CN106204514A

CN106204514A - A kind of liver localization method based on three-dimensional CT image and device

Info

Publication number: CN106204514A
Application number: CN201510219291.8A
Authority: CN
Inventors: 贾富仓; 龚本伟; 贺宝春; 胡庆茂
Original assignee: Shenzhen Institute of Advanced Technology of CAS
Current assignee: Shenzhen Institute of Advanced Technology of CAS
Priority date: 2015-04-30
Filing date: 2015-04-30
Publication date: 2016-12-07
Anticipated expiration: 2035-04-30
Also published as: CN106204514B

Abstract

The embodiment of the invention discloses a kind of liver localization method based on CT image and device, the method includes: extract the tissue points feature of test sample three-dimensional CT image；Use thoracic cavity and the separating surface in abdominal cavity in difference model assignment test sample three-dimensional CT image；Separating surface is chosen reference body vegetarian refreshments；In calculating test sample three-dimensional CT image, each tissue points is relative to the relative coordinate of this reference body vegetarian refreshments；For tissue points to be measured, every the random forest decision tree being utilized respectively training in advance classification is classified；In the leaf node of the random forest decision tree fallen in tissue points to be measured, k nearest neighbor model is used to obtain the tissue points to be measured classification results every random forest decision tree；The classification results of all of random forest decision tree is averaged, determines the category attribution of tissue points to be measured according to meansigma methods；According to the category attribution of all tissue points to be measured, determine the tissue points to be measured belonging to liver area.Implement the embodiment of the present invention, the accuracy of liver location can be improved.

Description

A kind of liver localization method based on three-dimensional CT image and device

Technical field

The present invention relates to field of medical image processing, be specifically related to a kind of liver based on three-dimensional CT image location Method and device.

Background technology

Computed tomography (Computed Tomography, CT) be utilize Accurate collimation X-ray beam, Gamma-rays, ultrasound wave etc., the detector high with sensitivity is together made one around a certain position of human body and is connect one Individual profile scanning.The features such as CT, for the inspection of multiple disease, has sweep time fast, and image is clear. At present, the location of the liver in three-dimensional CT image is typically based on probability collection of illustrative plates method for registering.Join based on probability collection of illustrative plates Accurate liver localization method needs three-dimensional CT image sample to have in terms of bodily form form and Target organ form Preferably concordance, random due in terms of the difference of people's liver surface shape and three-dimensional CT image intercepting Property so that localization method based on probability atlas registration cannot be accurately positioned liver position.

Summary of the invention

The embodiment of the present invention provides a kind of liver localization method based on three-dimensional CT image and device, can improve The liver Position location accuracy of three-dimensional CT image.

First aspect, it is provided that a kind of liver localization method based on three-dimensional CT image, including:

Extract the tissue points feature of the three-dimensional CT image of test sample, the tissue points feature bag of described test sample Include the original coordinates of tissue points；

Difference model is used to position thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of described test sample；

The separating surface that the three-dimensional CT image of described test sample is corresponding chooses a reference body vegetarian refreshments, wherein, The X value of the original coordinates of the reference body vegetarian refreshments that described test sample is corresponding and each training sample chosen in advance The X value of the original coordinates of corresponding reference body vegetarian refreshments is identical, the reference body vegetarian refreshments that described test sample is corresponding former The Y value of the original coordinates of the reference body vegetarian refreshments that the Y value of beginning coordinate is corresponding with each training sample chosen in advance Identical；

Calculate the ginseng that the tissue points in the three-dimensional CT image of described test sample is corresponding relative to described test sample Examine the relative coordinate of tissue points, obtain the relative seat of each tissue points in the three-dimensional CT image of described test sample Mark；

For the tissue points to be measured in the three-dimensional CT image of described test sample, it is utilized respectively training in advance classification Every random forest decision tree described tissue points to be measured is classified, so that described tissue points to be measured is respectively Fall in the leaf node of every the random forest decision tree training classification；

In the leaf node of the random forest decision tree fallen in described tissue points to be measured, use k nearest neighbor model to institute State tissue points to be measured to classify, obtain described tissue points to be measured and tie in the classification of every random forest decision tree Really；

The classification results of all of random forest decision tree is averaged, and determines institute according to described meansigma methods Stating the category attribution of tissue points to be measured, described category attribution includes belonging to liver area or belonging to non-liver area；

According to the category attribution of all tissue points to be measured, determine the tissue points to be measured belonging to liver area.

Second aspect, it is provided that a kind of liver positioner based on three-dimensional CT image, including:

First extraction unit, for extracting the tissue points feature of the three-dimensional CT image of test sample, described test The tissue points feature of sample includes the original coordinates of tissue points；

First positioning unit, for using difference model to position thoracic cavity in the three-dimensional CT image of described test sample Separating surface with abdominal cavity；

First chooses unit, for choosing one on the separating surface that the three-dimensional CT image of described test sample is corresponding Individual reference body vegetarian refreshments, wherein, the X value of the original coordinates of the reference body vegetarian refreshments that described test sample is corresponding is with in advance The X value of the original coordinates of the reference body vegetarian refreshments that each training sample of choosing is corresponding is identical, described test sample pair The reference voxel that the Y value of the original coordinates of the reference body vegetarian refreshments answered is corresponding with each training sample chosen in advance The Y value of the original coordinates of point is identical；

First computing unit, the tissue points in the three-dimensional CT image calculating described test sample is relative to institute State the relative coordinate of reference body vegetarian refreshments corresponding to test sample, obtain in the three-dimensional CT image of described test sample The relative coordinate of each tissue points；

First taxon, for for the tissue points to be measured in the three-dimensional CT image of described test sample, divides Described tissue points to be measured is classified by every the random forest decision tree not utilizing training in advance to classify, so that Described tissue points to be measured respectively falls in the leaf node of every the random forest decision tree training classification；

Second taxon, in the leaf node of the random forest decision tree fallen in described tissue points to be measured, Use k nearest neighbor model that described tissue points to be measured is classified, obtain described tissue points to be measured random gloomy at every The classification results of woods decision tree；

First determines unit, for the classification results of all of random forest decision tree is averaged, and root Determine the category attribution of described tissue points to be measured according to described meansigma methods, described category attribution includes belonging to liver district Territory or belong to non-liver area；

Second determines unit, for the category attribution according to all tissue points to be measured, determines and belongs to liver area Tissue points to be measured.

In the embodiment of the present invention, extract the tissue points feature of the three-dimensional CT image of test sample, test sample Tissue points feature includes the original coordinates of tissue points；Use the three-dimensional CT image of difference model assignment test sample Middle thoracic cavity and the separating surface in abdominal cavity；The separating surface that the three-dimensional CT image of test sample is corresponding is chosen a ginseng Examining tissue points, wherein, the X value of the original coordinates of the reference body vegetarian refreshments that test sample is corresponding is every with choose in advance The X value of the original coordinates of the reference body vegetarian refreshments that individual training sample is corresponding is identical, the reference voxel that test sample is corresponding The original coordinates of the reference body vegetarian refreshments that the Y value of the original coordinates selected is corresponding with each training sample chosen in advance Y value identical；Calculate the ginseng that the tissue points in the three-dimensional CT image of test sample is corresponding relative to test sample Examine the relative coordinate of tissue points, obtain the relative coordinate of each tissue points in the three-dimensional CT image of test sample； For the tissue points to be measured in the three-dimensional CT image of test sample, be utilized respectively every of training in advance classification with Tissue points to be measured is classified by machine forest decision tree, so that tissue points to be measured respectively falls in trains classification In the leaf node of every random forest decision tree；Leaf segment the random forest decision tree that tissue points to be measured falls into In point, use k nearest neighbor model that tissue points to be measured is classified, obtain tissue points to be measured at every random forest The classification results of decision tree；The classification results of all of random forest decision tree is averaged, and according to flat Average determines the category attribution of tissue points to be measured, and category attribution includes belonging to liver area or belonging to non-liver district Territory；According to the category attribution of all tissue points to be measured, determine the tissue points to be measured belonging to liver area.This Tissue points to be measured is classified, at body to be measured by bright every the stochastic decision tree being utilized respectively training in advance classification In the leaf node of the random forest decision tree that vegetarian refreshments falls into, k nearest neighbor model is used tissue points to be measured to be carried out point Class, the method using random forest classification and k nearest neighbor model to combine, can improve the classification of tissue points to be measured The accuracy of ownership, such that it is able to improve the accuracy of liver location in three-dimensional CT image.

Accompanying drawing explanation

In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to enforcement In example or description of the prior art, the required accompanying drawing used is briefly described, it should be apparent that, describe below In accompanying drawing be only some embodiments of the present invention, for those of ordinary skill in the art, do not paying On the premise of going out creative work, it is also possible to obtain other accompanying drawing according to these accompanying drawings.

Fig. 1 is the flow chart of a kind of liver localization method based on three-dimensional CT image disclosed in the embodiment of the present invention；

Fig. 2 is the flow process of another kind of liver localization method based on three-dimensional CT image disclosed in the embodiment of the present invention Figure；

Fig. 3 be multiple training samples in the embodiment of the present invention CT image in thoracic cavity and interfacial location, abdominal cavity Result figure；

Fig. 4 is that disclosed in the embodiment of the present invention, the structure of a kind of liver positioner based on three-dimensional CT image is shown It is intended to；

Fig. 5 is the structure of another kind of liver positioner based on three-dimensional CT image disclosed in the embodiment of the present invention Schematic diagram；

Fig. 6 is the structure of another kind of liver positioner based on three-dimensional CT image disclosed in the embodiment of the present invention Schematic diagram；

Fig. 7 is the structure of another kind of liver positioner based on three-dimensional CT image disclosed in the embodiment of the present invention Schematic diagram；

Fig. 8 is the disclosed liver positioning experiment Comparative result figure using different localization method of the embodiment of the present invention.

Detailed description of the invention

Below in conjunction with the accompanying drawing in embodiment of the present invention, the technical scheme in embodiment of the present invention is entered Row clearly and completely describes.Obviously, described embodiment is a part of embodiment of the present invention, Rather than whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art are not having The every other embodiment obtained on the premise of making creative work, all should belong to present invention protection Scope.

The embodiment of the present invention provides a kind of liver localization method based on three-dimensional CT image and device, can improve The liver Position location accuracy of three-dimensional CT image.It is described in detail individually below.

Referring to Fig. 1, Fig. 1 is a kind of liver location side based on three-dimensional CT image disclosed in the embodiment of the present invention The flow chart of method.As it is shown in figure 1, the liver localization method based on three-dimensional CT image described in the present embodiment, Including step:

S101, extracts the tissue points feature of the three-dimensional CT image of test sample, the tissue points feature of test sample Original coordinates including tissue points.

In the embodiment of the present invention, penetrate when the certain thickness aspect in the regions such as the thoracic cavity of human body, abdominal part is carried out X After line (can also be gamma-rays, ultrasound wave ray etc.) scanning, detector receives the X-ray through this aspect, Excitation of X-rays fluorescent material is luminous, and after converting optical signals to the signal of telecommunication, reconvert becomes digital signal input meter Calculation machine processes, and obtains three-dimensional CT image, and wherein, three-dimensional CT image is made up of multiple tissue points, and tissue points is The minimum component unit of three-dimensional CT image, the tissue points feature of test sample can include the original seat of tissue points Mark, the gray value of tissue points, the local histogram of cube bounding box centered by tissue points, tissue points Contextual features etc., wherein, cube bounding box centered by tissue points is taken from the radius of neighbourhood of tissue points It it is cube bounding box of 5 (length of side of corresponding cube bounding box is 2*5+1=11)；The context of tissue points is special Levying is that the tissue points of certain deviation amount around tissue points is carried out gray value extraction, upper and lower as this tissue points Literary composition feature.Some preferred embodiment in, will centered by tissue points, that the radius of neighbourhood is 9 is (corresponding The length of side of cube bounding box is 2*9+1=19) upper as this tissue points of the local histogram of cube bounding box Following traits；The original coordinates of tissue points includes X, Y and Z, for the three-dimensional CT image of human body, and body The gray value span of vegetarian refreshments is from-1024 to 3075.

S102, uses thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of difference model assignment test sample.

In the embodiment of the present invention, the tissue points feature of test sample also includes tissue points gray value, due to three In Vc T image, the grey value profile scope of the tissue points being positioned at Different Organs is different, such as, is positioned at liver The grey value profile scope of the tissue points in region is generally 75-150, is positioned at the gray value of the tissue points of lung areas Distribution is less than-64, and the grey value profile scope of the tissue points being positioned at other organs of abdominal part typically exists -20-250, it is possible to use the difference of the gray value of Different Organs, uses the three of difference model assignment test sample Thoracic cavity and the separating surface in abdominal cavity in Vc T image.

In the embodiment that some are feasible, use the three-dimensional CT image mesothorax of difference model assignment test sample The separating surface in chamber and abdominal cavity specifically may include that

Thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of employing equation below assignment test sample:

L_I ^*=min{z | 0 ＜ Δ R_I(z) ＜ δ, Δ R_I(z)=R_I(z)-R_I(z-1) (1)；

R_I(L_I ^*)=α max{R_I(z), 0 ＜ α ＜ 1} (2)；

Wherein, L_I ^*For thoracic cavity in the three-dimensional CT image of test sample and the separating surface in abdominal cavity, z is test sample The axial number of plies corresponding to three-dimensional CT image axial plane, R_I(z) be the axial number of plies be z axial plane in set in advance Put gray value in region to account for less than the tissue points number presetting gray value axial plane pre-sets voxel in region The ratio of some number, R_I(z-1) be the axial number of plies be (z-1) axial plane in pre-set gray value in region Account for less than the tissue points number presetting gray value and axial plane pre-sets the ratio of tissue points number in region； R_I(L_I ^*) it is that in pre-setting region in separating surface, gray value accounts for boundary less than the tissue points number presetting gray value Face pre-sets the ratio of tissue points number in region.

During the present invention implements, the difference model solved in the differential equation in difference model and numerical computation method has Distinguished.In the present invention difference model refer on transverse section adjacent to CT image the ratio of specific pixel be distributed into Row statistics also makees difference analysis (i.e. R_I(z)—R_I(z-1) separating surface in thoracic cavity and abdominal cavity), is finally obtained.Due to not With thoracic cavity, the diversity (such as the diversity of different people's lung shape) in abdominal cavity of sample, at formula of the present invention (1), (2) combine under definition, can accurately position the thoracic cavity of sample and the separating surface in abdominal cavity, almost Do not affected by sample variation.

In the embodiment of the present invention, the tissue points of CT image midriff tissue (such as liver, spleen, kidney etc.) Gray value be typically larger than 0, be positioned at the gray value of tissue points of pulmonary's (lung belongs to chest area) typically smaller than -64, the liver area of human body is positioned at the right side of the human abdominal cavity half side right hypochondriac region of human body (liver be positioned at), is in The Left half-plane of CT image, and the gray value being positioned at the heart of human body left thoracic cavity is close with liver, can be by Pre-set region and be set to the Left half-plane (i.e. the RHP of human body) of CT image, can be by default gray value It is set to-64, i.e. R_IZ () pre-sets the tissue points of region pulmonary and accounts for for being positioned at and pre-set region in axial plane The ratio of interior tissue points number.For CT image, from abdominal cavity region to chest area, z value is gradually increased, Formula (1) substantially determines the separating surface of pulmonary and liver, and wherein δ represents adjacent two R_I(z) close to time One threshold value, δ is empirical value, may select different δ-value according to the difference of sample, and δ is typically chosen Between 0.002-0.015.0<ΔR_IZ () ＜ δ represents along with the continuation of axial plane z rises, in adjacent two axial planes Pre-set gray value in region to account for less than the tissue points number presetting gray value axial plane pre-sets region Difference DELTA R of the ratio of interior tissue points number_IZ () will not significantly change, and think reached thoracic cavity and abdominal cavity point Interface, i.e. liver and the separating surface of lung.

In certain span of z, the z value meeting formula (1) can have multiple, is meeting the z of condition In value, choose the z of minimum as thoracic cavity in three-dimensional CT image and the separating surface in abdominal cavity, simultaneously because the breast of sample The separating surface L in chamber and abdominal cavity_I ^*With R_IThere is certain error, therefore root in the z axial number of plies that the maximum of () is corresponding According to formula (2) according to the maximum R determined_IZ being multiplied by a factor alpha on the basis of () correspondence again, wherein α is Empirical value, is typically chosen between 0.4-0.9, and more preferably between 0.4-0.7, formula (2) illustrates boundary Face pre-sets gray value in region account in separating surface less than the tissue points number presetting gray value and pre-set Ratio R of tissue points number in region_I(L_I ^*) with this sample in R_IZ the maximum of () to be more or less the same.

S103, chooses a reference body vegetarian refreshments on the separating surface that the three-dimensional CT image of test sample is corresponding, its In, the X value of the original coordinates of the reference body vegetarian refreshments that test sample is corresponding and each training sample pair chosen in advance The X value of the original coordinates of the reference body vegetarian refreshments answered is identical, the original coordinates of the reference body vegetarian refreshments that test sample is corresponding The Y value of original coordinates of the Y value reference body vegetarian refreshments corresponding with each training sample chosen in advance identical.

In the embodiment of the present invention, the separating surface L that three-dimensional CT image in test sample is corresponding_I ^*On choose a ginseng Examining tissue points, the original coordinates of this reference body vegetarian refreshments can be P (0,0, z₀), this reference body vegetarian refreshments original The X value of coordinate is 0, and Y value is 0, and Z value is z₀, z₀For thoracic cavity in the three-dimensional CT image of test sample and abdominal cavity Separating surface L_I ^*The corresponding axial number of plies.

S104, calculates the reference that the tissue points in the three-dimensional CT image of test sample is corresponding relative to test sample The relative coordinate of tissue points, obtains the relative coordinate of each tissue points in the three-dimensional CT image of test sample.

In the embodiment of the present invention, the tissue points in the three-dimensional CT image of test sample can be calculated relative to test The relative coordinate of the reference body vegetarian refreshments that sample is corresponding, obtains each tissue points in the three-dimensional CT image of test sample Relative coordinate, such as, if the original coordinates of the reference body vegetarian refreshments of the test sample chosen is P (0,0, z₀), Then in the three-dimensional CT image of test sample, the concrete calculation of the relative coordinate of each tissue points can be: if The original coordinates of some tissue points is P₁(x₁, y₁, z₁), then its relative coordinate is P₁'(x₁, y₁, z₁-z₀)。

S105, for the tissue points to be measured in the three-dimensional CT image of test sample, is utilized respectively training in advance and divides Tissue points to be measured is classified by every random forest decision tree of class, so that tissue points to be measured respectively falls in In the leaf node of every random forest decision tree of training classification.

In the embodiment of the present invention, every random forest decision tree of training in advance classification be pass through a large amount of with The training sample of machine trains random forest decision tree classify, it is possible to use classify every of training in advance with Machine forest decision tree for carrying out preliminary category attribution to the tissue points to be measured in test sample, and every random Forest decision tree all includes some leaf nodes, and it is identical that the tissue points in same leaf node all has some Feature, after tissue points to be measured falls into a leaf node of certain random forest decision tree, shows voxel to be measured Point has some identical features with the training tissue points in this leaf node.

S106, in the leaf node of the random forest decision tree fallen in tissue points to be measured, uses k nearest neighbor model pair Tissue points to be measured is classified, and obtains the tissue points to be measured classification results every random forest decision tree.

In the embodiment of the present invention, k nearest neighbor model is the random forest decision tree by falling in tissue points to be measured Immediate K the training tissue points with the relative tertiary location of tissue points to be measured is chosen in leaf node, and according to K Whether the number of the training tissue points belonging to liver area in individual training tissue points exceedes predetermined threshold value judges this The classification results of tissue points to be measured.For example, if in first random forest decision tree, voxel to be measured Point fall in one of them leaf node, in this leaf node, uses k nearest neighbor model to find to be measured with this Immediate K the training tissue points in tissue points locus, calculates in this K training tissue points and belongs to liver district The number of the training tissue points in territory and the ratio of K, this ratio is random gloomy at first as this tissue points to be measured The classification results of woods decision tree.

In the embodiment that some are feasible, at the leaf node of the random forest decision tree that tissue points to be measured falls into In, use k nearest neighbor model that tissue points to be measured is classified, obtain tissue points to be measured at every random forest certainly The classification results of plan tree, may include that

11), in the leaf node that tissue points to be measured falls in every random forest decision tree, choose respectively and treat Survey immediate K the tissue points of relative coordinate of tissue points；

12) calculate respectively in the leaf node that test sample falls in every stochastic decision tree, with voxel to be measured In immediate K the tissue points of relative coordinate of point, belong to the tissue points number of liver area and the ratio of K, Using ratio as classification results, obtain the tissue points to be measured classification results at every stochastic decision tree.

In the embodiment of the present invention, tissue points to be measured, in every stochastic decision tree, has a classification results.

S107, averages to the classification results of all of random forest decision tree, and determines according to meansigma methods The category attribution of tissue points to be measured, category attribution includes belonging to liver area or belonging to non-liver area.

In the embodiment of the present invention, in every random forest decision tree, this tissue points to be measured all has a classification As a result, the classification results of all of random forest decision tree is averaged, if this meansigma methods exceedes default threshold Value, belongs to liver area by this tissue points to be measured, if this meansigma methods is less than predetermined threshold value, this is to be measured Tissue points belongs to non-liver area.

In the embodiment that some are feasible, the classification results of all of random forest decision tree is averaged, And determine that the category attribution of tissue points to be measured may include that according to meansigma methods

The classification results of all of random forest decision tree is averaged, obtains tissue points to be measured all of The average proportions value of random forest decision tree；

Judge that whether average proportions value is more than predetermined threshold value；

If so, tissue points to be measured is belonged to liver area；

If it is not, tissue points to be measured is belonged to non-liver area.

S108, according to the category attribution of all tissue points to be measured, determines the tissue points to be measured belonging to liver area.

In the embodiment of the present invention, determine that in the three-dimensional CT image of test sample, the classification of all tissue points to be measured is returned After genus, it may be determined that belong to the tissue points to be measured of liver area, obtain liver in the three-dimensional CT image of test sample The profile in dirty region.

As shown in Figure 8, Fig. 8 is the disclosed liver positioning experiment using different localization method of the embodiment of the present invention Comparative result figure.Wherein, figure (a) is original image, and figure (b) is the experiment using AdBoost localization method As a result, figure (c) is the experimental result using random forest location, and figure (d) is to use random forest near with K The adjacent experimental result combining location.From the contrast images of Fig. 8, can clearly see, use the most gloomy Woods is combined the experimental result ratio of location and uses AdBoost location and only with random forest location with k nearest neighbor Experimental result to be got well.

In the embodiment of the present invention, extract the tissue points feature of the three-dimensional CT image of test sample, test sample Tissue points feature includes the original coordinates of tissue points；Use the three-dimensional CT image of difference model assignment test sample Middle thoracic cavity and the separating surface in abdominal cavity；The separating surface that the three-dimensional CT image of test sample is corresponding is chosen a ginseng Examining tissue points, wherein, the X value of the original coordinates of the reference body vegetarian refreshments that test sample is corresponding is every with choose in advance The X value of the original coordinates of the reference body vegetarian refreshments that individual training sample is corresponding is identical, the reference voxel that test sample is corresponding The original coordinates of the reference body vegetarian refreshments that the Y value of the original coordinates selected is corresponding with each training sample chosen in advance Y value identical；Calculate the ginseng that the tissue points in the three-dimensional CT image of test sample is corresponding relative to test sample Examine the relative coordinate of tissue points, obtain the relative coordinate of each tissue points in the three-dimensional CT image of test sample； For the tissue points to be measured in the three-dimensional CT image of test sample, be utilized respectively every of training in advance classification with Tissue points to be measured is classified by machine forest decision tree, so that tissue points to be measured respectively falls in trains classification In the leaf node of every random forest decision tree；Leaf segment the random forest decision tree that tissue points to be measured falls into In point, use k nearest neighbor model that tissue points to be measured is classified, obtain tissue points to be measured at every random forest The classification results of decision tree；The classification results of all of random forest decision tree is averaged, and according to flat Average determines the category attribution of tissue points to be measured, and category attribution includes belonging to liver area or belonging to non-liver district Territory；According to the category attribution of all tissue points to be measured, determine the tissue points to be measured belonging to liver area.Use The present invention, be utilized respectively training in advance classification every stochastic decision tree tissue points to be measured is classified, In the leaf node of the random forest decision tree that tissue points to be measured falls into, use k nearest neighbor model that tissue points to be measured is entered Row classification, the method using random forest classification and k nearest neighbor model to combine, can improve tissue points to be measured The accuracy of category attribution, such that it is able to improve the liver Position location accuracy of three-dimensional CT image.

Referring to Fig. 2, Fig. 2 is another kind of liver based on three-dimensional CT image location disclosed in the embodiment of the present invention The flow chart of method.As in figure 2 it is shown, the liver location side based on three-dimensional CT image described in the present embodiment Method, including step:

S201, extracts the tissue points feature of the three-dimensional CT image of multiple training sample, the body of training sample respectively Vegetarian refreshments feature includes original coordinates and the tissue points gray value of tissue points.

In the embodiment of the present invention, before test sample is classified, first have to training sample is instructed Practice classification.Three-dimensional CT image is made up of multiple tissue points, and tissue points is the minimum component unit of three-dimensional CT image, The tissue points feature of training sample can include the original coordinates of tissue points, the gray value of tissue points, with voxel The local histogram of cube bounding box centered by Dian and the contextual feature etc. of tissue points, wherein, with voxel It is the 5 (length of sides of corresponding cube bounding box that cube bounding box centered by Dian is taken from the radius of neighbourhood of tissue points For 2*5+1=11) cube bounding box；The contextual feature of tissue points is to certain deviation amount around tissue points Tissue points carry out gray value extraction, as the contextual feature of this tissue points.Some preferred embodiment party In formula, by centered by tissue points, the radius of neighbourhood be 9 (length of side of corresponding cube bounding box is as 2*9+1=19) The local histogram of cube bounding box as the contextual feature of this tissue points；The original coordinates bag of tissue points Including X, Y and Z, for the three-dimensional CT image of human body, the gray value span of tissue points is from-1024 to 3075.

S202, uses difference model to position thoracic cavity and abdominal cavity in the three-dimensional CT image of each training sample respectively Separating surface.

In the embodiment of the present invention, owing to, in three-dimensional CT image, being positioned at the gray value of the tissue points of Different Organs Distribution is different, and such as, the grey value profile scope of the tissue points being positioned at liver area is generally 75-150, The grey value profile scope of the tissue points being positioned at lung areas is less than-64, is positioned at the tissue points of other organs of abdominal part Grey value profile scope typically at-20-250, it is possible to use the difference of the gray value of Different Organs, it is poor to use Sub-model positions thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of each training sample respectively.

In the embodiment that some are feasible, use the three-dimensional CT image mesothorax of difference model location training sample The separating surface in chamber and abdominal cavity specifically may include that

Thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of employing equation below location training sample:

L_I ^*=min{z | 0 ＜ Δ R_I(z) ＜ δ, Δ R_I(z)=R_I(z)-R_I(z-1) (1)；

R_I(L_I ^*)=α max{R_I(z), 0 ＜ α ＜ 1} (2)；

Wherein, L_I ^*For thoracic cavity in the three-dimensional CT image of training sample and the separating surface in abdominal cavity, z is training sample The axial number of plies corresponding to three-dimensional CT image axial plane, R_I(z) be the axial number of plies be z axial plane in set in advance Put gray value in region to account for less than the tissue points number presetting gray value axial plane pre-sets voxel in region The ratio of some number, R_I(z-1) be the axial number of plies be (z-1) axial plane in pre-set gray value in region Account for less than the tissue points number presetting gray value and axial plane pre-sets the ratio of tissue points number in region； R_I(L_I ^*) it is that in pre-setting region in separating surface, gray value accounts for boundary less than the tissue points number presetting gray value Face pre-sets the ratio of tissue points number in region.

The present invention implement in difference model and numerical computation method in the difference model that solves in the differential equation have Distinguished.Difference model in the present invention refers to the ratio distribution of specific pixel on transverse section adjacent to CT image Carry out adding up and make difference analysis (i.e. R_I(z)—R_I(z-1) separating surface in thoracic cavity and abdominal cavity), is finally obtained.Due to The different thoracic cavity of sample, the diversityes (such as different people pulmonarys, the diversity of liver shape) in abdominal cavity, at this Invention formula (1), (2) combine under definition, can accurately position the thoracic cavity of sample and the boundary in abdominal cavity Face, is little affected by the impact of sample variation.

For example, participating in accompanying drawing 3, Fig. 3 is the CT image mesothorax of multiple training samples in the embodiment of the present invention Chamber and abdominal cavity interfacial positioning result figure；Certain sample a is carried out difference model location, obtains location three-dimensional The image of the interfacial sample a of CT image, as shown in Fig. 3 (a), the solid line in Fig. 3 (a) represents this sample This thoracic cavity and abdominal cavity separating surface.For sample a, calculate axial numbers of plies z pair different in the CT image of sample a In pre-setting region in the axial plane answered, gray value accounts for axle less than the tissue points number presetting gray value (-64) Plane pre-sets ratio R of tissue points number in region_IZ (), obtains the character pixel distribution of sample, as Shown in Fig. 3 (c), in Fig. 3 (c), abscissa is axial number of plies z, and vertical coordinate is R_I(z)。

Curve a from Fig. 3 (c) is it can be seen that R in sample a_IThe z axial number of plies that the maximum of () is corresponding For z_c, z_cIt is the catastrophe point (being not drawn in figure) in curve a, z_cIn corresponding axial plane, the picture of lung areas The ratio of vegetarian refreshments is maximum, if at z_cOn the basis of along with z continuation increase, R_IZ () will start to diminish；But by Separating surface L in the thoracic cavity of sample and abdominal cavity_I ^*With R_IZ axial number of plies z that the maximum of () is corresponding_cExist certain Error, it is therefore desirable at R_IFactor alpha (for sample a, α=0.9) it is multiplied by the basis of the maximum of (z), R_I(L_I ^*) and R_IZ the maximum of () is closer to, and then obtain thoracic cavity and abdominal cavity in the three-dimensional CT image of this sample Separating surface L_I ^*, the z that the vertical line that marks on curve a in solid line on Fig. 3 (a) and Fig. 3 (c) is corresponding is this The separating surface L of sample a_I ^*。

Generally speaking, separating surface L_I ^*Needing to meet two primary conditions: first, the adjacent axle of two-layer up and down is put down In Left half-plane in face, the gray value tissue points number less than-64 accounts in axial plane tissue points in Left half-plane Difference DELTA R of the ratio of number_IZ () will not significantly change, i.e. 0 < Δ R_I(z) ＜ δ, and meet the z value of condition at these In take wherein minimum z value；Secondly, the tissue points number being less than-64 in separating surface in Left half-plane accounts for separating surface Ratio R of tissue points number in middle Left half-plane_I(L_I ^*) with this sample in R_IZ the maximum of () to be more or less the same.

Similarly, another sample b is also carried out difference model location, obtains positioning the separating surface of three-dimensional CT image The image of sample b, as shown in Fig. 3 (b), the solid line on Fig. 3 (b) is the separating surface L of this sample b_I ^*, Dash-dotted gray line on Fig. 3 (b) is R_IThe z axial number of plies that the maximum of () is corresponding is z_cAxial plane.Fig. 3 (c) The z that the vertical line that marks on curve b is corresponding is the thoracic cavity of this sample and the separating surface L in abdominal cavity_I ^*, liver is at this point Under interface, lung is on this separating surface.The separating surface L of sample b_I ^*With catastrophe point z_cSlightly deviation, for Sample b, α=0.8.

S203, chooses a reference body on the separating surface that the three-dimensional CT image of each training sample is corresponding respectively Vegetarian refreshments, wherein, the X value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding is identical, each training The Y value of the original coordinates of the reference body vegetarian refreshments that sample is corresponding is identical.

In the embodiment of the present invention, the separating surface L that three-dimensional CT image at each training sample is corresponding_I ^*Upper select respectively The reference body vegetarian refreshments taken, wherein, the X value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding Identical, the Y value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding is identical.Such as, at first Choosing a reference body vegetarian refreshments on the separating surface that training sample is corresponding, the original coordinates of this reference body vegetarian refreshments is permissible For P₁(0,0, z₁), the X value of the original coordinates of this reference body vegetarian refreshments is 0, and Y value is 0, and Z value is z₁, z₁ It it is the axial number of plies that in the three-dimensional CT image of first training sample, thoracic cavity is corresponding with the separating surface in abdominal cavity；? A reference body vegetarian refreshments, the original coordinates of this reference body vegetarian refreshments is chosen on the separating surface that two training samples are corresponding Can be P₂(0,0, z₂), the X value of the original coordinates of this reference body vegetarian refreshments is 0, and Y value is 0, and Z value is z₂, z₂It it is the axial number of plies that in the three-dimensional CT image of second training sample, thoracic cavity is corresponding with the separating surface in abdominal cavity；? A reference body vegetarian refreshments, the original coordinates of this reference body vegetarian refreshments is chosen on the separating surface that n-th training sample is corresponding Can be P_n(0,0, z_n), the X value of the original coordinates of this reference body vegetarian refreshments is 0, and Y value is 0, and Z value is z_n, z_nIt it is the axial number of plies that in the three-dimensional CT image of second training sample, thoracic cavity is corresponding with the separating surface in abdominal cavity；This In the X value of each training sample chosen be 0, Y value is 0.

S204, calculates the tissue points in the three-dimensional CT image of each training sample respectively relative to tissue points place The relative coordinate of the reference body vegetarian refreshments that training sample is corresponding, obtains in the three-dimensional CT image of each training sample every The relative coordinate of individual tissue points.

In the embodiment of the present invention, calculate respectively tissue points in the three-dimensional CT image of each training sample relative to The relative coordinate of the reference body vegetarian refreshments that tissue points place training sample is corresponding, obtains the three-dimensional of each training sample The relative coordinate of each tissue points in CT image, such as, for first training sample, if the reference chosen The original coordinates of tissue points is P (0,0, z₁), then each voxel in the three-dimensional CT image of the first training sample The concrete calculation of the relative coordinate of point can be: if the original coordinates of some tissue points is P₁(x₁, y₁, z₂), then its relative coordinate is P₁'(x₁, y₁, z₂-z₁)；For second training sample, if the reference chosen The original coordinates of tissue points is P (0,0, z₂), then each voxel in the three-dimensional CT image of the second training sample The concrete calculation of the relative coordinate of point can be: if the original coordinates of some tissue points is P₁(x₁, y₁, z₁), then its relative coordinate is P₁'(x₁, y₁, z₁-z₂)。

S205, randomly selects the liver of the first predetermined number from the three-dimensional CT image of each training sample respectively The tissue points in region and the tissue points of the non-liver area of the second predetermined number, as training tissue points at random.

In the embodiment of the present invention, owing to each training sample is carried out random forest classification, from each training sample The tissue points and second randomly selecting the liver area of the first predetermined number in this three-dimensional CT image respectively is preset The tissue points of the non-liver area of quantity, as training tissue points at random.First predetermined number and second is preset Quantity can pre-set as required, and the number of random training tissue points is generally large, for example, 400,000 random tissue points can be chosen.

S206, has the feature randomly selecting the 3rd predetermined number put back to from the tissue points feature of training sample Build random forest decision tree, for every random forest decision tree, put from random training tissue points The random training tissue points of the 4th predetermined number randomly selected returned is trained classification, so that the 4th presets The random training tissue points of quantity falls in the leaf node of random forest decision tree.

In the embodiment of the present invention, the tissue points feature of training sample includes tissue points coordinate, tissue points gray value, The local histogram of cube bounding box centered by tissue points and the contextual feature etc. of tissue points, tissue points The number of feature can be more than 100, the X-coordinate of such as tissue points, the Y coordinate of tissue points, the Z of tissue points Coordinate, the pixel value of tissue points, cube bounding box centered by tissue points local histogram in gray value Number from-1000 to-950, cube bounding box centered by tissue points local histogram gray value from The number of-950 to-900, cube bounding box centered by tissue points local histogram in gray value from-900 To the number of-850, etc., the 3rd predetermined number can be 10.If the number of training tissue points is 40 at random Ten thousand, the 4th predetermined number can be 20,000.For example, to every random forest decision tree, Ke Yicong The tissue points feature of training sample have that puts back to randomly select 10 features spy as random forest decision tree Levy, from random tissue points, then randomly select 20,000 tissue points be trained classification, when random forest decision-making The tissue points the purest (tissue points such as falling into this node belongs to liver area) of the node of tree, or node When tissue points number is less than 64, stop division, using this node as leaf node.

S207, extracts the tissue points feature of the three-dimensional CT image of test sample, the tissue points feature of test sample Original coordinates including tissue points.

S208, uses thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of difference model assignment test sample.

S209, chooses a reference body vegetarian refreshments on the separating surface that the three-dimensional CT image of test sample is corresponding, its In, the X value of the original coordinates of the reference body vegetarian refreshments that test sample is corresponding and each training sample pair chosen in advance The X value of the original coordinates of the reference body vegetarian refreshments answered is identical, the original coordinates of the reference body vegetarian refreshments that test sample is corresponding The Y value of original coordinates of the Y value reference body vegetarian refreshments corresponding with each training sample chosen in advance identical.

S210, calculates the reference that the tissue points in the three-dimensional CT image of test sample is corresponding relative to test sample The relative coordinate of tissue points, obtains the relative coordinate of each tissue points in the three-dimensional CT image of test sample.

S211, for the tissue points to be measured in the three-dimensional CT image of test sample, is utilized respectively training in advance and divides Tissue points to be measured is classified by every random forest decision tree of class, so that tissue points to be measured respectively falls in In the leaf node of every random forest decision tree of training classification.

S212, in the leaf node of the random forest decision tree fallen in tissue points to be measured, uses k nearest neighbor model pair Tissue points to be measured is classified, and obtains the tissue points to be measured classification results every random forest decision tree.

S213, averages to the classification results of all of random forest decision tree, and determines according to meansigma methods The category attribution of tissue points to be measured, category attribution includes belonging to liver area or belonging to non-liver area.

S214, according to the category attribution of all tissue points to be measured, determines the tissue points to be measured belonging to liver area.

In the embodiment of the present invention, step S207-step S214 is referred to step S101-step S108 of Fig. 1. The embodiment of the present invention repeats no more.

In the embodiment of the present invention, extract the tissue points feature of the three-dimensional CT image of multiple training sample, body respectively Vegetarian refreshments feature includes original coordinates and the tissue points gray value of tissue points；Difference model is used to position each respectively Thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of training sample；Three-dimensional CT image at each training sample A reference body vegetarian refreshments, wherein, the reference body that each training sample is corresponding is chosen respectively on corresponding separating surface The X value of the original coordinates of vegetarian refreshments is identical, the Y value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding Identical；The tissue points calculated respectively in the three-dimensional CT image of each training sample is trained relative to tissue points place The relative coordinate of the reference body vegetarian refreshments that sample is corresponding, obtains in the three-dimensional CT image of each training sample every individual The relative coordinate of vegetarian refreshments；The first predetermined number is randomly selected respectively from the three-dimensional CT image of each training sample The tissue points of liver area and the tissue points of non-liver area of the second predetermined number, as training body at random Vegetarian refreshments；The feature construction randomly selecting the 3rd predetermined number put back to is had from the tissue points feature of training sample Random forest decision tree, for every random forest decision tree, puts back to having from random training tissue points The random training tissue points of the 4th predetermined number randomly selected is trained classification, so that the 4th predetermined number Random training tissue points fall in the leaf node of random forest decision tree；Extract the three dimensional CT figure of test sample The tissue points feature of picture, tissue points feature includes the original coordinates of tissue points；Use difference model assignment test Thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of sample；Corresponding the dividing of three-dimensional CT image in test sample A reference body vegetarian refreshments, wherein, the X of the original coordinates of the reference body vegetarian refreshments that test sample is corresponding is chosen on interface The X value of the original coordinates of the reference body vegetarian refreshments that value is corresponding with each training sample chosen in advance is identical, test specimens The reference that the Y value of the original coordinates of the reference body vegetarian refreshments of this correspondence is corresponding with each training sample chosen in advance The Y value of the original coordinates of tissue points is identical；Calculate test sample three-dimensional CT image in tissue points relative to The relative coordinate of the reference body vegetarian refreshments that test sample is corresponding, obtains in the three-dimensional CT image of test sample every individual The relative coordinate of vegetarian refreshments；For the tissue points to be measured in the three-dimensional CT image of test sample, it is utilized respectively in advance Tissue points to be measured is classified by every random forest decision tree of training classification, so that tissue points to be measured is respectively Fall in the leaf node of every the random forest decision tree training classification；Tissue points to be measured fall into random In the leaf node of forest decision tree, use k nearest neighbor model that tissue points to be measured is classified, obtain voxel to be measured Point is at the classification results of every random forest decision tree；The classification results of all of random forest decision tree is asked Meansigma methods, and the category attribution of tissue points to be measured is determined according to meansigma methods, category attribution includes belonging to liver district Territory or belong to non-liver area；According to the category attribution of all tissue points to be measured, determine and belong to liver area Tissue points to be measured.Use the present invention, the liver Position location accuracy of three-dimensional CT image can be improved.

Referring to Fig. 4, Fig. 4 is a kind of liver based on three-dimensional CT image location dress disclosed in the embodiment of the present invention The structural representation put.As shown in Figure 4, the device described in the present embodiment, including the first extraction unit 301, First positioning unit 302, first choose unit the 303, first computing unit the 304, first taxon 305, Second grouping sheet 306, first determine that unit 307 and second determines unit 308, wherein:

First extraction unit 301, for extracting the tissue points feature of the three-dimensional CT image of test sample, test specimens This tissue points feature includes the original coordinates of tissue points.

In the embodiment of the present invention, penetrate when the certain thickness aspect in the regions such as the thoracic cavity of human body, abdominal part is carried out X After line (can also be gamma-rays, ultrasound wave ray etc.) scanning, detector receives the X-ray through this aspect, Excitation of X-rays fluorescent material is luminous, and after converting optical signals to the signal of telecommunication, reconvert becomes digital signal input meter Calculation machine processes, and obtains three-dimensional CT image, and wherein, three-dimensional CT image is made up of multiple tissue points, and tissue points is The minimum component unit of three-dimensional CT image, the tissue points feature of test sample can include the original seat of tissue points Mark, the gray value of tissue points, the local histogram of cube bounding box centered by tissue points and tissue points Contextual features etc., wherein, cube bounding box centered by tissue points is taken from the radius of neighbourhood of tissue points It it is cube bounding box of 5 (length of side of corresponding cube bounding box is 2*5+1=11)；The context of tissue points is special Levying is that the tissue points of certain deviation amount around tissue points is carried out gray value extraction, upper and lower as this tissue points Literary composition feature.Some preferred embodiment in, will centered by tissue points, that the radius of neighbourhood is 9 is (corresponding The length of side of cube bounding box is 2*9+1=19) upper as this tissue points of the local histogram of cube bounding box Following traits；The original coordinates of tissue points includes X, Y and Z, for the three-dimensional CT image of human body, and body The gray value span of vegetarian refreshments is from-1024 to 3075.

First positioning unit 302, for use in the three-dimensional CT image of difference model assignment test sample thoracic cavity and The separating surface in abdominal cavity.

In the embodiment of the present invention, the tissue points feature of test sample also includes tissue points gray value, due to three In Vc T image, the grey value profile scope of the tissue points being positioned at Different Organs is different, such as, is positioned at liver The grey value profile scope of the tissue points in region is generally 75-150, is positioned at the gray value of the tissue points of lung areas Distribution is less than-64, and the grey value profile scope of the tissue points being positioned at other organs of abdominal part typically exists -20-250, it is possible to use the difference of the gray value of Different Organs, the first positioning unit 302 uses difference model fixed Thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of bit test sample.

Optionally, the first positioning unit 302 with thoracic cavity in the three-dimensional CT image of difference model assignment test sample and The separating surface in abdominal cavity specifically may include that

L_I ^*=min{z | 0 ＜ Δ R_I(z) ＜ δ, Δ R_I(z)=R_I(z)-R_I(z-1) (1)；

R_I(L_I ^*)=α max{R_I(z), 0 ＜ α ＜ 1} (2)；

In the embodiment of the present invention, the tissue points of CT image midriff tissue (such as liver, spleen, kidney etc.) Gray value be typically larger than 0, be positioned at the gray value of tissue points of pulmonary's (lung belongs to chest area) typically smaller than -64, the liver of human body is positioned at the right side of the human abdominal cavity half side right hypochondriac region of human body (liver be positioned at), is in CT figure The Left half-plane of picture, and the gray value being positioned at the heart of human body left thoracic cavity is close with liver, can be by advance Setting area is set to the Left half-plane (i.e. the RHP of human body) of CT image, can be set to by default gray value -64, i.e. R_IZ () pre-sets the tissue points of region pulmonary and accounts for axial plane for being positioned at and pre-set region voxel The ratio of some number.For CT image, from abdominal cavity region to chest area, z value is gradually increased, formula (1) substantially determining the separating surface of pulmonary and liver, wherein δ represents adjacent two R_I(z) close to time one Threshold value, δ is empirical value, may select different δ-value according to the difference of sample, and δ is typically chosen 0.002-0.015 Between.0<ΔR_IZ () ＜ δ represents along with the continuation of axial plane z rises, pre-set district in adjacent two axial planes In territory, gray value accounts for less than the tissue points number presetting gray value and pre-sets tissue points in region in axial plane Difference DELTA R of the ratio of number_IZ () will not significantly change, and think and reached the separating surface in thoracic cavity and abdominal cavity, i.e. liver Separating surface with lung.

First chooses unit 303, for choosing one on the separating surface that the three-dimensional CT image of test sample is corresponding Reference body vegetarian refreshments, wherein, the X value of the original coordinates of the reference body vegetarian refreshments that test sample is corresponding with choose in advance The X value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding is identical, the reference body that test sample is corresponding The original seat of the reference body vegetarian refreshments that the Y value of the original coordinates of vegetarian refreshments is corresponding with each training sample chosen in advance Target Y value is identical.

In the embodiment of the present invention, first chooses unit 303 at separating surface corresponding to the three-dimensional CT image of test sample On a reference body vegetarian refreshments choosing, the original coordinates of this reference body vegetarian refreshments can be P (0,0, z₀), this ginseng The X value of the original coordinates examining tissue points is 0, and Y value is 0, and Z value is z₀, z₀Three dimensional CT figure for test sample The axial number of plies that in Xiang, thoracic cavity is corresponding with the separating surface in abdominal cavity.

First computing unit 304, the tissue points in the three-dimensional CT image calculating test sample is relative to test The relative coordinate of the reference body vegetarian refreshments that sample is corresponding, obtains each tissue points in the three-dimensional CT image of test sample Relative coordinate.

In the embodiment of the present invention, the first computing unit 304 can calculate the body in the three-dimensional CT image of test sample The relative coordinate of the reference body vegetarian refreshments that vegetarian refreshments is corresponding relative to test sample, obtains the three dimensional CT figure of test sample The relative coordinate of each tissue points in Xiang, such as, if the original seat of the reference body vegetarian refreshments of the test sample chosen It is designated as P (0,0, z₀), then the relative coordinate of each tissue points concrete in the three-dimensional CT image of test sample Calculation can be: if the original coordinates of some tissue points is P₁(x₁, y₁, z₁), then its relative coordinate For P₁'(x₁, y₁, z₁-z₀)。

First taxon 305, for for the tissue points to be measured in the three-dimensional CT image of test sample, difference Tissue points to be measured is classified by every the random forest decision tree utilizing training in advance to classify, so that body to be measured Vegetarian refreshments respectively falls in the leaf node of every the random forest decision tree training classification.

Second grouping sheet 306, in the leaf node of the random forest decision tree fallen in tissue points to be measured, adopts With k nearest neighbor model, tissue points to be measured is classified, obtain tissue points to be measured every random forest decision tree Classification results.

Optionally, it is another kind of based on three-dimensional CT image disclosed in the embodiment of the present invention for referring to Fig. 5, Fig. 5 The structural representation of liver positioner, as it is shown in figure 5, the second grouping sheet 306 may include that

Choose subelement 3061, for the leaf node that tissue points to be measured falls in every random forest decision tree In, choose immediate K the tissue points with the relative coordinate of tissue points to be measured respectively.

Computation subunit 3062, for calculating the leaf segment that test sample in every stochastic decision tree falls into respectively In point, with the relative coordinate of tissue points to be measured in immediate K tissue points, belong to the tissue points of liver area The ratio of number and K, using ratio as classification results, obtain tissue points to be measured every stochastic decision tree point Class result.

First determines unit 307, for the classification results of all of random forest decision tree is averaged, and Determine the category attribution of tissue points to be measured according to meansigma methods, category attribution includes belonging to liver area or belonging to non- Liver area.

Optionally, it is another kind of based on three-dimensional CT image disclosed in the embodiment of the present invention for referring to Fig. 6, Fig. 6 The structural representation of liver positioner, as shown in Figure 6, first determines that unit 307 can include that average son is single Unit 3071, judgment sub-unit 3072, first belong to subelement 3073 and the second ownership subelement 3074, wherein:

Average subelement 3071, for the classification results of all of random forest decision tree is averaged, To tissue points to be measured in the average proportions value of all of random forest decision tree.

Judgment sub-unit 3072, is used for judging that whether average proportions value is more than predetermined threshold value.

First ownership subelement 3073, for when judgment sub-unit 3072 judged result is for being, by body to be measured Vegetarian refreshments belongs to liver area.

Second ownership subelement 3074, for when judgment sub-unit 3072 judged result is no, by body to be measured Vegetarian refreshments belongs to non-liver area.

Second determines unit 308, for the category attribution according to all tissue points to be measured, determines and belongs to liver district The tissue points to be measured in territory.

In the embodiment of the present invention, the tissue points that the first extraction unit 301 extracts the three-dimensional CT image of test sample is special Levying, the tissue points feature of test sample includes the original coordinates of tissue points；First positioning unit 302 uses difference Thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of model orientation test sample；First chooses unit 303 is surveying A reference body vegetarian refreshments, wherein, test sample pair is chosen on the separating surface that sample three-dimensional CT image originally is corresponding The reference voxel that the X value of the original coordinates of the reference body vegetarian refreshments answered is corresponding with each training sample chosen in advance The X value of the original coordinates of point is identical, and the Y value of the original coordinates of the reference body vegetarian refreshments that test sample is corresponding is with in advance The Y value of the original coordinates of the reference body vegetarian refreshments that each training sample of choosing is corresponding is identical；First computing unit The tissue points reference body vegetarian refreshments corresponding relative to test sample in 304 three-dimensional CT image calculating test samples Relative coordinate, obtain the relative coordinate of each tissue points in the three-dimensional CT image of test sample；First classification Unit 305, for the tissue points to be measured in the three-dimensional CT image of test sample, is utilized respectively training in advance classification Tissue points to be measured is classified by every random forest decision tree, so that tissue points to be measured respectively falls in and trains In the leaf node of every random forest decision tree of classification；Second grouping sheet 306 tissue points to be measured fall into In the leaf node of machine forest decision tree, use k nearest neighbor model that tissue points to be measured is classified, obtain body to be measured Vegetarian refreshments is at the classification results of every random forest decision tree；First determines that unit 307 is to all of random forest certainly The classification results of plan tree is averaged, and determines the category attribution of tissue points to be measured according to meansigma methods, and classification is returned Belong to and include belonging to liver area or belonging to non-liver area；Second determines that unit 308 is according to all tissue points to be measured Category attribution, determine the tissue points to be measured belonging to liver area.Use the present invention, be utilized respectively and instruct in advance Practice classification every stochastic decision tree tissue points to be measured is classified, tissue points to be measured fall into the most gloomy In the leaf node of woods decision tree, use k nearest neighbor model that tissue points to be measured is classified, use random forest to divide The method that class and k nearest neighbor model combine, can improve the accuracy of the category attribution of tissue points to be measured, thus The liver Position location accuracy of three-dimensional CT image can be improved.

Referring to Fig. 7, Fig. 7 is another kind of liver based on three-dimensional CT image location disclosed in the embodiment of the present invention The structural representation of device.As it is shown in fig. 7, the device described in the present embodiment, except shown in Fig. 3 first Extraction unit the 301, first positioning unit 302, first choose unit the 303, first computing unit 304, first Taxon the 305, second grouping sheet 306, first determine that unit 307 and second determines outside unit 308, also Including second extraction unit the 309, second positioning unit 310, second choose unit the 311, second computing unit 312, 3rd chooses unit 313 and the 3rd taxon 314, wherein:

Second extraction unit 309, special for extracting the tissue points of the three-dimensional CT image of multiple training sample respectively Levying, the tissue points feature of training sample includes original coordinates and the tissue points gray value of tissue points.

In the embodiment of the present invention, before test sample is classified, first have to training sample is instructed Practice classification.Three-dimensional CT image is made up of multiple tissue points, and tissue points is the minimum component unit of three-dimensional CT image, Tissue points feature can include the original coordinates of tissue points, the gray value of tissue points, centered by tissue points The local histogram of cube bounding box and the contextual feature etc. of tissue points, wherein, centered by tissue points It is that 5 (length of side of corresponding cube bounding box is that cube bounding box is taken from the radius of neighbourhood of tissue points Cube bounding box 2*5+1=11)；The contextual feature of tissue points is to certain deviation amount around tissue points Tissue points carries out gray value extraction, as the contextual feature of this tissue points.At some preferred embodiment In, by centered by tissue points, the radius of neighbourhood be 9 (length of side of corresponding cube bounding box is as 2*9+1=19) The local histogram of cube bounding box as the contextual feature of this tissue points；The original coordinates bag of tissue points Including X, Y and Z, for the three-dimensional CT image of human body, the gray value span of tissue points is from-1024 to 3075.

Second positioning unit 310, for using difference model to position the three-dimensional CT image of each training sample respectively Middle thoracic cavity and the separating surface in abdominal cavity.

In the embodiment of the present invention, owing to, in three-dimensional CT image, being positioned at the gray value of the tissue points of Different Organs Distribution is different, and such as, the grey value profile scope of the tissue points being positioned at liver area is generally 75-150, The grey value profile scope of the tissue points being positioned at lung areas is less than-64, is positioned at the tissue points of other organs of abdominal part Grey value profile scope typically at-20-250, it is possible to use the difference of the gray value of Different Organs, second is fixed Bit location 310 uses difference model to position thoracic cavity and abdominal cavity in the three-dimensional CT image of each training sample respectively Separating surface.

Second chooses unit 311, for difference on the separating surface that the three-dimensional CT image of each training sample is corresponding Choose a reference body vegetarian refreshments, wherein, the X value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding Identical, the Y value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding is identical.

In the embodiment of the present invention, second chooses the unit 311 corresponding the dividing of three-dimensional CT image at each training sample The reference body vegetarian refreshments chosen respectively on interface, wherein, the reference body vegetarian refreshments that each training sample is corresponding The X value of original coordinates is identical, and the Y value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding is identical. Such as, the separating surface that first training sample is corresponding chooses a reference body vegetarian refreshments, this reference body vegetarian refreshments Original coordinates can be P₁(0,0, z₁), the X value of the original coordinates of this reference body vegetarian refreshments is 0, and Y value is 0, Z value is z₁, z₁It is axial corresponding with the separating surface in abdominal cavity in thoracic cavity in the three-dimensional CT image of first training sample The number of plies；The separating surface that second training sample is corresponding chooses a reference body vegetarian refreshments, this reference body vegetarian refreshments Original coordinates can be P₂(0,0, z₂), the X value of the original coordinates of this reference body vegetarian refreshments is 0, and Y value is 0, Z value is z₂, z₂It is axial corresponding with the separating surface in abdominal cavity in thoracic cavity in the three-dimensional CT image of second training sample The number of plies；The separating surface that n-th training sample is corresponding is chosen a reference body vegetarian refreshments, this reference body vegetarian refreshments Original coordinates can be P_n(0,0, z_n), the X value of the original coordinates of this reference body vegetarian refreshments is 0, and Y value is 0, Z value is z_n, z_nIt is axial corresponding with the separating surface in abdominal cavity in thoracic cavity in the three-dimensional CT image of second training sample The number of plies；The X value of each training sample chosen here is 0, and Y value is 0.

Second computing unit 312, the tissue points phase in the three-dimensional CT image calculating each training sample respectively For the relative coordinate of reference body vegetarian refreshments corresponding to tissue points place training sample, obtain each training sample The relative coordinate of each tissue points in three-dimensional CT image.

In the embodiment of the present invention, the second computing unit 312 calculates in the three-dimensional CT image of each training sample respectively The relative coordinate of the tissue points reference body vegetarian refreshments corresponding relative to tissue points place training sample, obtain each The relative coordinate of each tissue points in the three-dimensional CT image of training sample, such as, for first training sample, If the original coordinates of the reference body vegetarian refreshments chosen is P (0,0, z₁), then the three dimensional CT figure of the first training sample In Xiang, the concrete calculation of the relative coordinate of each tissue points can be: if the original seat of some tissue points It is designated as P₁(x₁, y₁, z₂), then its relative coordinate is P₁'(x₁, y₁, z₂-z₁)；For second training sample This, if the original coordinates of the reference body vegetarian refreshments chosen is P (0,0, z₂), then the three-dimensional of the second training sample In CT image, the concrete calculation of the relative coordinate of each tissue points can be: if some tissue points is former Beginning coordinate is P₁(x₁, y₁, z₁), then its relative coordinate is P₁'(x₁, y₁, z₁-z₂)。

3rd chooses unit 313, for randomly selecting first from the three-dimensional CT image of each training sample respectively The tissue points of the liver area of predetermined number and the tissue points of the non-liver area of the second predetermined number, as with Machine training tissue points.

In the embodiment of the present invention, owing to each training sample is carried out random forest classification, the 3rd chooses unit 313 liver areas randomly selecting the first predetermined number from the three-dimensional CT image of each training sample respectively The tissue points of the non-liver area of tissue points and the second predetermined number, as training tissue points at random.First is pre- If quantity and the second predetermined number can pre-set as required, owing to the present invention is used for positioning liver Region, so the first predetermined number is typically greater than the second predetermined number.The random number one training tissue points As bigger, for example, 400,000 random tissue points can be chosen.

3rd taxon 314, for having that puts back to randomly select the 3rd from the tissue points feature of training sample The feature construction random forest decision tree of predetermined number, for every random forest decision tree, to from random instruction Practice in tissue points and have the random training tissue points of the 4th predetermined number randomly selected put back to be trained point Class, so that the random training tissue points of the 4th predetermined number falls in the leaf node of random forest decision tree.

In the embodiment of the present invention, the tissue points feature of training sample includes tissue points coordinate, tissue points gray value, The local histogram of cube bounding box centered by tissue points and the contextual feature etc. of tissue points, train sample The number of this tissue points feature can more than 100, the X-coordinate of such as tissue points, the Y coordinate of tissue points, The Z coordinate of tissue points, the pixel value of tissue points, the local histogram of cube bounding box centered by tissue points The middle gray value number from-1000 to-950, cube bounding box centered by tissue points local histogram The gray value number from-950 to-900, cube bounding box centered by tissue points local histogram gray scale The value number from-900 to-850, etc., the 3rd predetermined number can be 10.If training tissue points at random Number is 400,000, and the 4th predetermined number can be 20,000.For example, to every random forest decision tree, Can have that puts back to randomly select 10 features as random forest decision-making from the tissue points feature of training sample The feature of tree, then randomly selects 20,000 tissue points from random tissue points and is trained classification, when the most gloomy The tissue points of the node of woods decision tree is the purest (tissue points such as falling into this node belongs to liver area), or When the tissue points number of node is less than 64, stop division, using this node as leaf node.

In the embodiment of the present invention, the second extraction unit 309 extracts the three-dimensional CT image of multiple training sample respectively Tissue points feature, the tissue points feature of training sample includes original coordinates and the tissue points gray value of tissue points； Second positioning unit 310 use difference model position respectively in the three-dimensional CT image of each training sample thoracic cavity and The separating surface in abdominal cavity；Second chooses unit 311 on the separating surface that the three-dimensional CT image of each training sample is corresponding Choose a reference body vegetarian refreshments, wherein, the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding respectively X value identical, the Y value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding is identical；Second calculates The tissue points that unit 312 calculates in the three-dimensional CT image of each training sample respectively is instructed relative to tissue points place Practice the relative coordinate of reference body vegetarian refreshments corresponding to sample, obtain in the three-dimensional CT image of each training sample each The relative coordinate of tissue points；3rd chooses unit 313 distinguishes random from the three-dimensional CT image of each training sample Choose the tissue points of the liver area of the first predetermined number and the voxel of the non-liver area of the second predetermined number Point, as training tissue points at random；3rd taxon 314 is put back to from the tissue points feature of training sample The feature construction random forest decision tree randomly selecting the 3rd predetermined number, for every random forest decision-making Tree, trains voxel at random to have the 4th predetermined number randomly selected put back to from random training tissue points Point is trained classification, so that the random training tissue points of the 4th predetermined number falls into random forest decision tree In leaf node；First extraction unit 301 extracts the tissue points feature of the three-dimensional CT image of test sample, tissue points Feature includes the original coordinates of tissue points；First positioning unit 302 uses the three of difference model assignment test sample Thoracic cavity and the separating surface in abdominal cavity in Vc T image；First chooses the unit 303 three-dimensional CT image in test sample A reference body vegetarian refreshments is chosen on corresponding separating surface, wherein, the reference body vegetarian refreshments that test sample is corresponding former The X value of the original coordinates of the reference body vegetarian refreshments that the X value of beginning coordinate is corresponding with each training sample chosen in advance Identical, the Y value of the original coordinates of the reference body vegetarian refreshments that test sample is corresponding and each training sample chosen in advance The Y value of the original coordinates of corresponding reference body vegetarian refreshments is identical；First computing unit 304 calculates the three of test sample The relative coordinate of the reference body vegetarian refreshments that tissue points in Vc T image is corresponding relative to test sample, is tested The relative coordinate of each tissue points in the three-dimensional CT image of sample；First taxon 305 is for test sample Tissue points to be measured in three-dimensional CT image, is utilized respectively every random forest decision tree pair of training in advance classification Tissue points to be measured is classified, so that tissue points to be measured respectively falls in trains every random forest of classification certainly In the leaf node of plan tree；The leaf node of the random forest decision tree that the second grouping sheet 306 falls in tissue points to be measured In, use k nearest neighbor model that tissue points to be measured is classified, obtain tissue points to be measured at every random forest certainly The classification results of plan tree；First determines that the classification results of all of random forest decision tree is averaging by unit 307 Value, and determine according to meansigma methods the category attribution of tissue points to be measured, category attribution include belonging to liver area or Belong to non-liver area；Second determines the unit 308 category attribution according to all tissue points to be measured, determines and belongs to The tissue points to be measured of liver area.Use the present invention, the liver Position location accuracy of three-dimensional CT image can be improved.

One of ordinary skill in the art will appreciate that all or part of step in the various methods of above-described embodiment The program that can be by completes to instruct relevant hardware, and this program can be stored in a computer-readable and deposit In storage media, storage medium may include that flash disk, read only memory (Read-Only Memory, ROM), Random access device (Random Access Memory, RAM), disk or CD etc..

A kind of based on three-dimensional CT image the liver the localization method above embodiment of the present invention provided and dress Putting and be described in detail, principle and the embodiment of the present invention are carried out by specific case used herein Illustrating, the explanation of above example is only intended to help to understand method and the core concept thereof of the present invention；Meanwhile, For one of ordinary skill in the art, according to the thought of the present invention, in detailed description of the invention and range of application On all will change, in sum, this specification content should not be construed as limitation of the present invention.

Claims

1. a liver localization method based on three-dimensional CT image, it is characterised in that including:

Method the most according to claim 1, it is characterised in that in the three-dimensional of described extraction test sample Before the tissue points feature of CT image, described method also includes:

Extracting the tissue points feature of the three-dimensional CT image of multiple training sample respectively, the tissue points of training sample is special Levy original coordinates and the tissue points gray value including tissue points；

Difference model is used to position thoracic cavity and the boundary in abdominal cavity in the three-dimensional CT image of each training sample respectively Face；

The separating surface that the three-dimensional CT image of each training sample is corresponding is chosen a reference body vegetarian refreshments respectively, Wherein, the X value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding is identical, each training sample pair The Y value of the original coordinates of the reference body vegetarian refreshments answered is identical；

The tissue points calculated respectively in the three-dimensional CT image of each training sample is instructed relative to described tissue points place Practice the relative coordinate of reference body vegetarian refreshments corresponding to sample, obtain in the three-dimensional CT image of each training sample each The relative coordinate of tissue points；

The liver area of the first predetermined number is randomly selected respectively from the three-dimensional CT image of each training sample The tissue points of the non-liver area of tissue points and the second predetermined number, as training tissue points at random；

The feature structure randomly selecting the 3rd predetermined number put back to is had from the tissue points feature of described training sample Build random forest decision tree, for every random forest decision tree, have from described random training tissue points The random training tissue points of the 4th predetermined number randomly selected put back to is trained classification, so that described the The random training tissue points of four predetermined numbers falls in the leaf node of random forest decision tree.

Method the most according to claim 1, it is characterised in that the tissue points feature of described test sample Also including tissue points gray value, described employing difference model positions the three-dimensional CT image mesothorax of described test sample Chamber and the separating surface in abdominal cavity, including:

Thoracic cavity and the separating surface in abdominal cavity in the three-dimensional CT image of the described test sample in employing equation below location:

L_I ^*=min{z | 0 ＜ Δ R_I(z) ＜ δ, Δ R_I(z)=R_I(z)-R_I(z-1) (1)；

R_I(L_I ^*)=α max{R_I(z), 0 ＜ α ＜ 1} (2)；

Wherein, L_I ^*For thoracic cavity in the three-dimensional CT image of described test sample and the separating surface in abdominal cavity, z is described The axial number of plies that the three-dimensional CT image axial plane of test sample is corresponding, R_I(z) be the axial number of plies be the axial plane of z In pre-set gray value in region and account for less than the tissue points number presetting gray value axial plane pre-sets district The ratio of tissue points number, R in territory_I(z-1) be the axial number of plies be (z-1) axial plane in pre-set region Interior gray value accounts for less than the tissue points number of described default gray value and pre-sets tissue points in region in axial plane The ratio of number；R_I(L_I ^*) it is separating surface to pre-set in region gray value less than the body of described default gray value Vegetarian refreshments number accounts for and pre-sets the ratio of tissue points number in region in separating surface.

Method the most according to claim 1, it is characterised in that described fall in described tissue points to be measured Random forest decision tree leaf node in, use k nearest neighbor model described tissue points to be measured is classified, To described tissue points to be measured at the classification results of every random forest decision tree, including:

In the leaf node that tissue points to be measured falls into described in every random forest decision tree, choose respectively and institute State immediate K the tissue points of relative coordinate of tissue points to be measured；

Calculate in the leaf node that test sample falls into described in every stochastic decision tree respectively, to be measured with described In immediate K the tissue points of relative coordinate of tissue points, belong to the tissue points number of liver area and the ratio of K Value, using described ratio as classification results, obtains the classification at every stochastic decision tree of the described tissue points to be measured Result.

Method the most according to claim 1, it is characterised in that described to all of random forest decision-making The classification results of tree is averaged, and determines the category attribution of described tissue points to be measured according to described meansigma methods, Including:

The classification results of all of random forest decision tree is averaged, obtains described tissue points to be measured in institute The average proportions value of some random forest decision trees；

Judge that whether described average proportions value is more than predetermined threshold value；

If so, described tissue points to be measured is belonged to liver area；

If it is not, described tissue points to be measured is belonged to non-liver area.

6. a liver positioner based on three-dimensional CT image, it is characterised in that including:

Device the most according to claim 6, it is characterised in that described device also includes:

Second extraction unit, for extracting the tissue points feature of the three-dimensional CT image of multiple training sample respectively, The tissue points feature of training sample includes original coordinates and the tissue points gray value of tissue points；

Second positioning unit, for using in the three-dimensional CT image that difference model positions each training sample respectively Thoracic cavity and the separating surface in abdominal cavity；

Second chooses unit, for selecting respectively on the separating surface that the three-dimensional CT image of each training sample is corresponding Take a reference body vegetarian refreshments, wherein, the X value phase of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding With, the Y value of the original coordinates of the reference body vegetarian refreshments that each training sample is corresponding is identical；

Second computing unit, the tissue points in the three-dimensional CT image calculating each training sample respectively is relative In the relative coordinate of reference body vegetarian refreshments corresponding to described tissue points place training sample, obtain each training sample Three-dimensional CT image in the relative coordinate of each tissue points；

3rd chooses unit, pre-for randomly selecting first from the three-dimensional CT image of each training sample respectively If the tissue points of the non-liver area of the tissue points of the liver area of quantity and the second predetermined number, as at random Training tissue points；

3rd taxon, for having that puts back to randomly select the from the tissue points feature of described training sample The feature construction random forest decision tree of three predetermined numbers, for every random forest decision tree, to from described The random training tissue points having the 4th predetermined number randomly selected put back in random training tissue points is instructed Practice classification, so that the random training tissue points of described 4th predetermined number falls into the leaf segment of random forest decision tree In point.

Device the most according to claim 6, it is characterised in that the tissue points feature of described test sample Also include that tissue points gray value, described first positioning unit use difference model to position the three of described test sample In Vc T image thoracic cavity and abdominal cavity interfacial mode particularly as follows:

L_I ^*=min{z | 0 ＜ Δ R_I(z) ＜ δ, Δ R_I(z)=R_I(z)-R_I(z-1) (1)；

R_I(L_I ^*)=α max{R_I(z), 0 ＜ α ＜ 1} (2)；

Device the most according to claim 6, it is characterised in that described second taxon includes:

Choose subelement, for the leaf node that tissue points to be measured falls into described in every random forest decision tree In, choose immediate K the tissue points with the relative coordinate of described tissue points to be measured respectively；

Computation subunit, for calculating the leaf segment that test sample falls into described in every stochastic decision tree respectively In point, with the relative coordinate of described tissue points to be measured in immediate K tissue points, belong to the body of liver area The ratio of vegetarian refreshments number and K, using described ratio as classification results, obtain described tissue points to be measured every with The classification results of machine decision tree.

Device the most according to claim 6, it is characterised in that described first determines that unit includes:

Average subelement, for averaging the classification results of all of random forest decision tree, obtains institute State the tissue points to be measured average proportions value all of random forest decision tree；

Judgment sub-unit, is used for judging that whether described average proportions value is more than predetermined threshold value；

First ownership subelement, for when described judgment sub-unit sentence read result is for being, by described body to be measured Vegetarian refreshments belongs to liver area；

Second ownership subelement, for when described judgment sub-unit sentence read result is no, by described body to be measured Vegetarian refreshments belongs to non-liver area.