CN104809723A - Three-dimensional liver CT (computed tomography) image automatically segmenting method based on hyper voxels and graph cut algorithm - Google Patents

Abstract

Disclosed is a three-dimensional liver CT image automatically segmenting method based on hyper voxels and the graph cut algorithm. The method comprises analyzing a volume data histogram to adaptively enhance the image contrast; performing primary liver contour segmentation layer by layer through an adaptive threshold and an morphological method, selecting the largest liver segment to compute and extract a liver interest area; selecting seed points on the largest liver segment according to a primary liver contour, modelling foreground and background colors through a Gaussian mixed model; generating hyper voxels on the contrast-enhanced liver interest area through an SLIC (simple linear iterative clustering) algorithm, structuring an undirected weighted graph by taking the hyper voxels as the vertex, and segmenting the undirected weighted graph through the graph cut algorithm; performing postprocessing on segmentation results through a morphological algorithm. The three-dimensional liver CT image automatically segmenting method based on the hyper voxels and the graph cut algorithm can achieve rapid and accurate automatic segmentation of the liver in a three-dimensional abdominal CT image.

Description

The three-dimensional CT image for liver automatic division method of algorithm is cut based on super voxel and figure

Technical field

The present invention relates to field of medical image processing, specifically, relate to the three-dimensional CT image for liver automatic division method cutting algorithm based on super voxel and figure.

Background technology

Primary carcinoma of liver is one of common in the world malignant tumour, has very high M & M.Anatomic information is accurate, resolution is high, sweep time is short, popularity rate advantages of higher owing to having for computed tomography (Computed Tomography, CT) imaging technique, is widely used in the Clinics and Practices of liver cancer.Accurate liver three-dimensional segmentation is a basic work in computer-aided diagnosis, is the important prerequisite of three-dimensional visualization, quantitative test, surgery planning etc.The segmentation of liver is generally rule of thumb completed by hand by doctor clinically at present, and this is time and effort consuming not only, and accuracy varies with each individual.Therefore the liver auto Segmentation of efficient stable becomes study hotspot, and this is for the working strength alleviating doctor, and improving diagnosis speed has great help.

The auto Segmentation realizing liver is a challenging research work, its difficult point is: liver structure is complicated, own form is changeable, the property of there are differences between individuality, the liver organization gray scale caused due to tumour, cirrhosis is uneven, with the obscurity boundary of the histoorgans such as adjacent diaphram, heart, kidney.In recent decades, in CT image for liver segmentation, proposed many methods both at home and abroad, mainly comprised: region growing, active contour, level set, figure have cut, cluster, statistical shape model and probability collection of illustrative plates etc.The advantages such as algorithm of region growing has fast, easy realization, but easily cause segmentation result inaccurate when liver organization gray scale is uneven.The partitioning algorithm of based upon activities profile and level set need provide initial profile and calculation of complex, splitting speed are slower.Although segmentation result more accurately can be obtained based on the partitioning algorithm of model, but the generation of probability collection of illustrative plates or statistical shape model needs a large amount of training images and corresponding artificial segmentation standard, and segmentation result may be caused inaccurate when processing non-standard shapes liver.Figure cuts algorithm and is widely used in medical image segmentation because obtaining globally optimal solution, however directly in units of voxel design of graphics carry out cutting and can cause too high calculated amount and satisfied segmentation result cannot be obtained.Accurately, the segmentation changed fast and automatically is the target of liver segmentation.

Summary of the invention

The object of the invention is to overcome the shortcoming of prior art and deficiency, propose a kind of three-dimensional CT image for liver automatic division method cutting algorithm based on simple linear iteration cluster (Simple Linear Iterative Clustering, SLIC) and figure.The method self-adaptation can strengthen picture contrast, automatically extracts liver area-of-interest and generates super voxel, and with super voxel for elementary cell builds undirected weighted graph, reduce computation complexity and calculated amount greatly, raising treatment effeciency.Automatically can use restraint to segmentation by selected seed point again simultaneously, and adopt gauss hybrid models to set up the color model of liver and background area, avoid the impact of man-machine interactively selected seed point on Algorithm robustness.The method employing figure cuts algorithm and surpasses the undirected weighted graph of voxel to liver area-of-interest and cut, and can realize fast, accurately and the liver segmentation of robotization, thus alleviates working doctor amount, provides auxiliary to medical diagnosis.

Invention increases and a kind ofly surpass based on SLIC the three-dimensional CT image for liver automatic division method that voxel and figure cut algorithm, comprise the following steps:

1.1. histogram analysis is carried out to a routine abdominal CT images I of input, thus self-adaptation strengthens picture contrast, the CT image I ' after the contrast that is enhanced;

1.2. adopt adaptive threshold and morphological method successively to carry out the segmentation of liver initial profile to image I, choose maximum liver section l _max, calculate and liver area-of-interest I ' extracted to image I ' after enhancing _rOI;

1.3. at maximum liver section l _maxon choose liver Seed Points S according to liver initial profile _fwith background Seed Points S _b, adopt gauss hybrid models to set up the color model P of liver and background area _fg, P _bkg;

1.4. to I ' _rOIutilize SLIC algorithm to carry out over-segmentation and generate super voxel;

1.5. with super voxel for summit builds undirected weighted graph G, utilize figure cut algorithm to figure G cut, obtain liver area segmentation result bianry image I _mask;

1.6. utilize morphology opening operation, medium filtering to I _maskdo aftertreatment, obtain level and smooth liver segmentation results.

Described step 1.1 comprises,

1.1.1. analyzing peak value number and peak C T value in image histogram, if there are 2 obvious peak values, is high-contrast image I _highif only there is 1 peak value, is soft image I _low;

1.1.2. threshold range [the I of self-adaptation contrast stretching is calculated _min, I _max], for high-contrast image I _highthere is I _min=(peak1+peak2)/2-60, I _max=peak2+250; For low comparison diagram image I _lowthere is I _min=peak1-60, I _max=peak1+250, wherein peak1, peak2 represent peak C T value;

1.1.3. contrast strengthen is done, I '=255 (I-I _min)/(I _max-I _min).

Described step 1.2 comprises,

1.2.1. adaptive threshold [peak1,400Hu] method is utilized to extract liver Probability Area l for every one deck section l _possible;

1.2.2. to l _possibleadopt morphological erosion successively, choose maximum subregion, holes filling, morphological dilations method, obtain the initial profile l of liver area _liver;

1.2.3. l is chosen _liverthe maximum liver section l of conduct that middle liver area area is maximum _max;

1.2.4. according to every layer of l _livertwo-dimentional liver area bounding box calculate the area-of-interest minimum bounding box ROI that can surround whole three-dimensional liver, extraction liver area-of-interest I image I ' after enhancing ^' _rOI.

Described step 1.3 comprises,

1.3.1. for maximum liver section l _max, at liver initial profile l _liverinside chooses liver Seed Points S at regular intervals _f, l _liverbounding box chooses background Seed Points S at regular intervals with exterior domain _b;

1.3.2. according to Seed Points S _f, S _bcalculate gauss hybrid models P _fg, P _bkg,

$P\left(I_p\right)=\underset{k}{\mathrm{\Σ}}{\mathrm{\ω}}_{k}N\left(I_p\right|{\mathrm{\μ}}_k,{\mathrm{\σ}}_k^2),N\left(I_p\right|{\mathrm{\μ}}_k,{\mathrm{\σ}}_k^2)=1/\sqrt{2\mathrm{\π}{\mathrm{\δ}}^2}\·\exp [-{(I_p-\mathrm{\μ})}^2/2{\mathrm{\σ}}^2]$

Wherein, I _pfor the gray-scale value that voxel is corresponding, k is the number of components of gauss hybrid models, ω, μ and σ ²be respectively the weight of gaussian component, average and variance.

Described step 1.4 comprises,

1.4.1. SLIC clustering algorithm formula is provided, N=kS ³, wherein N is I ^' _rOItotal number of middle voxel, S is super voxel spacing, and k is the last super voxel number generated;

1.4.2. cluster is carried out according to distance measure D,

$D=\sqrt{d_c^2+{(m/S)}^2{d}_s^2}$

Wherein, d _c, d _sbe respectively color and distance spatially, m is weight, and setting m=20, S=3, the gray-scale value of each super voxel is set to the gray average in this super voxel areas.

In described step 1.5, cut algorithm energy equation to publishing picture, r (l _p=0)=-log (P _fg(I _p)), R (l _p=1)=-log (P _bkg(I _p)), B (l _p, l _q)=δ (l _p, l _q)/((I _p-I _q) ²+ 1), wherein R (l _p) and B (l _p, l _q) be respectively area item and border item, $\mathrm{\δ}(l_p,l_q)=\left\{\begin{array}{cc}1& \mathrm{if}{l}_p\≠l_q\\ 0& \mathrm{if}{l}_q=l_q\end{array}\right..$

Compared with prior art, tool has the following advantages and beneficial effect in the present invention:

1. the inventive method automatization level when splitting CT image for liver is high, is complete, sufficient automatic segmentation, avoids the impact on Algorithm robustness that most of partitioning algorithm needs man-machine interactively to cause.

2. the present invention effectively reduces the computation complexity that figure cuts algorithm, improves real-time, a processing speed order of magnitude faster than existing three-dimensional liver automatic division method.

3. the 30 routine clinical abdominal CT images that algorithm of the present invention utilizes MICCAI 2007 Workshop to provide are verified, the CT image that exist the situations such as tumour in lower for contrast all can carry out auto Segmentation quickly and accurately.

Accompanying drawing explanation

Fig. 1: the process flow diagram of the inventive method;

Fig. 2: CT image histogram example in the inventive method, wherein: (a) is high-contrast image histogram, (b) is soft image histogram;

Fig. 3: the process flow diagram choosing maximum liver section and calculating liver region of interest border frame in the inventive method;

Fig. 4: a certain sectioning image in original CT sequence to be split;

Fig. 5: the design sketch after Adaptive contrast enhancement is done to original CT image;

Fig. 6: use adaptive threshold and morphological method to do the bianry image after the segmentation of liver initial profile to the image after strengthening contrast;

Fig. 7: to the design sketch after the image zooming-out liver area-of-interest strengthened after contrast;

Fig. 8: use SLIC algorithm to carry out over-segmentation to the liver region of interest area image extracted and generate super voxel;

Fig. 9: final liver segmentation results bianry image;

Figure 10: to the segmentation effect of the uneven image of liver organization gray scale, wherein (a) is transversal section, and (b) is coronal-plane, and (c) is sagittal plane.

Figure 11: to the segmentation effect with non-standard liver shape image, wherein (a) is transversal section, (b) is coronal-plane, and (c) is sagittal plane.

Embodiment

With concrete instance, leaching process is specifically described by reference to the accompanying drawings.Use view data to come from belly Enhanced CT image in MICCAI 2007Workshop database.The mean size of every routine CT image is 512*512*208 pixel, and average resolution rate is 0.68*0.68*1.6 millimeter.

The process flow diagram cutting the CT image for liver automatic division method of algorithm based on super voxel and figure of the present invention as shown in Figure 1, comprises the following steps:

Step 1, carries out histogram analysis to a routine abdominal CT images I (as shown in Figure 4) of input, and self-adaptation strengthens picture contrast, the CT image I ' (as shown in Figure 5) after the contrast that is enhanced.Concrete implementation step is as follows:

1.1. analyzing peak value number in image histogram, if there are 2 obvious peak values, is high-contrast image I _high(as shown in Figure 2 a) if only there is 1 peak value, be, soft image I _low(as shown in Figure 2 b), each peak C T value peak1, peak2 is recorded;

1.2. threshold range [the I of self-adaptation contrast stretching is calculated _min, I _max],

For high-contrast image I _highhave:

I _min＝(peak1+peak2)/2-60

I _max＝peak2+250

For soft image I _lowhave:

I _min＝peak1-60

I _max＝peak1+250

Do linear contrast's stretching to original image, in image, CT value is less than I _minset to 0, CT value is greater than I _maxbe set to 255, CT value at [I _min, I _max] having in scope:

$I^{\text{'}}=255\frac{I-I_{\min }}{I_{\max }-I_{\min }}$

Step 2, adopts adaptive threshold and morphological method successively to carry out the segmentation of liver initial profile to image I, chooses maximum liver section l _max, calculate three-dimensional liver region of interest ROI (process flow diagram as shown in Figure 3).Concrete implementation step is as follows:

2.1. adaptive threshold method is utilized to extract liver Probability Area l for every one deck section l _possible:

1) threshold value is set as [peak1,400Hu], and voxel CT value is less than peak1 or is greater than setting to 0 of 400Hu, and CT value is set to 255 in threshold range, is target area.The liver Probability Area extracted further comprises the similar tissue of some gray-scale values and organ except liver;

2) actionradius is that the circular mask of 7 is to l _possibledo morphological erosion;

3) retain the maximum subregion be positioned on the left of image, remove other subregions;

4) fill up the cavity of intra-zone, actionradius be 2 circular mask morphological dilations is done to image, finally obtain liver initial profile (as shown in Figure 6).

2.2. l is chosen _liverthe maximum liver section l of conduct that middle liver area area is maximum _max;

2.3. according to every layer of l _livertwo-dimentional liver area bounding box calculate the area-of-interest minimum bounding box ROI that can surround whole three-dimensional liver;

2.4. from the rear image I ' of enhancing, liver area-of-interest I ' is extracted _rOI(as shown in Figure 7).

Step 3, at maximum liver section l _maxon choose target liver Seed Points S _fwith background Seed Points S _b.Gauss hybrid models is adopted to set up the color model P of liver and background area _fg, P _bkg.Concrete implementation step is as follows:

3.1. at liver initial profile l _liverliver Seed Points S is chosen by the interval of 30 pixels in inside _f.The non-zero background Seed Points S of gray-scale value is chosen by the interval of 20 pixels in two-dimentional liver area bounding box outside _b;

3.2. calculate the gray probability distribution of liver and background according to the gray-scale value of seed point set, be expressed as gauss hybrid models P _fg, P _bkg.Gauss hybrid models formula is:

$P\left(I_p\right)=\underset{k}{\mathrm{\Σ}}{\mathrm{\ω}}_{k}N\left(I_p\right|{\mathrm{\μ}}_k,{\mathrm{\σ}}_k^2)$

The number of components k=5 of setting gauss hybrid models.Each gaussian component is expressed as:

$N\left(I_p\right|{\mathrm{\μ}}_k,{\mathrm{\σ}}_k^2)=\frac{1}{\sqrt{2\mathrm{\π}{\mathrm{\δ}}^2}}\exp [-\frac{1}{{\mathrm{\σ}}^2}{(I_p-\mathrm{\μ})}^2]$

Wherein, ω, μ and σ ²be respectively the weight of gaussian component, average and variance,

${\mathrm{\ω}}_k=\frac{N_k}{N}$

${\mathrm{\μ}}_k=\frac{1}{N_k}\underset{k}{\mathrm{\Σ}}{I}_p$

${\mathrm{\σ}}_k^2=\frac{1}{N_k}\underset{k}{\mathrm{\Σ}}{I}_p^2-{\mathrm{\μ}}_k^2$

N represents the number of Seed Points.

Step 4, to the liver area-of-interest I ' strengthened after contrast _rOIutilize SLIC algorithm to carry out over-segmentation and generate super voxel (as shown in Figure 8).Concrete implementation step is as follows:

4.1.N=kS ³, the super voxel spacing S=3 that setting generates.N is I ' _rOItotal number of middle voxel.Then at I ' _rOIin generate k initial cluster center with fixing interval S;

4.2., centered by each cluster centre point, search in the spatial dimension being limited to 2S*2S*2S size.For each voxel being positioned at search volume, estimate D according to minor increment and carry out cluster,

$D=\sqrt{d_c^2+{\left(\frac{m}{S}\right)}^2{d}_s^2}$

$d_c=\sqrt{{(I_j-I_i)}^2}$

$d_s=\sqrt{{(x_j-x_i)}^2+{(y_j-y_i)}^2+{(z_j-z_i)}^2}$

Wherein, d _c, d _sbe respectively color I _i, I _jwith the distance on space [x, y, z].M is weight.Setting m=20, S=3.Each super voxel is a little cluster areas, and its gray-scale value is set to the average of all voxel gray values in this cluster areas.

Step 5, with super voxel for summit V builds undirected weighted graph G.Utilize figure to cut algorithm to cut figure G, obtain liver area segmentation result bianry image I _mask.Concrete implementation step is as follows:

5.1. segmentation problem is converted into minimization of energy equation E (L),

$E\left(L\right)=\mathrm{\α}\underset{p\∈P}{\mathrm{\Σ}}R\left(l_p\right)+\mathrm{\β}\underset{(p,q)\∈N}{\mathrm{\Σ}}B(l_p,l_q)$

α, β are weighting parameters.For each super voxel p, there is l _p{ 0,1}, wherein label 0 represents background to ∈, and label 1 represents prospect (liver).R (l _p) be area item,

R(l _p＝0)＝-log(P _fg(I _p))，R(l _p＝1)＝-log(P _bkg(I _p))

B (l _p, l _q) be border item,

$B(l_p,l_q)=\mathrm{\δ}(l_p,l_q)\·\frac{1}{{(I_p-I_q)}^2+1},\mathrm{\δ}(l_p,l_q)=\left\{\begin{array}{cc}1& \mathrm{if}{l}_p\≠l_q\\ 0& \mathrm{if}{l}_p=l_q\end{array}\right.$

5.2. with super voxel for summit V builds undirected weighted graph G=(V, N), N represents the nonoriented edge collection of figure, adds 2 end point S, T.The weight on each limit is as shown in table 1.Setting S corresponding label 0 (background), T corresponding label 1 (target liver);

The weight on each limit of table 1 figure

5.3. utilize figure to cut algorithm to cut the weighted graph built, obtain liver area segmentation result bianry image I _mask.

Step 6, utilizes morphology opening operation, median filter successively to the smoothing process of segmentation result bianry image.Obtain final three-dimensional liver segmentation results (as shown in Figure 9).

Uneven for liver organization gray scale, to there is tumour CT image, this method all can carry out auto Segmentation quickly and accurately, and segmentation result is as shown in Figure 10,11.

Claims (6)

1. cut the three-dimensional CT image for liver automatic division method of algorithm based on super voxel and figure, it is characterized in that: the method comprises the following steps,

1.1. histogram analysis is carried out to a routine abdominal CT images I of input, thus self-adaptation strengthens picture contrast, the CT image I ' after the contrast that is enhanced;

1.2. adopt adaptive threshold and morphological method successively to carry out the segmentation of liver initial profile to image I, choose maximum liver section l _max, calculate and to enhancing after image I ^'extract liver area-of-interest I ' _rOI;

1.3. at maximum liver section l _maxon choose liver Seed Points S according to liver initial profile _fwith background Seed Points S _b, adopt gauss hybrid models to set up the color model P of liver and background area _fg, P _bkg;

1.4. to I ' _rOIutilize SLIC algorithm to carry out over-segmentation and generate super voxel;

1.5. with super voxel for summit builds undirected weighted graph G, utilize figure cut algorithm to figure G cut, obtain liver area segmentation result bianry image I _mask;

1.6. utilize morphology opening operation, medium filtering to I _maskdo aftertreatment, obtain level and smooth liver segmentation results.

2. the three-dimensional CT image for liver automatic division method cutting algorithm based on super voxel and figure according to claim 1, is characterized in that: described step 1.1 comprises,

1.1.1. analyzing peak value number and peak C T value in image histogram, if there are 2 obvious peak values, is high-contrast image I _highif only there is 1 peak value, is soft image I _low;

1.1.2. threshold range [the I of self-adaptation contrast stretching is calculated _min, I _max], for high-contrast image I _highthere is I _min=(peak1+peak2)/2-60, I _max=peak2+250; For low comparison diagram image I _lowthere is I _min=peak1-60, I _max=peak1+250, wherein peak1, peak2 represent peak C T value;

1.1.3. contrast strengthen is done, I '=255 (I-I _min)/(I _max-I _min).

3. the three-dimensional CT image for liver automatic division method cutting algorithm based on super voxel and figure according to claim 1, is characterized in that: described step 1.2 comprises,

1.2.1. adaptive threshold [peak1,400Hu] method is utilized to extract liver Probability Area l for every one deck section l _possible;

1.2.2. to l _possibleadopt morphological erosion successively, retain maximum subregion, holes filling, morphological dilations method, obtain the initial profile l of liver area _liver;

1.2.3. l is chosen _liverthe maximum liver section l of conduct that middle liver area area is maximum _max;

1.2.4. according to every layer of l _livertwo-dimentional liver area bounding box calculate the area-of-interest minimum bounding box ROI that can surround whole three-dimensional liver, from enhancing after image I ^'middle extraction liver area-of-interest I ' _rOI.

4. the three-dimensional CT image for liver automatic division method cutting algorithm based on super voxel and figure according to claim 1, is characterized in that: described step 1.3 comprises,

1.3.1. for maximum liver section l _max, at liver initial profile l _liverinside chooses liver Seed Points S at regular intervals _f, l _liverbounding box chooses background Seed Points S at regular intervals with exterior domain _b;

1.3.2. according to Seed Points S _f, S _bcalculate gauss hybrid models P _fg, P _bkg,

$P\left(I_p\right)=\underset{k}{\mathrm{\Σ}}{\mathrm{\ω}}_{k}N\left(I_p\right|{\mathrm{\μ}}_k,{\mathrm{\σ}}_k^2),N\left(I_p\right|{\mathrm{\μ}}_k,{\mathrm{\σ}}_k^2)=1/\sqrt{2\mathrm{\π}{\mathrm{\σ}}^2}\·\exp [-{(I_p-\mathrm{\μ})}^2/2{\mathrm{\σ}}^2]$

Wherein, I _pfor the gray-scale value that voxel is corresponding, k is the number of components of gauss hybrid models, ω, μ and σ ²be respectively the weight of gaussian component, average and variance.

5. the three-dimensional CT image for liver automatic division method cutting algorithm based on super voxel and figure according to claim 1, is characterized in that: described step 1.4 comprises,

1.4.1. SLIC clustering algorithm formula is provided, N=kS ³, wherein N is I ' _rOItotal number of middle voxel, S is super voxel spacing, and k is the last super voxel number generated;

1.4.2. cluster is carried out according to distance measure D,

$D=\sqrt{d_c^2+{(m/S)}^2{d}_s^2}$

Wherein, d _c, d _sbe respectively color and distance spatially, m is weight, and setting m=20, S=3, the gray-scale value of each super voxel is set to the gray average in this super voxel areas.

6. the three-dimensional CT image for liver automatic division method cutting algorithm based on super voxel and figure according to claim 1, is characterized in that: in described step 1.5, cutting algorithm energy equation to publishing picture, r (l _p=0)=-log (P _fg(I _p)), R (l _p=1)=-log (P _bkg(I _p)), B (l _p, l _q)=δ (l _p, l _q)/((I _p-I _q) ²+ 1); Wherein, R (l _p) and B (l _p, l _q) be respectively area item and border item, $\mathrm{\δ}(l_p,l_q)=\left\{\begin{array}{ccc}1& \mathrm{if}& l_p\≠l_q\\ 0& \mathrm{if}& l_p=l_q\end{array}\right..$