A kind of fast robust automatic division method of abdominal CT sequence image liver
Technical field
The invention belongs to technical field of image processing, relate to the segmentation of abdominal CT sequence image organ, particularly the fast robust auto Segmentation of abdominal CT sequence image liver, can be used for medical image auxiliary diagnosis and treatment.
Background technology
Liver segmentation is the prerequisite realizing liver diseases computer-aided diagnosis and hepatic transplantation preplanning.Utilize split and rebuild the hepatic model that obtains can Auxiliary Liver lesion analysis, cubing, vessels analysis, liver subsection, the work such as medical diagnosis on disease and assessment.CT angiographic image due to resolution high, little to injury of human, can image, reflect liver and lesion locations thereof exactly and be subject to the generally favor of doctor.Due to the image slice quantity very large (the Hepatic CT sequence of every patient is about about 200) that three-dimensional imaging uses, each section of artificial segmentation is very consuming time and segmentation result has very large subjectivity.Therefore, in research abdominal CT sequence image, the fast robust automatic division method of liver is to the Clinics and Practices of liver diseases, significant to the precision and efficiency improving computer-aided diagnosis.
Because liver organ complex structure, out-of-shape, Different Individual differ greatly, and be subject to the impact of noise, skew and histokinesis etc. during imaging, the Hepatic CT sequence image obtained has the feature such as complicacy and diversity usually.In addition, in CT image, the abdominal muscles of liver and surrounding, stomach, diaphram, spleen, the organ such as kidney and heart lack good intensity contrast, and this all brings very large difficulty by the accurate auto Segmentation of liver in CT sequence image.
Existing Hepatic CT sequences segmentation method generally can be divided into based on image and the large class of statistical model two.The simple dividing method based on image refers to and directly uses brightness, texture and other image self informations to carry out the method split, mainly comprise threshold method, cluster, region growing, movable contour model and figure to cut, mainly there is following shortcoming and defect in these cutting techniques: (1) needs to carry out complicated pre-service, comprises surrounding tissue or organs such as removing rib, vertebra, kidney; (2) be difficult to split the CT sequence image that contrast is low, liver boundary is fuzzy.First the method for Corpus--based Method model uses a large amount of CT sequence image establishing target prior models, then the segmentation of current sequence is applied to, these class methods have good segmentation effect for the image that contrast is lower, but poor for out-of-shape liver segmentation effect, and length consuming time, responsive to data initialization and registration.
Summary of the invention
The shortcoming that the present invention has taken into full account above-mentioned prior art, with not enough, its object is to, provide a kind of accurately and fast, the abdominal CT sequence image liver automatic division method of robust.And abdominal CT sequence image liver segmentation method of the present invention can be generalized in the segmentation of other abdomen organs.
Abdominal CT sequence image liver robust automatic division method of the present invention, comprises the following steps:
According to the brightness Probability Characteristics of liver area, set up the Intensity model based on gaussian curve approximation, suppress complex background, outstanding liver area;
Utilize image local information, set up the liver display model of Based PC A, strengthen liver weak boundary further, strengthen the discrimination of liver and background;
Utilize the spatial coherence of contiguous slices, in conjunction with brightness and display model, design of graphics cuts energy function, realizes the fast robust auto Segmentation of liver area.
Described Intensity model is set up and is comprised:
A. manual selected part liver area from list entries, adopts its brightness probability distribution of Gaussian function fitting, and obtains the roughly brightness range [I of this sequence image liver according to the fiducial interval of Gaussian distribution
min, I
max].
B. for liver brightness range [I
min, I
max], build following Intensity model:
001"/>
Wherein, k is the normal number of 0.5 ~ 5, and I is brightness of image,
i minwith
i maxbe respectively minimum value and the maximal value of the liver brightness that step a obtains.I is more close to brightness range [I
min, I
max] central value, Intensity model f
intensity(I) value is larger, shows that this pixel belongs to the probability of liver larger.
Described display model is set up and is comprised:
Centered by each pixel, get the subimage that size is (2n+1) × (2n+1), and represent the feature of this central pixel point with the monochrome information of this subimage, wherein n is the natural number of 1 ~ 8;
Use PCA that partial liver region all pixel neighborhoods of a point subimage that user specifies is carried out to principal component analysis (PCA) and calculates the average of its top n major component, obtain the external appearance characteristic p of liver
pCA_average, wherein N is 1 ~ (2
n+ 1) × (2
n+ 1) natural number;
Adopt distance function as outward appearance similarity measurement, calculate the appearance characteristics difference of all pixels and liver in whole sequence:
d(p)=||p
PCA(p)-p
PCA_average||
Wherein, p
pCAp () represents and carries out to the neighborhood subimage of pixel p the top n major component that PCA feature extraction obtains, d (p) represents major component feature p
pCA(p) and p
pCA_averagebetween distance;
Build the liver display model of Based PC A:
002"/>
Wherein, mean_d
prepresent the average to the distance d (p) that all pixel p calculate.F
pCAp the span of () is 0 ~ 1, f
pCAp probability that this pixel of () larger expression p belongs to liver is larger.
Abdominal CT sequence image liver robust automatic division method feature of the present invention is also, the fast robust auto Segmentation of described liver area comprises:
User specifies initial slice.Meet in sequence that liver area is relatively large and the sectioning image only comprising a liver connected region all can be designated as initial slice, acquiescence is positioned at list entries 1/3rd places from top to bottom, such as, if input CT sequence adds up to 210, then initial slice is defaulted as the 70th, if and the slice numbers of CT sequence can not be divided exactly by 3, then carry out round to division result.
Initial slice is split.First the brightness of this section and display model design of graphics is used to cut energy function:
003"/>
Wherein, P represents all set of pixels in image f; N
prepresent the neighborhood territory pixel collection of pixel p; F
intensity(f
p) and F
pCA(f
p) be respectively brightness penalty term and outward appearance penalty term, calculated by brightness and display model respectively; α and β is respectively the weight controlling brightness and outward appearance penalty term, and span is 0 ~ 1, and meets alpha+beta=1; B (f
p, f
q) be border penalty term, obtained by the gradient calculation between neighbor, be used for control segmentation curve smoothness.Then adopt optimization algorithm to minimize this energy function, obtain segmentation result.Finally get the liver final segmentation result of its largest connected region as this sequence initial slice.
Sequence section is split.The present invention adopts the mode of iteration to take initial slice as starting point respectively other all sections in sequence of partitions up and down.Because the size of liver between contiguous slices and position can not have significant change, therefore, in iterative segmentation process, the liver position information between contiguous slices is also used the accuracy increasing segmentation as energy penalty term.Now, energy function can be expressed as:
004"/>
Wherein, F
locationrepresent the position penalty term calculated by upper a slice segmentation result.The present invention, by utilizing the liver position information of contiguous slices, effectively can remove with liver incoherent analogous tissue, as spleen, pancreas, muscle etc., obtain segmentation result more accurately.
Compared with prior art, the inventive method has following advantage:
The invention provides a kind of adaptive liver sequences dividing method, data characteristics for each patient is set up corresponding brightness and display model and is carried out auto Segmentation in conjunction with the spatial coherence between contiguous slices, effectively can process the liver that shape, brightness are different, there is stronger robustness;
Do not need complicated preprocessing process, do not need to split in advance muscle, rib, vertebra, kidney etc.;
Do not need to carry out loaded down with trivial details training and statistical model builds, splitting speed is very fast;
Abdominal CT sequence image liver segmentation method of the present invention can be generalized in the segmentation application of other organs of abdominal CT sequence image, as the segmentation of spleen and kidney.
Accompanying drawing explanation
The abdominal CT sequence image liver robust automatic division method process flow diagram of Fig. 1 embodiment of the present invention;
The partial liver area schematic that Fig. 2 embodiment of the present invention user specifies;
The user of Fig. 3 embodiment of the present invention specifies liver area gaussian curve approximation result figure;
Fig. 4 embodiment of the present invention liver Intensity model exemplary plot;
Fig. 5 embodiment of the present invention liver display model exemplary plot;
The liver segmentation results exemplary plot of Fig. 6 embodiment of the present invention.
Embodiment
Embodiment 1
Figure 1 shows that the abdominal CT sequence image liver robust automatic division method process flow diagram of embodiment of the present invention.First from the CT sequence of input, selected part liver area sets up liver Intensity model and display model respectively arbitrarily, then cut algorithm in conjunction with brightness and display model utilization figure to split the liver in initial slice, finally adopt the mode of iteration with initial segmentation section for starting point is respectively to the liver of other all sections in upper, downward sequence of partitions.In iterative segmentation process, the figure that the liver position information of upper a slice segmentation result incorporates current slice is cut energy function, to increase the accuracy of segmentation result.Until all section segmentations complete, program end of run, exports segmentation result.
Below in conjunction with Fig. 1, describe abdominal CT sequence image liver robust automatic division method of the present invention in detail with a preferred embodiment.
1. liver Intensity model is set up.Specific implementation step is as follows:
According to CT scan characteristic, the brightness of certain certain organs is roughly positioned at a narrower interval range usually, and in Gaussian distribution on this interval.If liver brightness range is [I in abdominal CT sequence
min, I
max], then liver Intensity model can build as follows:
005"/>
Wherein, k is the normal number of 0.5 ~ 5, and it is brightness of image that the present embodiment is preferably 1.5, I.I is more close to brightness range [I
min, I
max] central value, Intensity model f
intensity(I) value is larger, shows that this pixel belongs to the probability of liver larger.
In order to obtain brightness range [I
min, I
max], the Luminance Distribution of the present invention to the partial liver region that user specifies carries out Gauss curve fitting:
006"/>
Wherein, a represents the center of Gaussian distribution, and b controls the width of Gaussian distribution, and c is the peak value of Gaussian distribution.Fig. 2 is the arbitrary sectioning image in original CT sequence, and wherein rectangular area represents the partial liver region that user specifies.Figure 3 shows that and utilize least-squares algorithm to carry out the result of gaussian curve approximation to this liver area, can see that the brightness probability of liver area can meet Gaussian distribution well.According to the probability theory of Gaussian distribution, the brightness range of [a-b, a+b], [a-2b, a+2b] and [a-3b, a+3b] can cover the pixel of liver area 68%, 95%, 99% respectively.Overlapping owing to there is brightness between liver with its Near tissue most probably, therefore, in order at utmost suppress complex background, the preferred liver brightness range of this programme is [a-b, a+b].Fig. 4 is the Intensity model result adopting the present embodiment to obtain, and wherein the first behavior is from the original image of different CT sequence, the liver Intensity model that the second behavior adopts the present embodiment method to obtain.
2. liver display model is set up.Specific implementation step is as follows:
(1) centered by each pixel, get the subimage of 9 × 9, and represent the feature of this central pixel point with the monochrome information of this subimage;
(2) use PCA that partial liver region all pixels neighborhood of a point subimage that user specifies is carried out to principal component analysis (PCA) and calculates the average of its front 6 major components, obtain the proper vector p of one 1 × 6 dimension thus
pCA_average, i.e. the external appearance characteristic of liver;
(3) adopt Euclidean distance as outward appearance similarity measurement, calculate the external appearance characteristic difference of all pixels and liver in whole sequence:
007"/>
Wherein, p
pCAp () represents and carries out to the subimage of pixel p front 6 features that PCA feature extraction obtains, d (p) represents major component feature p
pCA(p) and p
pCA_averagebetween Euclidean distance;
(4) the liver display model of Based PC A is built:
008"/>
Wherein, mean_d
prepresent the average to the distance d (p) that all pixel p calculate.F
pCAp the span of () is 0 ~ 1, f
pCAp probability that this pixel of () larger expression p belongs to liver is larger.
Fig. 5 is the display model result adopting the present embodiment to obtain, and wherein the first behavior is from the original image of different CT sequence, the liver display model that the second behavior adopts the present embodiment method to obtain.
3. initial slice segmentation.Specific implementation step is as follows:
Adopt after above-mentioned specific implementation method obtains liver brightness and display model, the figure building initial slice cuts energy function:
009"/>
Wherein, P represents all set of pixels in image f; N
prepresent the neighborhood territory pixel collection of pixel p; α and β is respectively and controls penalty term F
intensity(f
p) and F
pCA(f
p) weight, meet alpha+beta=1, preferred α=0.8 of the present embodiment, β=0.2; F
intensity(f
p), F
pCA(f
p) and B (f
p, f
q) be respectively brightness, outward appearance and border penalty term, and be defined as follows respectively:
010"/>
011"/>
012"/>
Wherein
013"/>
I
pfor the brightness value of pixel p, d (p, q) is the Euclidean distance between pixel p and q, Num
pfor the number of pixel in set of pixels P.Then adopt this energy function of max-flow min-cut algorithmic minimizing, obtain segmentation result.Finally get the liver final segmentation result of its largest connected region as this sequence initial slice.
4. sequence section segmentation.Specific implementation step is as follows:
Adopt above-mentioned specific implementation method to split after initial slice, then to distinguish other all sections in sequence of partitions up and down for starting point with initial segmentation section by the mode of iteration.In this iterative segmentation process, a upper section segmentation result is utilized to increase the accuracy of segmentation as liver position penalty term.Now, figure cuts energy function and can be expressed as:
014"/>
Wherein
015"/>
μ is the normal number between 0 ~ 0.5, and the present embodiment preferably 0.1, Dis (p) is the range conversion of a upper section segmentation result.Fig. 6 is the part CT image liver segmentation results adopting the present embodiment to obtain, and wherein the first behavior is from the original image of different CT sequence, the liver segmentation results that the second behavior adopts the present embodiment method to obtain.
Embodiment 2
The method of embodiment 1 is adopted to test 10 Hepatic CT sequences that XHCSU14 database provides, and adopt five error criterions to evaluate test result, comprise: volume aliasing error (VolumetricOverlapError, VOE), relative volume difference (RelativeVolumeDifference, RVD), even symmetrical surface distance (AverageSymmetricSurfaceDistance, ASD), root mean square symmetrical surface distance (RootMeanSquareSymmetricSurfaceDistance, RMSD), and maximum symmetrical surface distance (MaximumSymmetricSurfaceDistance, MSD).
10 cycle testss of XHCSU14 database all derive from Philipsbrilliance64 and arrange multi-layer spiral CT machine, thered is provided by Xiangya Hospital, Central-South China Univ., slice plane number of pixels is 512 × 512, and planar pixel spacing range is 0.53-0.74mm, and interlamellar spacing is 1.0mm.Adopt the segmentation result of VOE, RVD, ASD, RMSD and MSD five error criterions to XHCSU14 database to evaluate, the result obtained is as shown in table 1.Can see, for the CT sequence of 10 different patients, the present invention split the average of error and standard deviation all less, show the liver area in the inventive method energy accurate and effective Ground Split abdominal CT sequence, and there is stronger robustness.
Table 1
The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, within the spirit and principles in the present invention all, any amendment made, equivalent replacement, improvement etc., all should be included within the scope of protection of the invention.