Summary of the invention
Present invention is primarily targeted at and overcome deficiency of the prior art, it is provided that a kind of detection multiple organ of abdominal part automatically
Position, and can solve the problem that the associating dividing method of multiple organ in the medical image of speed and precision problem.Above-mentioned for solving
Technical problem, the solution of the present invention is:
There is provided multiple organ dividing method based on degree of depth convolutional neural networks and region-competitive model, for abdominal CT A
(Computed Tomography Angiography, CT angiography) three-dimensional data, i.e. computed tomography blood vessel
Liver in contrastographic picture, spleen, left kidney, right kidney are split simultaneously, described competing based on degree of depth convolutional neural networks and region
Strive the multiple organ dividing method of model and include following process:
One, training Three dimensional convolution neutral net;
Two, liver, spleen, kidney and the background in the Three dimensional convolution neural network learning CTA volume data trained is utilized
Prior probability image;
Three, the initial segmentation region of each tissue is determined by the prior probability image of various organization;
Four, determine that in image, each pixel is belonging respectively to the probability of four kinds of tissues;
Five, multi_region model based on region-competitive is set up;
Six, with convex Optimization Method model;
Seven, carry out post processing, obtain the profile of each organ;
Described process one specifically includes following step:
Step A: prepare training set: collect the abdominal part Hepatic CT A volume data that size is 512 × 512 × n, and make these
The Standard Segmentation result of the liver of data, spleen and kidney, wherein n is the number of plies of volume data;
Step B: the structure of design convolutional neural networks, input picture block size is 496 × 496 × 279, and output image is
Four sizes are the image block of 496 × 496 × 256, and the most corresponding each pixel belongs to background, liver, spleen and kidney
Probit;
Step C: utilize the various parameters in training set training convolutional neural networks, by training set ready in step A
Put into the convolutional neural networks designed in step B to be trained, obtain the various parameters in convolutional neural networks, trained
Become;
Described process two specifically refers to:
If needing the image carrying out organ segmentation is three-dimensional data I (x), image definition territory isPixel is x=
(x1, x2, x3);
In the convolutional neural networks train image input process one to be tested, make each pixel quilt of image I (x)
Giving the probit belonging to four kinds of tissues, note belongs to the probit of background, liver, spleen and kidney and is respectively L0(x), L1(x),
L2(x), and L3X (), x ∈ Ω, the size of these four probability graphs is identical with original image size;
Wherein, symbolExpression is contained in, and symbol ∈ represents and belongs to set;
Described process three specifically refers to:
Respectively to probability graph LiX (), i=0, each pixel point value of 1,2,3 takes threshold value 0.5, more than the pixel of threshold value
Belong to this histioid prime area Si, i=0,1,2,3;
Described process four specifically refers to:
To every bit pixel x in original image I (x) of input, for the original area of background, liver, spleen and kidney
Territory Si, i=0,1,2,3, the grey level histogram H in statistics regionali, i=0,1,2,3;Then according to the gray scale of pixel x
Value, by this gray value at grey level histogram HiRatio shared by belongs to the probability of i class as this pixel, is designated as pi
(x);
Described process five specifically includes following step:
Step D: the label function of definition background, liver, spleen and kidney is respectively as follows:
For x ∈ Ω and meet
uiX the value of () belongs to the i-th class loading equal to 1 expression pixel x;
Wherein, symbol :=expression is defined as;Ω represents image-region;Region Ωi, i=0,1,2,3 represent respectively background,
Liver, spleen and kidney region;Symbol ∪ represents union of sets;∑ is summation symbol;
Step E: to original image I (x), calculates frontier probe function g (x) as follows:
Wherein, β is positive number (value is 0.2), symbol :=expression is defined as;SymbolRepresent gradient operator;Symbol | |
Represent L2 norm;
Step F: set up region-competitive model based on prior probability image and regional statistical information:
And calculate Ci(x) :=[α1(-log pi(x))+α2(-log Li(x))] g (x), i=0,1,2,3;
Wherein, g (x) is the frontier probe function of step E definition;Symbol :=expression is defined as;Ω represents image-region;
∫ΩRepresent the integration in the Ω of region;Dx represents domain integral unit;∑ is summation symbol;piX () is the pixel that process four calculates
Point x belongs to the probability of class i;LiX () belongs to the prior probability of class i for the pixel x obtained by process two;Log represents with 10
Logarithm is sought at the end;Symbol | | represent L1Norm;α1, α2For normal number, for regulating every weight, value all interval [20,
50] in;Represent gradient operator;Described λ refers to regularization parameter, and for regulating every weight, value is between [0,20];
(C in the Section 1 of above-mentioned energy functional (1)iX () is item based on region, utilize the prime area of each tissue
Grey-level statistics pi(x) and prior probability image LiX (), estimates each pixel x and belongs to the i-th histioid cost function Ci
(x);Section 2 is item based on border, it is possible to catch boundary information well, it is ensured that the slickness of the liver profile being partitioned into;
Area item and border item are by frontier probe function g (x) and weight α1, α2, λ adaptive regulation proportion so that close
The place on border, model depends on boundary information, and region smooth in liver, model depends on area information;Optimum
SolveCan be obtained by the above-mentioned energy functional of minimization)
Described process six comprises the steps:
Step G: the model (1) in process five is carried out convex lax, obtain:
And meet
Step H: solve the problem (2) of above-mentioned belt restraining with Augmented Lagrange method, obtain optimal solution
Step I: determine the cut zone of each tissue according to the value of regional label function, specifically can be expressed as:
Ωi=x ∈ Ω | ui(x)=max (u0(x), u1(x), u2(x), u3(x)) }, respectively for i=0,1,2,3;
Described process seven specifically refers to:
The tissue binary segmentation result obtained for process six, fills out hole operator by two dimension successively and processes, i.e. obtain final
Corresponding tissue regions, it is achieved to the liver in computed tomography angiography image, spleen, kidney segmentation.
In the present invention, the convolutional neural networks in described step B particularly as follows:
Convolutional neural networks has 10 layers of convolutional layer: the 1st layer is convolutional layer, and input is the former of 496 × 496 × 279 sizes
Beginning image, is output as the characteristic pattern that 96 sizes are 248 × 248 × 136 sizes, and convolution kernel size is 7 × 7 × 9, and step-length is 2;
2nd convolutional layer 96 sizes of input are the characteristic pattern of 124 × 124 × 68, and exporting 512 sizes is the feature of 62 × 62 × 64
Figure, convolution kernel size is 5 × 5 × 5, and step-length is 2;3rd layer to the 7th layer convolutional layer is distributed in computing on 4 GPU, on each GPU
All inputting 512 sizes is the characteristic pattern of 31 × 31 × 32, and exporting 512 sizes is the characteristic pattern of 31 × 31 × 32, convolution kernel
Size is 3 × 3 × 3, and step-length is 1;The 7th layer of convolutional layer by 4 GPU distribution 512 characteristic patterns add and;8th convolution
Layer 64 size of input are the characteristic pattern of 62 × 62 × 64, and exporting 512 sizes is the characteristic pattern of 62 × 62 × 64, and convolution kernel is big
Little is 3 × 3 × 3, and step-length is 1;9th convolutional layer input size is the characteristic pattern of 124 × 124 × 128, exports 128 sizes
It it is the characteristic pattern of 124 × 124 × 128;The input of 10th convolutional layer is the characteristic pattern of 16 248 × 248 × 256, exports 4
The characteristic pattern of 248 × 248 × 256, convolution kernel size is 3 × 3 × 3, and step-length is 1;Then one softmax function output of effect
The image block of 4 248 × 248 × 256;Finally, the image block of output is upsampled to 496 × 496 × 256 sizes, obtains
4 whole probability graphs;
Wherein, acting on average pond layer after the 1st, 2 convolutional layers, local domain size is 2 × 2 × 2;The 7th, 8,9
After individual convolutional layer, double size output layer is reset in effect respectively, and 8 passages of input are become 2 × 2 × 2, i.e. double size, and 1/
8 port numbers;And export consistent with the yardstick holding inputted after resetting for 3 times.
In the present invention, in described process seven, two dimension is filled out hole operator process and is specifically referred to:
Every tomographic image to volume data, detection hole in tissue regions, it is believed that the hole surrounded by this tissue also belongs to
In this tissue, the value of the binary segmentation result of its correspondence is set to 1, i.e. obtains this tissue regions of final complete and accurate.
Compared with prior art, the invention has the beneficial effects as follows:
The present invention utilizes convolutional neural networks can the most quickly detect the position of abdominal part liver, spleen and kidney, and
And study obtains the prior probability image of accurately each organ, then, utilize region-competitive model, it is possible to the most accurately
It is partitioned into the profile of liver, spleen and kidney.The present invention need not man-machine interactively, and algorithm speed is fast, and segmentation result is accurate.
Detailed description of the invention
With detailed description of the invention, the present invention is described in further detail below in conjunction with the accompanying drawings:
The following examples can make the professional and technical personnel of this specialty that the present invention is more fully understood, but not with any side
Formula limits the present invention.
As it is shown in figure 1, use multiple organ dividing method based on convolutional neural networks and region-competitive model, to computer
Liver, spleen and kidney in Tomography Angiography image are split simultaneously, concretely comprise the following steps:
Described process one specifically includes following step:
Step A: collect the abdominal part Hepatic CT A volume data that 140 sizes are 512 × 512 × N, and provided these by doctor
The liver segmentation standard results of data, wherein N is the number of plies of volume data.For the data of N > 286, delete in data without liver
The number of plies of dirty tissue so that this data number of plies is reduced to 286 layers;For the data of N < 286, increase some at its last layer
Layer without liver organization so that this data number of plies increases to 286 layers.
Step B, C: the image selecting size to be 496 × 496 × 279 sizes with machine frame from the image of 512 × 512 × 286
Block, in input network.Random initializtion network parameter, then utilizes propagated forward to obtain the probability graph of output.According to obtain
Cost function between probability graph and the label of Standard Segmentation, back propagation updates network parameter.After iteration 90 step, stop network
Training.
Described process two, particularly as follows: input three dimensional CT A sweep image I size is 512 × 512 × 245, adjusts window width and window level
Make liver intensity scope between 0 to 255.Image is tested in the convolutional neural networks trained, obtains four
With original image with the probability graph L of sizei, i=0,1,2,3.Probability graph gives image every and belongs to the probability L of class ii(x), its value
Scope is in [0.1,0.9].
Described process three is particularly as follows: selected threshold value is 0.5, to probability graph L1Take and block, it is believed that the probability point more than 0.5 belongs to
In liver;To probability graph L2Take and block, it is believed that the probability point more than 0.5 belongs to spleen;To probability graph L3Take and block, it is believed that probability
Point more than 0.5 belongs to kidney;To probability graph L0Take and block, it is believed that the probability point more than 0.5 belongs to background.So given four
Histioid initial segmentation result.
Described process four particularly as follows: to input original image I (x) in every bit pixel x, for background, liver, spleen
The dirty prime area S with kidneyi, i=0,1,2,3, the grey level histogram H in statistics regionali, i=0,1,2,3.Then root
According to the gray value of pixel x, by this gray value at grey level histogram HiRatio shared by belongs to i class as this pixel
Probability, be designated as pi(x)。
Described process five is particularly as follows: the formula be given according to process five step B, C in explanation calculates image in each pixel
The C of pointiThe value of (x) and g (x).In this example, take and determine parameter alpha1=40, α2=24, λ=12.
Described process six, particularly as follows: utilize convex lax and that primal-dual interior pointmethod, introduces variable ps, pt, p, by energy functional
It is converted into:
s.t.pi(x)≤Ci(x), qi(x)≤λg(x);I=0,1,2,3
For solving above-mentioned new model, utilize the technology of staggered iteration, respectively to i=0,1,2,3, calculate iteratively with following formula
Son:
1, fixed variableOptimization Solution
This step can be thrown with Chambolle
Shadow Algorithm for Solving;
2, fixed variableUpdate
3, fixed variableUpdate
4, fixed variableUpdate
Iterative steps k=k+1, repeats above 1-4 iteration until stopping after Shou Lian.
Wherein, c is the step-length in iteration, is set to 0.35 in this example.G is that the border defined in this explanation step 4 is visited
Survey function g (x).After finally solving above-mentioned model, optimum segmentation label function
5, to the label function obtainedTwo-value is talked about, and obtains final segmentation result, specifically,
Ωi=x ∈ Ω | ui(x)=max (u0(x), u1(x), u2(x), u3(x)) }, respectively for i=0,1,2,3.
In above-mentioned symbol, s.t. represents " constrained in ";Max represents that maximizing, min represent and minimizes;K and k+1
Represent iteration kth step and kth+1 step;Represent that iteration kth walks Deng the subscript k of symbol;Arg max represents to ask to be made
Obtain the variate-value that energy function is maximum;|| ||∞Represent Infinite Norm;Div represents divergence operator;:=represent " being defined as ";∫Ω
Represent the integration in the Ω of region;Dx represents domain integral unit;Set omega0, Ω1, Ω2, Ω3The most corresponding background, liver, spleen
Dirty and kidney region.
Described process seven, particularly as follows: the binary segmentation result that obtains process seven, is filled out hole operator by two dimension and is processed.Specifically
, every tomographic image to volume data, detection hole in liver area, it is believed that the hole surrounded by liver falls within liver,
The value of the binary segmentation result of its correspondence is set to 1.In like manner, same process is also done in spleen, kidney region.
Fig. 2,5 be respectively the 118th of exemplary three-dimensional volume data the, 189 layers.Fig. 3,6 it is by Three dimensional convolution Processing with Neural Network
After initial segmentation, Fig. 4,7 be the technology of the present invention process after final segmentation result.It can be seen that Three dimensional convolution network is permissible
Accurately navigate to the position of liver, kidney and spleen, but segmentation result inaccuracy;Dividing through region-competitive model
Cut, the accurate segmentation result of each organ can be obtained.
It is only the specific embodiment of the present invention finally it should be noted that listed above.It is clear that the invention is not restricted to
Above example, it is also possible to have many variations.Those of ordinary skill in the art directly can lead from present disclosure
The all deformation gone out or associate, are all considered as protection scope of the present invention.