and T_d is one-dimensional vector, and T_h is 20 dimensional vectors；
(2-2) tectonic setting dictionary：The super-pixel of image border surrounding is selected as background candidate's super-pixel, and by background candidate
10% and 10% minimum super-pixel that brightness is maximum in super-pixel remove, and remaining background candidate super-pixel is used as the super picture of background
Element, with the feature y of background super-pixelbBackground dictionary matrix D, i.e. D=[y are synthesized as Column vector groupsb1, yb2..., ybm], m is
Background super-pixel number；
(2-3) carries out sparse reconstruction using the background dictionary obtained in step (2-2), obtains notable figure；
(2-4) selects seed point：Binaryzation is carried out to the notable figure obtained in step (2-3) using Da-Jin algorithm, by maximum company
Continuous region is as area-of-interest, and the center of gravity super-pixel of the area-of-interest is then seed point；
(3) starting point increased by the use of the seed point obtained in step (2) as region, enters row constraint growth under energy function,
Reach that constraints then stops growth and obtains final segmentation result, be specially：
(3-1) represents the image as non-directed graph G=(V using each super-pixel in entire image as node；E), V represents node
Set, i.e. super-pixel set, E represents line set, and side is present if two super-pixel are adjacent, otherwise in the absence of side, each edge eI, j
There is a non-negative weight w (vi, vj) represent the contact between super-pixel, wherein eI, j∈ E, are shown below：
w(vi, vj) represent side right value, vi、vjAny super-pixel point respectively in image, I (vi)、I(vj) super picture is represented respectively
Plain vi、vjMean flow rate, i, j ∈ [0, N), N represents the sum of super-pixel included in image；
The super-pixel point of image is divided into 3 parts, integration region R super-pixel vr, the adjacent domain L that is connected with integration region
Super-pixel vlAnd background area B super-pixel vb；
The definition of (3-2) point diversity factor, internal diversity degree and contrast on border
Point diversity factor Node_diff (vl):Represent the node v in adjacent domain LlWith integration region R diversity factor, calculation formula
It is as follows,
Wherein, elrFor connection super-pixel vl、vrSide, w (elr) it is side elrWeights；
Internal diversity degree Int_dif：The difference degree of all nodes in integration region R is represented, calculation formula is as follows,
MST(R；E integration region R minimum spanning tree) is represented, e represents to belong to a line of minimum spanning tree, and w (e) is side e's
Weights；
Contrast on border Edge_dif：All nodes and the average value of integration region R diversity factor in expression adjacent domain L,
Wherein, k represents the number of adjacent domain L interior joints, vlRepresent the arbitrary node in adjacent domain L；
The definition of (3-3) energy function
The energy function is defined as follows：
Energy=Eint+Eext
Wherein, EintFor internal energy, specific formula for calculation is：
Wherein, C is the node number of integration region, and pa1 is normal number；
EextFor external energy, specific formula for calculation is：
Eext=Int_dif-pa2*Edge_dif
Wherein, pa2 is normal number；
(3-4) region propagation process
Region propagation process is constantly by diversity factor Node_diff (v in adjacent domain Ll) less super-pixel is added to fusion
In region, when the value of energy function reaches minimum, integration region stops growing, and comprises the following steps that：
(3-4-1) is initialized：With seed point super-pixel vsAs present fusion region R, neighbouring all super-pixel are with seed point
Adjacent domain L, to the node in L according to diversity factor Node_diff (vl) ascending sequence, and computation energy function Energy
Value, set vernier index=1, internal diversity degree Int_dif=0；
(3-4-2) is backed up with R ', L ', Energy to R, L, Energy above, after Fusion failure, is recovered before fusion
State；
(3-4-3) is by L i-th ndex node vindexIt is added to integration region R, R=R+vindex, L is updated, and in L
Node is according to Node_diff (vl) ascending sequence；Computation energy function Energy value；
(3-4-4) sets index=1, gone to (3-4-2) if Energy≤Energy'；Otherwise, Fusion failure, to fusion
State afterwards is reduced, even R=R ', L=L ', Energy=Energy', sets index=index+1；If index
Less than or equal to L interior joint number K, then (3-4-3) is gone to, if index is segmentation result more than L interior joint numbers K, R.
2. the region based on energy constraint according to claim requirement 1 increases ultrasonoscopy automatic division method, it is special
Levy and be, step (2-1) super-pixel v characteristic vector yv={ T_a, T_d, T_h }, calculation formula is as follows,
Super-pixel v mean flow rate T_a：
Super-pixel v brightness variance T_d：
Wherein, n is the pixel sum included in super-pixel v, stRepresent the brightness of t-th of pixel in super-pixel v, wherein t
∈(0,n]；
Luminance Distribution T_h is the vector of one 20 dimension, and specific extraction process is：By by the minimum brightness of image to high-high brightness
Between scope be evenly dividing 20 brightness sections, count each super-pixel and be used as brightness in the histogram of this 20 brightness sections
It is distributed T_h.
3. the region according to claim 1 based on energy constraint increases ultrasonoscopy automatic division method, its feature exists
In the background dictionary obtained in step (2-3) (2-2) using step carries out sparse reconstruction, obtains notable figure, is specially：
(2-3-1) calculates sparse coefficient xs of each super-pixel v under background dictionary according to Its Sparse Decomposition formulav, such as following formula institute
Show：
Each super-pixel v is expressed as in background dictionary D sparse reconstructive residual error on image：
Wherein, lambda parameter is used for distributing error termWith sparse item | | xv||1Weight, λ value is bigger, sparse item
||xv||1Constraint it is bigger, error termContribution is smaller；Conversely, sparse item | | xv||1Constraint it is smaller, error
Contribution is bigger；
(2-3-2) will calculate obtained residual epsilonvAs super-pixel v significance, notable figure is obtained.
4. the region according to claim 1 based on energy constraint increases ultrasonoscopy automatic division method, its feature exists
In step (1) is described to be pre-processed to raw ultrasound image, is specially：
Denoising is carried out to image first, over-segmentation then is carried out to the image after denoising, super-pixel is generated.
5. the region according to claim 4 based on energy constraint increases ultrasonoscopy automatic division method, its feature exists
In the denoising is specially：
Algorithm is expanded using spot noise reduction anisotropy denoising is carried out to image.
6. the region according to claim 4 based on energy constraint increases ultrasonoscopy automatic division method, its feature exists
In the image to after denoising carries out over-segmentation, is specially：
Image is split using simple linear iteraction clustering algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510212533.0A CN104915950B (en) | 2015-04-29 | 2015-04-29 | A kind of region based on energy constraint increases ultrasonoscopy automatic division method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510212533.0A CN104915950B (en) | 2015-04-29 | 2015-04-29 | A kind of region based on energy constraint increases ultrasonoscopy automatic division method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104915950A CN104915950A (en) | 2015-09-16 |
| CN104915950B true CN104915950B (en) | 2017-10-20 |
Family
ID=54084988
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510212533.0A Active CN104915950B (en) | 2015-04-29 | 2015-04-29 | A kind of region based on energy constraint increases ultrasonoscopy automatic division method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104915950B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106952278A (en) * | 2017-04-05 | 2017-07-14 | 深圳市唯特视科技有限公司 | A kind of automatic division method in dynamic outdoor environment based on super-pixel |
| CN107292312B (en) * | 2017-06-19 | 2021-06-22 | 中国科学院苏州生物医学工程技术研究所 | Tumor CT image processing method |
| CN107633522B (en) * | 2017-08-30 | 2020-04-03 | 山东财经大学 | Brain image segmentation method and system based on local similarity active contour model |
| CN108629286B (en) * | 2018-04-03 | 2021-09-28 | 北京航空航天大学 | Remote sensing airport target detection method based on subjective perception significance model |
| CN108961196B (en) * | 2018-06-21 | 2021-08-20 | 华中科技大学 | Significance fusion method for 3D fixation point prediction based on graph |
| CN110211078B (en) * | 2019-05-14 | 2021-01-19 | 大连理工大学 | Significance detection method based on anisotropic diffusion |
| CN110910401A (en) * | 2019-10-31 | 2020-03-24 | 五邑大学 | Semi-automatic image segmentation data annotation method, electronic device and storage medium |
| CN112329791B (en) * | 2020-10-30 | 2022-12-27 | 中国电子科技集团公司第五十四研究所 | Automatic extraction method for hyperspectral image water area |
| CN113191330A (en) * | 2021-05-27 | 2021-07-30 | 宜宾学院 | Method for identifying area growth pores by fusing secondary electrons and back scattering electron images |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102663728A (en) * | 2012-03-11 | 2012-09-12 | 西安电子科技大学 | Dictionary learning-based medical image interactive joint segmentation |
| CN102831614A (en) * | 2012-09-10 | 2012-12-19 | 西安电子科技大学 | Sequential medical image quick segmentation method based on interactive dictionary migration |
| CN103295224A (en) * | 2013-03-14 | 2013-09-11 | 北京工业大学 | Breast ultrasonoscopy automatic segmentation method based on mean shift and divide |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2008110013A1 (en) * | 2007-03-15 | 2008-09-18 | Centre Hospitalier De L'universite De Montreal | Image segmentation |
-
2015
- 2015-04-29 CN CN201510212533.0A patent/CN104915950B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102663728A (en) * | 2012-03-11 | 2012-09-12 | 西安电子科技大学 | Dictionary learning-based medical image interactive joint segmentation |
| CN102831614A (en) * | 2012-09-10 | 2012-12-19 | 西安电子科技大学 | Sequential medical image quick segmentation method based on interactive dictionary migration |
| CN103295224A (en) * | 2013-03-14 | 2013-09-11 | 北京工业大学 | Breast ultrasonoscopy automatic segmentation method based on mean shift and divide |
Non-Patent Citations (2)
| Title |
|---|
| 基于区域进化的区域增长图像分割;徐爱霞等;《光学技术》;20060831;第32卷;第482-484页 * |
| 支持向量机与区域增长相结合的CT图像并行分割;刘露;《计算机科学》;20100531;第37卷(第5期);第237-239页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104915950A (en) | 2015-09-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104915950B (en) | A kind of region based on energy constraint increases ultrasonoscopy automatic division method | |
| Mharib et al. | Survey on liver CT image segmentation methods | |
| CN106780518B (en) | A kind of MR image three-dimensional interactive segmentation method of the movable contour model cut based on random walk and figure | |
| US8379957B2 (en) | System and method for segmentation of anatomical structures in MRI volumes using graph cuts | |
| CN109754404A (en) | A kind of lesion segmentation approach end to end based on more attention mechanism | |
| CN104268873B (en) | Breast tumor partition method based on nuclear magnetic resonance images | |
| CN109584244B (en) | Hippocampus segmentation method based on sequence learning | |
| CN109934235A (en) | A kind of unsupervised abdominal CT sequence image multiple organ automatic division method simultaneously | |
| Huang et al. | A Chan-Vese model based on the Markov chain for unsupervised medical image segmentation | |
| Wu et al. | AAR-RT–a system for auto-contouring organs at risk on CT images for radiation therapy planning: principles, design, and large-scale evaluation on head-and-neck and thoracic cancer cases | |
| Kitrungrotsakul et al. | Liver segmentation using superpixel-based graph cuts and restricted regions of shape constrains | |
| Ben Ayed et al. | Vertebral body segmentation in MRI via convex relaxation and distribution matching | |
| Jin et al. | Object recognition in medical images via anatomy-guided deep learning | |
| Zhang et al. | SDResU-net: separable and dilated residual U-net for MRI brain tumor segmentation | |
| Amiri et al. | Bayesian Network and Structured Random Forest Cooperative Deep Learning for Automatic Multi-label Brain Tumor Segmentation. | |
| Banerjee et al. | A CADe system for gliomas in brain MRI using convolutional neural networks | |
| Imielinska et al. | Hybrid segmentation of anatomical data | |
| CN109285176B (en) | Brain tissue segmentation method based on regularization graph segmentation | |
| Ananth et al. | CLG for Automatic Image Segmentation | |
| Liu et al. | Automatic segmentation of the human brain ventricles from MR images by knowledge-based region growing and trimming | |
| Zhang et al. | An end-to-end segmentation network for the temporomandibular joints CBCT image based on 3D U-Net | |
| Wu et al. | Multi-scale segmentation using deep graph cuts: Robust lung tumor delineation in MVCBCT | |
| Jia et al. | Amo-net: abdominal multi-organ segmentation in mri with a extend unet | |
| Wang et al. | A robust statistics driven volume-scalable active contour for segmenting anatomical structures in volumetric medical images with complex conditions | |
| Kawala et al. | Eyeing the Human Brain’s Segmentation Methods |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |