CN109753997A - A kind of liver neoplasm automatic and accurate Robust Segmentation method in CT image - Google Patents
A kind of liver neoplasm automatic and accurate Robust Segmentation method in CT image Download PDFInfo
- Publication number
- CN109753997A CN109753997A CN201811553883.3A CN201811553883A CN109753997A CN 109753997 A CN109753997 A CN 109753997A CN 201811553883 A CN201811553883 A CN 201811553883A CN 109753997 A CN109753997 A CN 109753997A
- Authority
- CN
- China
- Prior art keywords
- pixel
- image
- liver
- super
- gray
- 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
Abstract
The invention discloses a kind of liver neoplasm automatic and accurate Robust Segmentation methods in CT image, comprising: (1) pre-processes to CT image, extract liver area therein；(2) multi-layer iterative segmentation is carried out to liver area with the image superpixel dividing method based on LI-SLIC, is same super-pixel by the more consistent region division of gray scale in liver and texture, obtains the boundary between normal liver parenchyma and liver neoplasm；(3) according to image local gray scale and textural characteristics, normal liver parenchyma/liver neoplasm two is carried out to each pixel of liver area and is classified；(4) classified according to the super-pixel that liver area pixel classification results generate step (2), obtain final liver neoplasm segmentation result.The present invention can effectively solve to improve the efficiency and precision of liver diseases computer-aided diagnosis because the segmentation of the brings such as liver neoplasm obscurity boundary, structure is complicated, gray scale multiplicity is difficult in CT imaging noise and CT image.
Description
Technical field
The present invention relates to image procossing, mode identification technology, in particular to the liver neoplasm in a kind of CT image is certainly
Move accurate Robust Segmentation method.
Background technique
Computed tomography (Computed Tomography, CT) has, image resolution ratio small to human body wound
Height, energy is intuitive, accurately reflects the features such as patient's liver and lesion region, is widely used in the clinical diagnosis of liver diseases.CT
The segmentation of liver neoplasm is to carry out the important prerequisite of liver neoplasm load Analysis in image, can be rapidly and accurately using segmentation result
Ground obtains the information such as shape, position, size, distribution, active degree and the transfer case of liver tumour, before liver diseases
Phase diagnosis, operative treatment and radiotherapy play a crucial role.
Since liver organ anatomical structure is complicated, Different Individual differs greatly, and when imaging by noise, deviate and make
The influence of shadow agent etc., obtained CT image for liver usually have complexity and diversity, and the liver neoplasm tissue in image is logical
Often with there is the features such as obscurity boundary, form of diverse, uneven gray scale.The tumor region manually divided in CT image not only consumes
When effort, and subjectivity is strong, and segmentation result depends critically upon the experience and skill of doctor.Currently, realizing liver using computer
The automatic segmentation of tumour has become the hot spot of the continuous research and discovery of lot of domestic and foreign scholar.Divide automatically to reduce liver neoplasm
The complexity and difficulty cut improve segmentation precision, and major part automatic division method is before carrying out liver neoplasm segmentation at present
First CT image can be pre-processed, obtain liver area-of-interest therein.Document " A hierarchical local
region-based sparse shape composition for liver segmentation in CT scans”
(pattern recognition, pp. 88-106,2016.) and " Medical image segmentation by
combining graph cuts and oriented active appearance models”(IEEE transaction
On image processing, pp.2035-2046,2012.) disclosed in method can effectively divide abdominal CT figure automatically
Liver area as in.On the basis of liver segmentation, existing liver neoplasm automatic division method be broadly divided into it is unsupervised and
There are supervision two major classes.Unsupervised dividing method refers to the letter that can directly obtain from image with gray scale, gradient or texture etc.
The method being split is ceased, mainly includes that threshold value, cluster, region growing, movable contour model and figure are cut.These methods are only sharp
Be split with image bottom data, be not bound with high-rise priori knowledge, be generally difficult to adapt to CT image for liver complexity and
Diversity.There is the method for supervision to be primarily referred to as the method being split using characteristics of image priori combination machine learning, such side
Although method can effectively distinguish tumor tissues and normal liver parenchyma by increasing training sample, liver neoplasm form of diverse, ash are solved
The problems such as degree is uneven, but accurate liver neoplasm boundary can not be obtained.
Summary of the invention
The present invention has fully considered the shortcomings that above-mentioned prior art and insufficient, it is intended that providing in a kind of CT image
Liver neoplasm automatic and accurate Robust Segmentation method, provide technical support and certainly for liver diseases computer-aided diagnosis and treatment
Plan service.
The present invention is realized by the following scheme:
A kind of liver neoplasm automatic and accurate Robust Segmentation method in CT image, comprising the following steps:
(1) it cuts algorithm using sparse combination of shapes or figure and CT image is pre-processed, obtain the liver area in image；
(2) it in order to effectively obtain the weak boundary between liver neoplasm and normal liver parenchyma, proposes with based on LI-SLIC
(Simple Liner Iterative Clustering Based on Local Information, based on local message
Simple linear iteration cluster) image superpixel dividing method multi-layer iterative segmentation is carried out to liver area, will be grey in liver
Degree and the more consistent region division of texture are same super-pixel, and super-pixel segmentation result is denoted as Si(i=1,2 ..., n),
Middle n is super-pixel number；
(3) utilize CT image local gray feature and textural characteristics, one normal liver parenchyma of training liver neoplasm two classify
Classifier, and classify to each pixel for the liver area that step (1) obtains with trained classifier, point
Class result is denoted as F, if pixel p is divided into normal liver parenchyma, F (p) value is 1, if being divided into tumor tissues, F (p) is taken
Value is -1；
(4) according to liver area pixel classification results, the super-pixel generated to step (2) is classified, and is obtained final
Liver neoplasm Accurate Segmentation as a result, method are as follows: for each super-pixel Si, calculate its all pixels classification knot for being included
The weighted sum of fruit:
Wherein, weight wpIt is defined respectively as with normalization factor M:
Wherein dpFor pixel p and super-pixel SiMass center between Euclidean distance, dmaxFor super-pixel SiMiddle all pixels
Maximum Euclidean distance of the point apart from its mass center, pixel p is closer apart from super-pixel mass center, weight wpValue is bigger, the pixel
It is also bigger to the contribution of super-pixel classification.If the weighted sum λ being calculated by above-mentioned formulaiLess than 0, then by its corresponding super picture
Plain SiLabeled as tumor region, it is otherwise marked as normal liver parenchyma.
A kind of liver neoplasm automatic and accurate Robust Segmentation method characteristic in CT image also resides in, in described (2) step,
Image superpixel dividing method based on LI-SLIC, specifically includes:
(I) original image is divided into continuous nonoverlapping image subblock that side length is h, h value is as follows:
Wherein, N is total number of image pixels mesh, n1For super-pixel number to be generated, it is set greater than 0 natural number；
(II) initial cluster center C is determined according to the average gray of all pixels in image subblock and average spatial informationk:
Ck=[lk,xk,yk]
Wherein, k=1,2, Λ, n1, lkFor the gray average of pixel in k-th of image subblock, xkAnd ykIt is then k-th of figure
As the mean value of pixel column coordinate and column coordinate in sub-block；
(III) it calculates with coordinate (xk,yk) centered on, each in subgraph corresponding to the rectangular box that 2h+1 is side length
Pixel p and cluster centre CkDistance metric D (p, Ck):
Wherein, dc(p,Ck) and ds(p,Ck) respectively indicate pixel p and cluster centre CkGray homogeneity and space length,
Parameter m is normal number, and preferably 10~30 normal number adjusts the distance for controlling Gray homogeneity and space length and measures D (p, Ck)
Influence, m value is bigger, and space length influences it bigger, and vice versa；
(IV) distance metric D (p, C are usedk) cluster is iterated to image pixel, and constantly update cluster centre Ck, directly
To the Euclidean distance between cluster centre before newly updated be less than preset threshold ε, can be obtained super-pixel segmentation as a result,
Wherein preferably 10 ε-4~3 normal number.
In described (III) step, in order to remove influence of the picture noise to super-pixel segmentation, while enhancing texture image
Segmentation robustness, present invention introduces neighborhood information calculate Gray homogeneity dc(p,Ck):
Wherein, L (p) indicate centered on pixel p, size be (2a+1) × (2a+1) neighborhood territory pixel collection, a be greater than
Natural number equal to 0, preferably 1~10 natural number, lqFor the gray value of pixel q, lkFor the ash of pixel in k-th of image subblock
Spend mean value, weight wpqMeetIt is defined as follows:
Wherein, Z is normalization factor:
γ is the pixel grey scale standard deviation of neighborhood territory pixel collection L (p).
In described (III) step, space length ds(p,Ck) calculate it is as follows:
Wherein, xpAnd ypThe respectively abscissa and ordinate of pixel p, xkAnd ykIt is then respectively kth image subblock
The mean value of middle pixel column coordinate and column coordinate.
A kind of liver neoplasm automatic and accurate Robust Segmentation method characteristic in CT image also resides in, in described (2) step
In, multi-layer iterative segmentation is carried out to liver area with the image superpixel dividing method based on LI-SLIC, is specifically included:
(I) super-pixel coarse segmentation is carried out to liver area using the image superpixel dividing method based on LI-SLIC, wherein
Super-pixel number n to be generated1It is preferred that 50~500 natural number；
(II) each super-pixel P is calculatediThe gray standard deviation σ of middle all pixelsiIf σiGreater than preset threshold σ (σ be greater than
0 normal number, preferably 1~35 normal number), then to super-pixel PiThe corresponding subgraph f of minimum circumscribed rectangleiCarry out LI-
SLIC super-pixel segmentation, wherein super-pixel number n to be generated1It is preferred that 2~10 natural number；
(III) by subgraph fiMiddle PiThe super-pixel segmentation result in region is assigned to Pi, realize PiAn iteration segmentation；
(IV) step (II) and (III) is repeated until the gray standard deviation σ of super-pixel all in imageiRespectively less than it is equal to threshold
The normal number of value σ, σ preferably 1~35；
(V) divide redundancy caused by multi-layer iterative segmentation to eliminate, using the gray feature between neighbouring super pixels into
Row super-pixel merges, and specifically includes: for each super-pixel Pi, calculate PiIt is adjacent super-pixel PjMinimal gray difference μi,
And it obtains and works as μiCorresponding adjacent super-pixel P when minimumopt:
Wherein,WithRespectively super-pixel PiAnd PjGray average, NPiFor PiAdjoining super-pixel collection.If μiLess than pre-
If threshold value μ, then it is assumed that super-pixel PiIt is adjacent super-pixel PoptGray scale very close to, should belong to same target area, thus by its
It merges, wherein the normal number of parameter μ preferably 5~30.
A kind of liver neoplasm automatic and accurate Robust Segmentation method characteristic in CT image also resides in, in described (3) step
In, one classification classifier of normal liver parenchyma/liver neoplasm two of training, and with the trained classifier to liver area
The method classified of each pixel, specifically include:
(I) classifier of one normal liver parenchyma/liver neoplasm two of training classification, method are as follows: divide from abdominal CT image
Not selecting sufficient amount of size is that (2b+1) × (2b+1) (b is the natural number greater than 0, preferably 3~30 natural number) includes
As training image, gray feature and the texture for extracting training image are special for normal liver parenchyma and the subgraph in liver neoplasm region
Sign, wherein gray feature includes average gray, standard deviation and entropy, and textural characteristics include invariable rotary LBP (Local Binary
Patterns, local binary patterns) feature and multiple dimensioned Gabor characteristic, the feature of extraction is inputted into support vector machines, training one
The classifier that a normal liver parenchyma/liver neoplasm two is classified；
(II) for liver area to be detected, centered on each pixel p, selection size is (2b+1) × (2b+1)
Subgraph fp, subgraph f is extracted using step (I) the methodpGray scale and textural characteristics, the feature of extraction is inputted into instruction
The classification classifier of normal liver parenchyma/tumor tissues two perfected is classified, and classification results are assigned to subgraph fpCenter
Pixel p.
Detailed description of the invention
Fig. 1 CT image preprocessing result example, wherein Fig. 1 (a)~Fig. 1 (c) be three width original CT images, Fig. 1 (d)~
Fig. 1 (f) is the result figure for carrying out liver segmentation to it using embodiment of the present invention；
Image superpixel multi-layer iterative segmentation result example of the Fig. 2 based on LI-SLIC, wherein Fig. 2 (a)~Fig. 2 (c) points
The result of super-pixel multi-layer iterative segmentation Wei not be carried out to image shown in Fig. 1 (a)~Fig. 1 (c) using embodiment of the present invention
Figure；
Fig. 3 LBP coding signal, wherein Fig. 3 (a) is that circle shaped neighborhood region pixel chooses signal, and Fig. 3 (b) is image original gradation
Signal, Fig. 3 (c) are binary pattern schematic diagram；
Fig. 4 LBP weight template, wherein Fig. 4 (a)~Fig. 4 (h) is respectively the 8 weight templates for rotating and generating
The Gabor filters of Fig. 5 different directions is illustrated, wherein Fig. 5 (a)~Fig. 5 (f) be respectively 0 °, 30 °, 60 °, 90 °,
The Gabor filter figure in 120 ° and 150 ° 6 directions；
Fig. 6 CT image liver area pixel classification results example, wherein Fig. 6 (a)~Fig. 6 (c) is respectively to use this hair
Bright embodiment is to the result figure that liver area pixel is classified shown in Fig. 1 (d)~Fig. 1 (f)；
Fig. 7 liver neoplasm segmentation result example, wherein Fig. 7 (a)~Fig. 7 (c) is respectively to use embodiment of the present invention pair
The result figure of the progress liver neoplasm segmentation of image shown in Fig. 1 (a)~Fig. 1 (c).
Specific embodiment
Embodiment 1
A kind of liver neoplasm automatic and accurate Robust Segmentation method in CT image, the specific implementation steps are as follows:
(1) original CT image is pre-processed using sparse combination of shapes, obtains liver area therein；Fig. 1 (a)~
Fig. 1 (c) is three width original CT images, and Fig. 1 (d)~Fig. 1 (f) is to carry out pretreated result to it using the present embodiment method；
(2) multi-layer iterative segmentation is carried out to liver area with the image superpixel dividing method based on LI-SLIC, it will
Gray scale and the more consistent region division of texture are same super-pixel in liver, are obtained between liver neoplasm and normal liver parenchyma
Boundary, super-pixel segmentation result are denoted as Si(i=1,2 ..., n), wherein n is super-pixel number；
In described (2) step, the image superpixel dividing method based on LI-SLIC is specifically included:
(I) original image is divided into continuous nonoverlapping image subblock that side length is h, h value is as follows:
Wherein, N is total number of image pixels mesh, n1For super-pixel number to be generated, it is set greater than 0 natural number；
(II) initial cluster center C is determined according to the average gray of all pixels in image subblock and average spatial informationk:
Ck=[lk,xk,yk]
Wherein, k=1,2, Λ, n1, lkFor the gray average of pixel in k-th of image subblock, xkAnd ykIt is then k-th of figure
As the mean value of pixel column coordinate and column coordinate in sub-block；
(III) it calculates with coordinate (xk,yk) centered on, each in subgraph corresponding to the rectangular box that 2h+1 is side length
Pixel p and cluster centre CkDistance metric D (p, Ck):
Wherein, dc(p,Ck) and ds(p,Ck) respectively indicate pixel p and cluster centre CkGray homogeneity and space length,
Parameter m is normal number, and preferably 10~30 normal number adjusts the distance for controlling Gray homogeneity and space length and measures D (p, Ck)
Influence, m value is bigger, and space length influences it bigger, and vice versa, the preferred m=20 of the present embodiment.It makes an uproar to remove image
Influence of the sound to super-pixel segmentation, while enhancing the segmentation robustness of texture image, it introduces neighborhood information and calculates Gray homogeneity dc
(p,Ck):
Wherein, L (p) indicate centered on pixel p, size be (2a+1) × (2a+1) neighborhood territory pixel collection, a be greater than
Natural number equal to 0, the natural number between preferably 1~10, the present embodiment preferred a=4, lqFor the gray value of pixel q, weight
wpqMeetIt is defined as follows:
Wherein, Z is normalization factor:
γ is the pixel grey scale standard deviation of neighborhood territory pixel collection L (p).In addition, space length ds(p,Ck) calculate it is as follows:
Wherein, xpAnd ypThe respectively abscissa and ordinate of pixel p, xkAnd ykIt is then respectively in k-th of image subblock
The mean value of pixel column coordinate and column coordinate.
(IV) cluster is iterated to image pixel using distance metric D, and constantly updates cluster centre Ck, until newly more
The new Euclidean distance between cluster centre before is less than preset threshold ε, can be obtained super-pixel segmentation as a result, this implementation
Example preferably ε=1；
In described (2) step, multilayer is carried out to liver area with the image superpixel dividing method based on LI-SLIC
Grade iterative segmentation, specifically includes:
(I) super-pixel coarse segmentation is carried out to image using the image superpixel dividing method based on LI-SLIC, wherein to be generated
At super-pixel number n1It is preferred that 50~500 natural number, the preferred n of the present embodiment1=200；
(II) each super-pixel P generated is calculatediThe gray standard deviation σ of included all pixelsiIf σiGreater than default threshold
Value σ (normal number of σ preferably 1~35, preferred σ=14 of the present embodiment), then using the image superpixel segmentation based on LI-SLIC
Method is to super-pixel PiThe corresponding subgraph f of minimum circumscribed rectangleiSuper-pixel segmentation is carried out, wherein super-pixel number to be generated
Mesh n1It is preferred that 2~10 natural number, the preferred n of the present embodiment1=6；
(III) by subgraph fiMiddle PiThe super-pixel segmentation result in region is assigned to Pi, realize PiAn iteration segmentation；
(IV) step (II) and (III) is repeated until the gray standard deviation σ of super-pixel all in imageiRespectively less than it is equal to threshold
Value σ；
(V) divide redundancy caused by multi-layer iterative segmentation to eliminate, using the gray feature between neighbouring super pixels into
Row super-pixel merges, and specifically includes: for each super-pixel Pi, calculate PiIt is adjacent super-pixel PjMinimal gray difference μi,
And it obtains and works as μiCorresponding adjacent super-pixel P when minimumopt:
Wherein,WithRespectively super-pixel PiAnd PjGray average, NPiFor PiAdjoining super-pixel collection.If μiLess than pre-
If threshold value μ, then it is assumed that super-pixel PiIt is adjacent super-pixel PoptGray scale very close to, should belong to same target area, thus by its
It merges, wherein the normal number of parameter μ preferably 5~30, preferred μ=11 of the present embodiment；
Fig. 2 (a)~Fig. 2 (c) is to be carried out CT image shown in Fig. 1 (a)~Fig. 1 (c) based on LI- using the present embodiment method
The result of the image superpixel multi-layer iterative segmentation of SLIC, it can be seen that the super-pixel of generation can be effectively bonded in image
Target weak boundary, such as liver neoplasm boundary, vessel borders and soft tissue boundary, and the pixel that each super-pixel is included
Point belongs to same target；
(3) utilize CT image local gray feature and textural characteristics, one normal liver parenchyma of training liver neoplasm two classify
Classifier, and classify to each pixel for the liver area that step (1) obtains with trained classifier, point
Class result is denoted as F, if pixel p is divided into normal liver parenchyma, F (p) value is 1, if being divided into tumor tissues, F (p) is taken
Value is -1；Specifically includes the following steps:
(I) selecting 500 width sizes respectively from abdominal CT images is that (2b+1) × (2b+1) includes normal liver parenchyma and liver
The subgraph of dirty tumor region is as training image, and wherein b is the natural number greater than 0, preferably 3~30 natural number, this implementation
The preferred b=10 of example；
(II) gray feature and textural characteristics of training image are extracted, wherein gray feature includes image grayscale mean value, mark
Quasi- difference and entropy, textural characteristics include invariable rotary LBP feature and multiple dimensioned Gabor characteristic.
Method particularly includes:
Following formulas Extraction image grayscale mean value avg, standard deviation std and entropy ety totally 3 gray scale spies are respectively adopted in (I)
Sign:
Wherein, f indicates training image, IpFor the gray value of pixel p, NfFor the sum of all pixels mesh of image f, HgIndicate gray scale
The probability that grade g occurs in image f, G is image grayscale number of stages, for one 8 gray level images, G=256；
The invariable rotary LBP feature of (II) extraction training image.Method are as follows: 1) using circle shaped neighborhood region to each in image
Pixel carries out LBP coding, specifically includes: firstly, the circle shaped neighborhood region pixel as shown in Fig. 3 (a) is chosen, in figure shown in black rectangle
Pixel is current pixel, and using it as the center of circle, r is radius, and 45 ° are interval, chooses circle shaped neighborhood region pixel shown in grey rectangle in figure,
Wherein r is the natural number greater than 0, then the preferred r=5 of the present embodiment carries out the gray scale of neighborhood territory pixel and Current central pixel
Compare, if the gray value of neighborhood territory pixel is greater than center pixel, corresponding binary pattern value is 1, and otherwise value is 0, Fig. 3
It (b) is image grayscale example, Fig. 3 (c) is obtained binary pattern, finally, in order to ensure obtained LBP coding has rotation
Invariance rotates binary pattern and as shown in Figure 48 kinds of obtained weight patterns and is weighted summation, takes minimum value therein
It in total include 36 kinds of different coding values as LBP encoded radio；2) probability statistics are carried out to the LBP encoded radio of image pixel, obtained
The invariable rotary LBP feature of one 36 dimension；
The multiple dimensioned Gabor characteristic of (III) extraction training image.Method are as follows: using the filter pair of different scale and direction
Training image carries out Gabor filtering, the present embodiment using 6 directions and size as shown in Fig. 5 (a)~Fig. 5 (f) be respectively 21 ×
21 and 41 × 41 12 Gabor filters are filtered image, then extract the mean value and variance of filtering image, amount to
24 characteristic values；
(III) the 3+36+24=63 dimensional feature of extraction is inputted into support vector machines, one normal liver parenchyma/liver of training
The classifier that tumour two is classified；
(IV) for liver area to be detected, centered on each pixel p, selection size is (2b+1) × (2b+1)
Subgraph fp, subgraph f is extracted using step (II) the methodpGray feature and textural characteristics, the feature of extraction is defeated
Enter the trained classification of normal liver parenchyma/tumor tissues two classifier to classify, and classification results are assigned to subgraph fp's
Central pixel point p, classification results are denoted as F, if pixel p is divided into normal liver parenchyma, F (p) value is 1, if being divided into swollen
Tumor tissue, then F (p) value is -1；
Fig. 6 (a)~Fig. 6 (c) is the knot classified using the present embodiment to liver area shown in Fig. 1 (d)~Fig. 1 (f)
Fruit, wherein white area indicates that normal liver parenchyma, black region indicate liver neoplasm, and gray area is uncorrelated background；
(4) according to liver area pixel classification results F, the super-pixel generated to step (2) is classified, and is obtained final
Liver neoplasm Accurate Segmentation as a result, method are as follows: for each super-pixel Si, calculate its all pixels classification knot for being included
The weighted sum of fruit:
Wherein, weight wpIt is defined respectively as with normalization factor M:
Wherein, dpFor pixel p and super-pixel SiMass center between Euclidean distance, dmaxFor super-pixel SiMiddle all pixels
Maximum Euclidean distance of the point apart from its mass center, pixel p is closer apart from super-pixel mass center, weight wpValue is bigger, the pixel
It is also bigger to the contribution of super-pixel classification.If the weighted sum λ being calculated by above-mentioned formulaiLess than 0, then by its corresponding super picture
Plain SiLabeled as tumor region, otherwise it is marked as normal liver parenchyma.
Fig. 7 (a)~Fig. 7 (c) be using the present embodiment to CT image shown in Fig. 1 (a)~Fig. 1 (c) be split as a result,
It can be seen that wherein form of diverse, obscurity boundary multiple liver neoplasm regions be complete effective Ground Split.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Within mind and principle, any modification, equivalent substitution, improvement and etc. done be should be included within the scope of the present invention.
Claims (10)
1. a kind of liver neoplasm automatic and accurate Robust Segmentation method in CT image, which comprises the following steps:
(1) CT image is pre-processed, obtains the liver area in image；
(2) multi-layer iterative segmentation is carried out to liver area with the image superpixel dividing method based on LI-SLIC, by liver
Middle gray scale and the more consistent region division of texture are same super-pixel, obtain the side between liver neoplasm and normal liver parenchyma
Boundary, super-pixel segmentation result are denoted as Si, i=1,2 ..., n, wherein n is super-pixel number；
(3) utilize CT image local gray feature and textural characteristics, one normal liver parenchyma of training point classified of liver neoplasm two
Class device, and classify with trained classifier to each pixel for the liver area that step (1) obtains, classification knot
Fruit is denoted as F, if pixel p is divided into normal liver parenchyma, F (p) value is 1, if being divided into tumor tissues, F (p) value
It is -1；
(4) according to liver area pixel classification results F, the super-pixel generated to step (2) is classified, and obtains final liver
Dirty tumour Accurate Segmentation is as a result, method are as follows: for each super-pixel Si, calculate its all pixels classification results for being included
Weighted sum:
Wherein, weight wpIt is defined respectively as with normalization factor M:
Wherein dpFor pixel p and super-pixel SiMass center between Euclidean distance, dmaxFor super-pixel SiMiddle all pixels point away from
Maximum Euclidean distance from its mass center, pixel p is closer apart from super-pixel mass center, weight wpValue is bigger, and the pixel is to super
The contribution of pixel classifications is also bigger；If the weighted sum λ being calculated by above-mentioned formulaiLess than 0, then by its corresponding super-pixel Si
Labeled as tumor region, it is otherwise marked as normal liver parenchyma.
2. the liver neoplasm automatic and accurate Robust Segmentation method in a kind of CT image as described in claim 1, it is characterised in that:
In described (1) step, preprocess method is that sparse combination of shapes or figure cut algorithm.
3. the liver neoplasm automatic and accurate Robust Segmentation method in a kind of CT image as described in claim 1, it is characterised in that:
In described (2) step, the image superpixel dividing method based on LI-SLIC, specifically includes the following steps:
(I) original image is divided into continuous nonoverlapping image subblock that side length is h, h value is as follows:
Wherein, N is total number of image pixels mesh, n1For super-pixel number to be generated, it is set greater than 0 natural number；
(II) initial cluster center C is determined according to the average gray of all pixels in image subblock and average spatial informationk:
Ck=[lk,xk,yk]
Wherein, k=1,2 ..., n1, lkFor the gray average of pixel in k-th of image subblock, xkAnd ykIt is then k-th of image subblock
The mean value of middle pixel column coordinate and column coordinate；
(III) it calculates with coordinate (xk,yk) centered on, each pixel in subgraph corresponding to the rectangular box that 2h+1 is side length
Point p and cluster centre CkDistance metric D (p, Ck):
Wherein, dc(p,Ck) indicate pixel p and cluster centre CkGray homogeneity, by pixel p local neighborhood information calculate
It obtains, ds(p,Ck) indicate pixel p and cluster centre CkSpace length, parameter m be normal number, for controlling Gray homogeneity
It adjusts the distance with space length and measures D (p, Ck) influence, m value is bigger, and space length influences it bigger, and vice versa；
(IV) distance metric D (p, C are usedk) cluster is iterated to image pixel, and constantly update cluster centre Ck, until new
The Euclidean distance between cluster centre before updated is less than preset threshold ε, can be obtained super-pixel segmentation result.
4. the liver neoplasm automatic and accurate Robust Segmentation method in a kind of CT image as claimed in claim 3, it is characterised in that:
In described (III) step, Gray homogeneity dc(p,Ck) it is calculated by following formula:
Wherein, L (p) is indicated centered on pixel p, size is the neighborhood territory pixel collection of (2a+1) × (2a+1), and a is more than or equal to 0
Natural number, lqFor the gray value of pixel q, lkFor the average gray of pixel in k-th of image subblock, weight wpqMeetIt is defined as follows:
Wherein, Z is normalization factor:
γ is the pixel grey scale standard deviation of neighborhood territory pixel collection L (p).
5. the liver neoplasm automatic and accurate Robust Segmentation method in a kind of CT image as claimed in claim 3, it is characterised in that:
In described (III) step, space length ds(p,Ck) it is calculated by following formula:
Wherein, xpAnd ypThe respectively abscissa and ordinate of pixel p, xkAnd ykIt is then respectively pixel in k-th of image subblock
The mean value of row coordinate and column coordinate.
6. the liver neoplasm automatic and accurate Robust Segmentation method in a kind of CT image as claimed in claim 4, it is characterised in that:
The normal number that the m is 10~30, the ε are 10-4~3 normal number, the natural number that a is 1~10.
7. the liver neoplasm automatic and accurate Robust Segmentation method in a kind of CT image as described in claim 1, it is characterised in that:
In described (2) step, multi-layer iteration is carried out to liver area with the image superpixel dividing method based on LI-SLIC
Segmentation, specifically includes the following steps:
(I) super-pixel coarse segmentation is carried out to liver area using the image superpixel dividing method based on LI-SLIC, wherein to be generated
At super-pixel number n1For 50~500 natural number；
(II) each super-pixel P generated is calculatediThe gray standard deviation σ of included all pixelsiIf σiGreater than preset threshold σ,
Then to super-pixel PiThe corresponding subgraph f of minimum circumscribed rectangleiLI-SLIC super-pixel segmentation is carried out, wherein super picture to be generated
Prime number mesh n1For 2~10 natural number；
(III) by subgraph fiMiddle PiThe super-pixel segmentation result in region is assigned to Pi, realize PiAn iteration segmentation；
(IV) step (II) and (III) is repeated until the gray standard deviation σ of super-pixel all in imageiRespectively less than it is equal to threshold value σ；
(V) divide redundancy caused by multi-layer iterative segmentation to eliminate, surpassed using the gray feature between neighbouring super pixels
Pixel combination specifically includes: for each super-pixel Pi, calculate PiIt is adjacent super-pixel PjMinimal gray difference μi, and obtain
It takes and works as μiCorresponding adjacent super-pixel P when minimumopt:
Wherein,WithRespectively super-pixel PiAnd PjGray average, NPiFor PiAdjoining super-pixel collection；If μiLess than default threshold
Value μ, then it is assumed that super-pixel PiIt is adjacent super-pixel PoptGray scale very close to should belong to same target area, closed
And.
8. the liver neoplasm automatic and accurate Robust Segmentation method in a kind of CT image as claimed in claim 7, it is characterised in that:
The normal number of the σ preferably 1~35, the normal number of the μ preferably 5~30.
9. the liver neoplasm automatic and accurate Robust Segmentation method in a kind of CT image as described in claim 1, it is characterised in that:
In described (3) step, one normal liver parenchyma of training the classifier classified of liver neoplasm two, and trained with this
The method classified to each pixel of liver area of classifier, specifically includes the following steps:
(I) selecting sufficient amount of size respectively from CT image is (2b+1) × (2b+1) swollen comprising normal liver parenchyma and liver
The subgraph in tumor region is as training image, and wherein b is the natural number greater than 0；
(II) gray feature and textural characteristics of training image are extracted, wherein gray feature includes image grayscale mean value, standard deviation
And entropy, textural characteristics include invariable rotary LBP feature and multiple dimensioned Gabor characteristic；
(III) feature of extraction is inputted into support vector machines, the classifier of training one normal liver parenchyma/liver neoplasm two classification；
(IV) for liver area to be detected, centered on each pixel p, the subgraph that size is (2b+1) × (2b+1) is chosen
As fp, subgraph f is extracted using step (II) the methodpGray feature and textural characteristics, the feature of extraction is inputted into training
The good classification classifier of normal liver parenchyma/tumor tissues two is classified, and classification results are assigned to subgraph fpMiddle imago
Vegetarian refreshments p.
10. the liver neoplasm automatic and accurate Robust Segmentation method in a kind of CT image as claimed in claim 9, feature exist
In: the natural number that the b is 3~30.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811553883.3A CN109753997B (en) | 2018-12-19 | 2018-12-19 | Automatic accurate robust segmentation method for liver tumor in CT image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811553883.3A CN109753997B (en) | 2018-12-19 | 2018-12-19 | Automatic accurate robust segmentation method for liver tumor in CT image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109753997A true CN109753997A (en) | 2019-05-14 |
| CN109753997B CN109753997B (en) | 2022-11-22 |
Family
ID=66402800
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811553883.3A Active CN109753997B (en) | 2018-12-19 | 2018-12-19 | Automatic accurate robust segmentation method for liver tumor in CT image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109753997B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110503649A (en) * | 2019-07-05 | 2019-11-26 | 陕西科技大学 | One kind is based on Multi scale U-net and the modified liver segmentation method of super-pixel |
| CN110647939A (en) * | 2019-09-24 | 2020-01-03 | 广州大学 | Semi-supervised intelligent classification method and device, storage medium and terminal equipment |
| CN110706794A (en) * | 2019-09-26 | 2020-01-17 | 中国科学院深圳先进技术研究院 | Medical image processing system and medical image processing method |
| CN111476292A (en) * | 2020-04-03 | 2020-07-31 | 北京全景德康医学影像诊断中心有限公司 | Small sample element learning training method for medical image classification processing artificial intelligence |
| CN115148330A (en) * | 2022-05-24 | 2022-10-04 | 中国医学科学院北京协和医院 | POP treatment scheme forming method and system |
| CN116188786A (en) * | 2023-05-04 | 2023-05-30 | 潍坊医学院附属医院 | Image segmentation system for hepatic duct and biliary tract calculus |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105957063A (en) * | 2016-04-22 | 2016-09-21 | 北京理工大学 | CT image liver segmentation method and system based on multi-scale weighting similarity measure |
| US20180114313A1 (en) * | 2016-10-21 | 2018-04-26 | Yuan Feng | Medical Image Segmentation Method and Apparatus |
| CN108596887A (en) * | 2018-04-17 | 2018-09-28 | 湖南科技大学 | A kind of abdominal CT sequence image liver neoplasm automatic division method |
-
2018
- 2018-12-19 CN CN201811553883.3A patent/CN109753997B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105957063A (en) * | 2016-04-22 | 2016-09-21 | 北京理工大学 | CT image liver segmentation method and system based on multi-scale weighting similarity measure |
| US20180114313A1 (en) * | 2016-10-21 | 2018-04-26 | Yuan Feng | Medical Image Segmentation Method and Apparatus |
| CN108596887A (en) * | 2018-04-17 | 2018-09-28 | 湖南科技大学 | A kind of abdominal CT sequence image liver neoplasm automatic division method |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110503649A (en) * | 2019-07-05 | 2019-11-26 | 陕西科技大学 | One kind is based on Multi scale U-net and the modified liver segmentation method of super-pixel |
| CN110647939A (en) * | 2019-09-24 | 2020-01-03 | 广州大学 | Semi-supervised intelligent classification method and device, storage medium and terminal equipment |
| CN110706794A (en) * | 2019-09-26 | 2020-01-17 | 中国科学院深圳先进技术研究院 | Medical image processing system and medical image processing method |
| CN111476292A (en) * | 2020-04-03 | 2020-07-31 | 北京全景德康医学影像诊断中心有限公司 | Small sample element learning training method for medical image classification processing artificial intelligence |
| CN115148330A (en) * | 2022-05-24 | 2022-10-04 | 中国医学科学院北京协和医院 | POP treatment scheme forming method and system |
| CN115148330B (en) * | 2022-05-24 | 2023-07-25 | 中国医学科学院北京协和医院 | POP treatment scheme forming method and system |
| CN116188786A (en) * | 2023-05-04 | 2023-05-30 | 潍坊医学院附属医院 | Image segmentation system for hepatic duct and biliary tract calculus |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109753997B (en) | 2022-11-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109753997A (en) | A kind of liver neoplasm automatic and accurate Robust Segmentation method in CT image | |
| He et al. | Pelvic organ segmentation using distinctive curve guided fully convolutional networks | |
| CN108364294B (en) | Multi-organ segmentation method for abdominal CT image based on superpixels | |
| Yang et al. | CT image denoising with perceptive deep neural networks | |
| CN108596887B (en) | Automatic segmentation method for liver tumor region image in abdominal CT sequence image | |
| CN105957063B (en) | CT image liver segmentation method and system based on multiple dimensioned weighting similarity measure | |
| CN101576997B (en) | Abdominal organ segmentation method based on secondary three-dimensional region growth | |
| US8290223B2 (en) | Method for segmenting digital medical image | |
| Jeyavathana et al. | A survey: analysis on pre-processing and segmentation techniques for medical images | |
| Campadelli et al. | A segmentation framework for abdominal organs from CT scans | |
| CN104268873B (en) | Breast tumor partition method based on nuclear magnetic resonance images | |
| Chen et al. | Pathological lung segmentation in chest CT images based on improved random walker | |
| CN104637056B (en) | Medicine CT image adrenal tumor dividing method based on rarefaction representation | |
| CN106778829B (en) | A kind of image detecting method of the hepar damnification classification of Active Learning | |
| 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 | |
| Fan et al. | Lung nodule detection based on 3D convolutional neural networks | |
| CN106504239B (en) | A kind of method of liver area in extraction ultrasound image | |
| Liu et al. | Automatic segmentation algorithm of ultrasound heart image based on convolutional neural network and image saliency | |
| Kumar et al. | Image processing in diabetic related causes | |
| Dakua | Performance divergence with data discrepancy: a review | |
| Abdel-massieh et al. | Automatic liver tumor segmentation from CT scans with knowledge-based constraints | |
| Wei et al. | A novel method for segmentation of CT head images | |
| Stough et al. | Regional appearance in deformable model segmentation | |
| Seo et al. | Efficient liver segmentation based on the spine | |
| Abdellatif et al. | K2. Automatic pectoral muscle boundary detection in mammograms using eigenvectors segmentation |
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 |