CN104715484A - Automatic tumor area partition method based on improved level set - Google Patents

Abstract

The invention provides an automatic tumor area partition method based on an improved level set. The automatic tumor area partition method comprises the steps of obtaining an original PET image, containing lesion areas, to be partitioned, preprocessing and locating the original PET image, deciding the preprocessed lesion area PET image to be partitioned, constructing a supergraph according to CT images of the lesion areas and the preprocessed lesion area PET image to be partitioned, initially deciding a general tumor area in the PET image as an initial zero level set, carrying out an improved level set method on the initial zero level set to decide the tumor areas and carrying out edge smoothing on the tumor areas according to morphological algorithm. The automatic tumor area partition method can quickly and accurately partition tumor areas and accordingly helps surgical doctors to diagnose, treat and evaluate therapeutic effect.

Description

Based on the automatic tumor region dividing method of the level set improved

Technical field

The present invention relates to image processing techniques, particularly a kind of automatic tumor region dividing method of the level set based on improving.

Background technology

Cervical carcinoma is one of large malignant tumour of female sex organ modal three, is one of crisis women life and the major malignant tumor affecting quality of life, occupy female sex organ malignant tumour first.According to 2014 worlds cancer report that international cancer research institution (the The InternationalAgency for Research on Cancer) February 3 of World Health Organization (WHO) subordinate delivers in the general headquarters being positioned at Lyons, France, cervical cancer 2 reaches more than 50 ten thousand examples in the new cases in the whole world in 012 year, breast cancer, the carcinoma of the rectum, lung cancer is only second in female malignant, ranked fourth position, the death that the same period, cervical carcinoma caused is more than more than 260,000 people, fatal rate is only second to breast cancer, lung cancer, the carcinoma of the rectum, occupies female cancer mortality ratio the 4th.In less developed country women, cervical carcinoma is modal cancer.In recent years, the incidence of disease of the cervical carcinoma of young woman has and increases trend, becomes one of three large major cancers that young woman easily suffers from.China is cervical cancer pathogenesis and dead big country, and M & M all accounts for 1/3rd of the world.Therefore the Accurate Diagnosis for cervical cancer patient is very important.Positron emission computed tomography (positron emission tomography, and CT scan (Computed Tomography PET), CT) as molecular image means, be the detection means that current clinical tumor field is commonly used, utilize PET/CT to carry out quantitative test to tumour and can provide diagnostic message also auxiliary formulation therapeutic scheme accurately for clinical.

The index of quantitative test the most frequently used clinically is at present standard uptake value (standard uptakevalue, SUV), the radioactive concentration (kBq/ml) that SUV equals focus divided by injected dose (MBq) again divided by body weight (kg), secondly gross tumor volume, i.e. MTV value, being also proved to be can the recurrence of predicting tumors and evaluate its prognosis.But these quantitative targets all depend on accurately delineating of tumor region.This is external in the Radiation treatment plans of cervical carcinoma, also depends on accurately delineating of target area.Consider the poor efficiency of manual segmentation and higher subjectivity, automatic accurate Cervical Tumor segmentation is very necessary.But, compared with other tumours, delineating then in the face of more challenges of Cervical Tumor region: on the one hand, because tumour is identical with the attenuation coefficient of uterine neck essence, be therefore difficult to accurate resolution on CT image; On the other hand, because the position of bladder and uterine neck is very close, and the radioactivity of urine in bladder is greater than or is approximately equal to the radioactivity of tumour, is therefore also difficult to automatically extract in PET image.

Summary of the invention

The invention provides a kind of automatic tumor region dividing method of the level set based on improving, being difficult to the problem in automatic distinguishing tumor region and bladder region for solving in Cervical Tumor segmentation, Cervical Tumor can being split fast and accurately to make user thus assisted surgery doctor carries out diagnoses and treatment and curative effect evaluation.

The automatic tumor region dividing method that the present invention is based on the level set of improvement comprises:

Obtain the original PET image to be split that comprises lesion region and carry out pre-service and location thus determine pretreated lesion region PET image to be split;

According to CT image and the described pretreated lesion region PET image structure hypergraph to be split of lesion region, thus tentatively determine that the rough tumor region in described PET image is initial zero level collection;

Described initial zero level collection is performed to the Level Set Method of improvement thus determines tumor region;

According to morphology operations, edge-smoothing process is performed to described tumor region.

Beneficial effect of the present invention is:

The present invention proposes a kind of automatic tumor region dividing method of the level set based on improving, solve the problem being difficult to automatic distinguishing tumor region and bladder region in Cervical Tumor segmentation, make user can split Cervical Tumor fast and accurately thus assisted surgery doctor carries out diagnoses and treatment and curative effect evaluation, it is fast that the method for the invention has speed, precision is high, the advantage of strong robustness, experimental result shows, this technology accurately automatically can delineate Cervical Tumor, realize tumour and bladder automatic distinguishing, clinical diagnosis and treatment has great practical value.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the automatic tumor region dividing method of the level set that the present invention is based on improvement;

Fig. 2 for the present invention is based on the level set of improvement automatic tumor region dividing method described in locate the schematic diagram of lesion region to be split;

Fig. 3 is for the present invention is based on the schematic diagram during Level Set Method iteration improved described in the automatic tumor region dividing method of the level set of improvement;

Fig. 4 is that application the present invention is based on the segmentation result of tumor region and the schematic diagram of goldstandard in 3 typical cervical carcinoma data of the automatic tumor region dividing method of the level set of improvement;

Fig. 5 is segmentation result and the consistance schematic diagram of goldstandard when carrying out quantitative test of the automatic tumor region dividing method of the level set that the present invention is based on improvement.

Embodiment

Fig. 1 is the process flow diagram of the automatic tumor region dividing method of the level set that the present invention is based on improvement, and as shown in Figure 1, the automatic tumor region dividing method that the present invention is based on the level set of improvement comprises:

S1, obtain the original PET image to be split that comprises lesion region and carry out pre-service and location thus determine pretreated lesion region PET image to be split;

Preferably, described acquisition comprises the original PET image to be split of lesion region and carries out pre-service and location thus determine that pretreated lesion region PET image to be split comprises:

Voxel gray values in the original PET image to be split of lesion region is comprised divided by injected developer dosage and patient body weight to be converted to SUV value to described, carry out gaussian filtering and up-sampling again, to make the resolution of PET image to be split identical with CT image, finally locate according to described SUV value and determine pretreated lesion region PET image to be split.

Preferably, describedly locate according to described SUV value and determine that pretreated lesion region PET image to be split comprises:

Preprocessing process, comprising:

By the gray-scale value of each for PET image voxel by being converted into SUV value divided by the dosage of 18F-FDG of injection and patient body weight, then carrying out gaussian filtering and up-sampling, making its resolution identical with CT image.Meanwhile, CT image also carries out identical gaussian filtering;

And position fixing process, comprising:

Calculate the SUV peak value (SUVpeak comprising each section in the original PET image to be split of lesion region, the mean value of the SUV value of each voxel in 26 neighborhoods representing the corresponding SUVmax also i.e. voxel of maximum SUV value), between adjacent two the minimum SUV values choosing the maximum SUV peak value of more than foot, the section of correspondence is cut into slices as lesion region place thus determines pretreated lesion region PET image to be split.

S2, according to the CT image of lesion region and described pretreated lesion region PET image structure hypergraph to be split, thus tentatively determine that the rough tumor region in described PET image is initial zero level collection;

Preferably, the described CT image according to lesion region and described pretreated lesion region PET image structure hypergraph to be split, thus tentatively determine that the rough tumor region in described PET image is that initial zero level collection comprises:

To the CT image normalization of described pretreated lesion region PET image to be split and lesion region and the result after being multiplied builds hypergraph, recycling fuzzy C-means clustering, morphological erosion and generic threshold value method tentatively determine that the rough tumor region in described PET image is initial zero level collection, comprising:

First build hypergraph, comprising:

Each voxel of hypergraph is by three structural feature, respectively: the normalized SUV value (i.e. SUV/SUVmax) of the corresponding voxel of PET image, the normalized HU value (HU/HUmax, HU represent the CT value of each voxel) of the corresponding voxel of CT image, and their product.According to root tissue specificity, hypergraph can be divided into four parts: the representative tumour that (a) SUV and HU is larger; The high representative bladder that still HU is low of (b) SUV; C () SUV is low but other soft tissues of representative that HU is high; D representative background that () SUV and HU is lower.

Recycling fuzzy C-means clustering is divided into 4 classes to hypergraph, the representative tumour that wherein three features are all larger, the method of morphological erosion is utilized to erode the bladder wall that by mistake may be divided into tumor region, then utilize the clinical SUVmax of conventional 40% to be threshold value to this region, obtain rough tumor region.

S3, the Level Set Method improved is performed to described initial zero level collection thus determines tumor region;

Preferably, describedly the Level Set Method improved is performed to initial zero level collection thus determines that tumor region comprises:

The gradient fields information of pretreated lesion region PET image to be split is joined described initial zero level concentrate, build new EVOLUTION EQUATION, with the resolution identical with the resolution of described original PET image, down-sampling is performed to the rough tumor region in described PET image and obtains initial zero level collection;

According to method of finite difference, successive ignition is carried out in described EVOLUTION EQUATION to described initial zero level collection and determine the final tumor region split;

Preferably, the Level Set Method of described improvement comprises:

Consider that the tumour after gaussian filtering and bladder all present the dark characteristic of middle bright limb, so the gradient fields direction at the edge of tumour and bladder is contrary, the new level set movements equation EVOLUTION EQUATION as described below that therefore can build is:

$\begin{array}{c}\frac{\&PartialD;\mathrm{\&phi;}\left(x\right)}{\&PartialD;t}=\{-\mathrm{\&lambda;}{(I_{\mathrm{\&sigma;}}-c_1)}^2+{(I_{\mathrm{\&sigma;}}-c_2)}^2\}\\ \&CenterDot;\tanh (\cot <\&dtri;\mathrm{\&phi;}\left(x\right),\&dtri;I_{\mathrm{\&sigma;}}>)\&CenterDot;\mathrm{sgn}\left({\mathrm{\&delta;}}_{{\mathrm{\&epsiv;}}_1}\left(\mathrm{\&phi;}\right)\right)\end{array}---\left(1\right)$

Wherein, I _σit is pretreated lesion region PET image to be split, (variance of gaussian kernel is σ), φ is the function that initial zero level collection or initial zero level collection carry out the level set after several times iteration, and <*> represents described I _σand the angle between the gradient vector of φ, | * | represent the amplitude of described vector, c ₁and c ₂represent the average gray value of the voxel inside and outside initial zero level collection (or initial zero level collection) respectively, that is:

$\left\{\begin{array}{c}{c}_1\left(\mathrm{\&phi;}\right)=\frac{{\&Integral;}_{\mathrm{\&Omega;}}{I}_{\mathrm{\&sigma;}}(x,y,z)\&CenterDot;H_{{\mathrm{\&epsiv;}}_2}\left(\mathrm{\&phi;}(x,y,z)\right)\mathrm{dxdydz}}{{\&Integral;}_{\mathrm{\&Omega;}}{H}_{{\mathrm{\&epsiv;}}_2}\left(\mathrm{\&phi;}(x,y,z)\right)\mathrm{dzdydz}}\\ c_2\left(\mathrm{\&phi;}\right)=\frac{{\&Integral;}_{\mathrm{\&Omega;}}{I}_{\mathrm{\&sigma;}}(x,y,z)\&CenterDot;(1-H_{{\mathrm{\&epsiv;}}_2}\left(\mathrm{\&phi;}(x,y,z)\right))\mathrm{dxdydz}}{{\&Integral;}_{\mathrm{\&Omega;}}(1-H_{{\mathrm{\&epsiv;}}_2}\left(\mathrm{\&phi;}(x,y,z)\right))\mathrm{dxdydz}}\end{array}\right.---\left(2\right)$

Wherein, Ω represents image-region,

$H_{\mathrm{\&epsiv;}}\left(z\right)=\left\{\begin{array}{cc}1& \mathrm{ifz}>\mathrm{\&epsiv;}\\ 0& \mathrm{ifz}<\mathrm{\&epsiv;}\\ 1/2[1+z/\mathrm{\&epsiv;}+1/\mathrm{\&pi;}\sin (\mathrm{\&pi;z}/\mathrm{\&epsiv;})]& \mathrm{if}\left|z\right|<\mathrm{\&epsiv;}\end{array}\right.---\left(3\right)$

δ function obtains by smooth function δ ε is below approximate:

${\mathrm{\&delta;}}_{\mathrm{\&epsiv;}}\left(x\right)=\left\{\begin{array}{cc}\frac{1}{2\mathrm{\&epsiv;}}[1+\cos \left(\frac{\mathrm{\&pi;x}}{\mathrm{\&epsiv;}}\right)],& \left|x\right|\&le;\mathrm{\&epsiv;}\\ 0,& \left|x\right|>\mathrm{\&epsiv;}\end{array}\right.---\left(4\right)$

Φ is the function of level set after the initial zero level collection of adjustment or initial zero level collection carry out several times iteration and makes it serialization, needs after each iteration to carry out following gaussian filtering to level set function: φ=G _σ* φ,

Wherein, G _σthe gaussian kernel of to be variance be σ, * represents convolution algorithm, and initial function φ 0 is defined as follows:

${\mathrm{\&phi;}}_0\left(x\right)=\left\{\begin{array}{c}{c}_0,\mathrm{ifx}\&Element;R_0\\ -c_0,\mathrm{otherwise}\end{array}\right.,---\left(5\right)$

Wherein, c0 is normal number, and R0 is the rough tumor region that down-sampling to original PET image resolution obtains;

Consider for 3-D view, level set function φ (x, y, z, t) can discretely turn to , wherein (i, j, k) is volume coordinate, and n is time coordinate, then EVOLUTION EQUATION can discretely turn to:

$\frac{{\mathrm{\&phi;}}_{i,j,k}^{n+1}-{\mathrm{\&phi;}}_{i,j,k}^n}{\mathrm{\&Delta;t}}=L\left({\mathrm{\&phi;}}_{i,j,k}^n\right),$

Wherein L is on the right of the equal sign of EVOLUTION EQUATION, then evolutionary process can carry out iteration by following formula:

${\mathrm{\&phi;}}_{i,j,k}^{n+1}={\mathrm{\&phi;}}_{i,j,k}^n+\mathrm{\&Delta;t}\&CenterDot;L\left({\mathrm{\&phi;}}_{i,j,k}^n\right),n=\mathrm{0,1,2},...$

Fig. 3 is the schematic diagram of iterative process.

S4, according to morphology operations, edge-smoothing process is performed to described tumor region.

Preferably, describedly according to morphology operations, edge-smoothing process is performed to described tumor region and comprises:

Chondritic element is used to eliminate edge protuberance and filling cavity to described tumor region execution morphology opening operation and closed operation;

Preferably, wherein concrete morphology operations refers to and utilizes chondritic element to utilize morphology opening operation and closed operation to eliminate edge protuberance to the tumor region that S103 obtains, and radius is that the chondritic element s (x, y, z) of r is as follows:

Described utilization chondritic element performs morphology opening operation to described tumor region and edge protuberance is eliminated in closed operation and filling cavity comprises: use formula (6) and (7) to carry out opening operation and closed operation process respectively:

Opening operation process:

Closed operation process: $\mathrm{Close}=M\&CenterDot;S=(M\&CirclePlus;B)\&CircleTimes;B---\left(7\right)$

Wherein, represent dilation operation symbol, represent erosion operation symbol, M represents the bianry image of described tumor region.

It should be noted that, for verifying validity of the present invention and practicality, we test on clinical PET/CT image, and goldstandard is by the mean value of two expert's manual segmentation results.

By a large amount of experiments, Dice likeness coefficient (i.e. Dice similarity coefficient, weigh the Duplication between segmentation result and goldstandard) be 91.80 ± 2.46%, goldstandard and the segmentation result of Hausdorff distance (weighing the most very much not matching degree between segmentation result and goldstandard) to be 77.79 ± 2.18mm, Fig. 4 be three groups of typical images; Fig. 5 (a) and (b) quantitative target SUVmean for utilizing this technology to split the tumour obtained, between MTV and goldstandard

Bland-Altman schemes, and the high consistency of this technology and goldstandard is described.Experiment shows, our method is well positioned to meet clinical diagnosis and the auxiliary demand formulating treatment plan, has huge practical value.

Last it is noted that above each embodiment is only in order to illustrate technical scheme of the present invention, be not intended to limit; Although with reference to foregoing embodiments to invention has been detailed description, those of ordinary skill in the art is to be understood that: it still can be modified to the technical scheme described in foregoing embodiments, or carries out equivalent replacement to wherein some or all of technical characteristic; And these amendments or replacement, do not make the essence of appropriate technical solution depart from the scope of various embodiments of the present invention technical scheme.

Claims (5)

1., based on an automatic tumor region dividing method for the level set improved, it is characterized in that, comprising:

Obtain the original PET image to be split that comprises lesion region and carry out pre-service and location thus determine pretreated lesion region PET image to be split;

According to CT image and the described pretreated lesion region PET image structure hypergraph to be split of lesion region, thus tentatively determine that the rough tumor region in described PET image is initial zero level collection;

Described initial zero level collection is performed to the Level Set Method of improvement thus determines tumor region;

According to morphology operations, edge-smoothing process is performed to described tumor region.

2. the automatic tumor region dividing method of the level set based on improving according to claim 1, it is characterized in that, described acquisition comprises the original PET image to be split of lesion region and carries out pre-service and location thus determine that pretreated lesion region PET image to be split comprises:

Voxel gray values in the original PET image to be split of lesion region is comprised divided by injected developer dosage and patient body weight to be converted to SUV value to described, carry out gaussian filtering and up-sampling again, finally locate according to described SUV value and determine pretreated lesion region PET image to be split.

The described CT image according to lesion region and described pretreated lesion region PET image structure hypergraph to be split, thus tentatively determine that the rough tumor region in described PET image is that initial zero level collection comprises:

To the CT image normalization of described pretreated lesion region PET image to be split and lesion region and the result after being multiplied builds hypergraph, recycling fuzzy C-means clustering, morphological erosion and generic threshold value method tentatively determine that the rough tumor region in described PET image is initial zero level collection.

Describedly the Level Set Method improved is performed to initial zero level collection thus determines that tumor region comprises:

The gradient fields information of pretreated lesion region PET image to be split is joined described initial zero level concentrate, build new EVOLUTION EQUATION, with the resolution identical with the resolution of described original PET image, down-sampling is performed to the rough tumor region in described PET image and obtains initial zero level collection;

According to method of finite difference, successive ignition determination tumor region is carried out in described EVOLUTION EQUATION to described initial zero level collection.

3. the automatic tumor region dividing method of level set based on improving according to claim 2, is characterized in that, describedly locates according to SUV value and determines that pretreated lesion region PET image to be split comprises:

Calculate the SUV peak value comprising each section in the original PET image to be split of lesion region, choose section corresponding between adjacent two minimum SUV values of the maximum SUV peak value of more than foot and cut into slices as lesion region place thus determine pretreated lesion region PET image to be split;

Accordingly, described EVOLUTION EQUATION is:

$\begin{array}{c}\frac{\&PartialD;\mathrm{\&phi;}\left(x\right)}{\&PartialD;t}=\{-{(I_{\mathrm{\&sigma;}}-c_1)}^2+{(I_{\mathrm{\&sigma;}}-c_2)}^2\}\\ \&CenterDot;\tanh (\cos <\&dtri;\mathrm{\&phi;}\left(x\right),{\&dtri;I}_{\mathrm{\&sigma;}}>)\&CenterDot;\mathrm{sgn}\left({\mathrm{\&delta;}}_{{\mathrm{\&epsiv;}}_1}\left(\mathrm{\&phi;}\right)\right)\end{array}---\left(1\right)$

Wherein, I _σit is pretreated lesion region PET image to be split, (variance of gaussian kernel is σ), φ is the function that initial zero level collection or initial zero level collection carry out the level set after several times iteration, and <*> represents described I _σand the angle between the gradient vector of φ, | * | represent the amplitude of described vector, c ₁and c ₂represent the average gray value of the voxel inside and outside initial zero level collection (or initial zero level collection) respectively, that is:

$\left\{\begin{array}{c}{c}_1\left(\mathrm{\&phi;}\right)=\frac{{\&Integral;}_{\mathrm{\&Omega;}}{I}_{\mathrm{\&sigma;}}(x,y,z)H_{{\mathrm{\&epsiv;}}_2}\left(\mathrm{\&phi;}(x,y,z)\right)\mathrm{dxdydz}}{{\&Integral;}_{\mathrm{\&Omega;}}{H}_{{\mathrm{\&epsiv;}}_2}\left(\mathrm{\&phi;}(x,y,z)\right)\mathrm{dxdydz}}\\ c_2\mathrm{\&phi;}=\frac{{\&Integral;}_{\mathrm{\&Omega;}}{I}_{\mathrm{\&sigma;}}(x,y,z)\&CenterDot;(1-H_{{\mathrm{\&epsiv;}}_2}\left(\mathrm{\&phi;}(x,y,z)\right))\mathrm{dxdydz}}{{\&Integral;}_{\mathrm{\&Omega;}}(1-H_{{\mathrm{\&epsiv;}}_2}\left(\mathrm{\&phi;}(x,y,z)\right))\mathrm{dxdydz}}\end{array}\right.---\left(2\right)$

Wherein, Ω represents image-region,

$H_{\mathrm{\&epsiv;}}\left(z\right)=\left\{\begin{array}{c}\begin{array}{cc}1& \mathrm{if\; z}>\mathrm{\&epsiv;}\end{array}\\ \begin{array}{cc}0& \mathrm{if\; z}<\mathrm{\&epsiv;}\end{array}\\ 1/2[1+z/\mathrm{\&epsiv;}+1/\mathrm{\&pi;}\sin (\mathrm{\&pi;z}/\mathrm{\&epsiv;})]\mathrm{if}\left|z\right|<\mathrm{\&epsiv;}\end{array}\right.---\left(3\right)$

δ function is by smooth function δ below _εbe similar to and obtain:

${\mathrm{\&delta;}}_{\mathrm{\&epsiv;}}\left(X\right)=\left\{\begin{array}{cc}\frac{1}{2\mathrm{\&epsiv;}}[1+\cos \left(\frac{\mathrm{\&pi;x}}{\text{\&epsiv;}}\right)\text{],}& \text{|x|\&le;\&epsiv;}\\ 0,& \left|x\right|>\mathrm{\&epsiv;}\end{array}\right.$

Φ is the function of level set after the initial zero level collection of adjustment or initial zero level collection carry out several times iteration and makes it serialization, needs after each iteration to carry out following gaussian filtering to level set function:

φ＝G _σ*φ，

Wherein, G _σthe gaussian kernel of to be variance be σ, * represents convolution algorithm, and initial function 0 is defined as follows:

${\mathrm{\&phi;}}_0\left(x\right)=\left\{\begin{array}{c}{c}_0,\mathrm{if\; x}\&Element;R_0\\ -c_0,\mathrm{otherwise}\end{array}\right.---\left(5\right)$

Wherein, c ₀normal number, R ₀for the rough tumor region that down-sampling to original PET image resolution obtains.

4. the automatic tumor region dividing method of level set based on improving according to claim 1, is characterized in that, describedly performs edge-smoothing process according to morphology operations to described tumor region and comprises:

Chondritic element is used to eliminate edge protuberance and filling cavity to described tumor region execution morphology opening operation and closed operation.

5. the automatic tumor region dividing method of the level set based on improving according to claim 4, it is characterized in that, described utilization chondritic element performs morphology opening operation to described tumor region and edge protuberance is eliminated in closed operation and filling cavity comprises: use formula (6) and (7) to carry out opening operation and closed operation process respectively:

Opening operation process:

Closed operation process: $\mathrm{Close}=M\&CenterDot;S=(M\&CirclePlus;B)\&CircleTimes;B---\left(7\right)$

Wherein, represent dilation operation symbol, represent erosion operation symbol, M represents the bianry image of described tumor region.