Summary of the invention
It is a primary object of the present invention to overcome deficiency in the prior art, a kind of automatic detection multiple organs of abdomen are provided
Position, and be able to solve the joint dividing method of multiple organ in the medical image of speed and precision problem.It is above-mentioned to solve
Technical problem, solution of the invention is:
The multiple organ dividing method based on depth convolutional neural networks and region-competitive model is provided, for abdominal CT A
(Computed Tomography Angiography, CT angiography) three-dimensional data, i.e. computed tomography blood vessel
Liver, spleen, left kidney, right kidney in contrastographic picture carry out while dividing, described competing based on depth convolutional neural networks and region
The multiple organ dividing method for striving model includes following processes:
One, training Three dimensional convolution neural network;
Two, liver, spleen, kidney and the background in trained Three dimensional convolution neural network learning CTA volume data are 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 each pixel is belonging respectively to four kinds of probability organized in image;
Five, the multi_region model based on region-competitive is established;
Six, with convex optimization method solving model;
Seven, it is post-processed, obtains the profile of each organ;
The process one specifically include the following steps:
Step A: prepare training set: collecting size and be the abdomen Hepatic CT A volume data of 512 × 512 × n, and make these
The Standard Segmentation of the livers of data, spleen and kidney is as a result, wherein n is the number of plies of volume data;
Step B: designing the structure of convolutional neural networks, and input picture block size is 496 × 496 × 279, and output image is
The image block that four sizes are 496 × 496 × 256, respectively corresponds each pixel and belongs to background, liver, spleen and kidney
Probability value;
Step C: using the various parameters in training set training convolutional neural networks, by training set ready in step A
It is put into designed convolutional neural networks in step B to be trained, obtains the various parameters in convolutional neural networks, trained
At;
The process two specifically refers to:
If the image for needing to carry out organ segmentation is three-dimensional data I (x), image definition domain isPixel is x=
(x1, x2, x3);
By in the trained convolutional neural networks of image input process one to be tested, make each pixel quilt of image I (x)
The probability value for belonging to four kinds of tissues is assigned, it is respectively L that note, which belongs to the probability value of background, liver, spleen and kidney,0(x), L1(x),
L2(x) and L3(x), the size of x ∈ Ω, this four probability graphs are identical as original image size;
Wherein, symbolExpression is contained in, and symbol ∈ expression belongs to set;
The process three specifically refers to:
Respectively to probability graph Li(x), i=0,1,2,3 each pixel point value takes threshold value 0.5, greater than the pixel of threshold value
Belong to the histioid prime area Si, i=0,1,2,3;
The process four specifically refers to:
To the every bit pixel x in the original image I (x) of input, for the original area of background, liver, spleen and kidney
Domain Si, i=0,1,2,3, count the grey level histogram H in each regioni, i=0,1,2,3;Then according to the gray scale of pixel x
Value, by this gray value in grey level histogram HiIn shared ratio belong to the probability of i class as this pixel, be denoted as pi
(x);
The process five specifically include the following steps:
Step D: the label function for defining background, liver, spleen and kidney is respectively as follows:
For x ∈ Ω and satisfaction
ui(x) value is equal to 1 expression pixel x and belongs to the i-th class loading;
Wherein, symbol :=indicate to be defined as;Ω indicates image-region;Region Ωi, i=0,1,2,3 respectively represent background,
Liver, spleen and kidney region;Symbol ∪ indicates union of sets;∑ is summation symbol;
Step E: to original image I (x), it is as follows to calculate frontier probe function g (x):
Wherein, β is positive number (value 0.2), and symbol :=expression is defined as;SymbolIndicate gradient operator;Symbol | |
Indicate L2 norm;
Step F: the region-competitive model based on prior probability image and regional statistical information is established:
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 that step E is defined;Symbol :=indicate to be defined as;Ω indicates image-region;
∫ΩIndicate the integral in the Ω of region;Dx indicates domain integral member;∑ is summation symbol;pi(x) pixel calculated for process four
Point x belongs to the probability of class i;Li(x) the pixel x to be obtained by process two belongs to the prior probability of class i;Log is indicated with 10
Seek logarithm in bottom;Symbol | | indicate L1Norm;α1, α2For normal number, for adjusting every weight, value section [20,
50] in;Indicate gradient operator;The λ refers to regularization parameter, and for adjusting every weight, value is between [0,20];
(C in the first item of above-mentioned energy functional (1)i(x) it is the item based on region, utilizes the prime area of each tissue
Grey-level statistics pi(x) and prior probability image Li(x), it estimates each pixel x and belongs to the i-th histioid cost function Ci
(x);Section 2 is the item based on boundary, can capture boundary information well, guarantees the slickness of liver profile being partitioned into;
Area item and border item pass through a frontier probe function g (x) and weight α1, α2, λ adaptive adjusting specific gravity, so that close
The place on boundary, model depend on boundary information, the smooth region in liver, and model depends on area information;It is optimal
SolutionIt is available by the above-mentioned energy functional of minimization)
The process six includes the following steps:
Step G: convex relaxation is carried out to the model (1) in process five, is obtained:
And meet
Step H: the problem of solving above-mentioned belt restraining with Augmented Lagrange method (2) obtains optimal solution
Step I: the cut zone of each tissue is determined according to the value of each region label function, can specifically be indicated are as follows:
Ωi=x ∈ Ω | ui(x)=max (u0(x), u1(x), u2(x), u3(x)) }, respectively for i=0,1,2,3;
The process seven specifically refers to:
For the tissue binary segmentation result that process six obtains, the processing of hole operator successively is filled out with two dimension to get arriving finally
Corresponding tissue regions are realized and are divided to liver, spleen, the kidney in computed tomography angiography image.
In the present invention, the convolutional neural networks in the step B specifically:
Convolutional neural networks share 10 layers of convolutional layer: the 1st layer is convolutional layer, is inputted as the original of 496 × 496 × 279 sizes
Beginning image, export be for 96 sizes 248 × 248 × 136 sizes characteristic pattern, convolution kernel size be 7 × 7 × 9, step-length 2;
2nd convolutional layer inputs the characteristic pattern that 96 sizes are 124 × 124 × 68, exports the feature that 512 sizes are 62 × 62 × 64
Figure, convolution kernel size are 5 × 5 × 5, step-length 2;3rd layer to the 7th layer convolutional layer is distributed in operation on 4 GPU, on each GPU
The characteristic pattern that 512 sizes are 31 × 31 × 32 is all inputted, the characteristic pattern that 512 sizes are 31 × 31 × 32, convolution kernel are exported
Size is 3 × 3 × 3, step-length 1;512 characteristic patterns being distributed on 4 GPU are summed it up in the 7th layer of convolutional layer;8th convolution
Layer inputs the characteristic pattern that 64 sizes are 62 × 62 × 64, exports the characteristic pattern that 512 sizes are 62 × 62 × 64, and convolution kernel is big
Small is 3 × 3 × 3, step-length 1;The characteristic pattern that 9th convolutional layer input size is 124 × 124 × 128, exports 128 sizes
For 124 × 124 × 128 characteristic pattern;The characteristic pattern that 10th convolutional layer input is 16 248 × 248 × 256, exports 4
248 × 248 × 256 characteristic pattern, convolution kernel size are 3 × 3 × 3, step-length 1;Then a softmax function output is acted on
4 248 × 248 × 256 image blocks;Finally, the image block of output is upsampled to 496 × 496 × 256 sizes, obtain most
4 whole probability graphs;
Wherein, average pond layer is acted on after the 1st, 2 convolutional layer, local domain size is 2 × 2 × 2;In the 7th, 8,9
Double size output layer is reset in effect respectively after a convolutional layer, and 8 channels of input are become 2 × 2 × 2, i.e. double size, and 1/
8 port numbers;And the scale for exporting and inputting after 3 rearrangements is consistent.
In the present invention, in the process seven, two dimension is filled out the processing of hole operator and is specifically referred to:
To every tomographic image of volume data, the hole in tissue regions is detected, it is believed that also belonged to by the hole that the tissue surrounds
In the tissue, the value of its corresponding binary segmentation result is set as 1 to get the tissue regions for arriving final complete and accurate.
Compared with prior art, the beneficial effects of the present invention are:
The present invention can quickly detect the position of abdomen liver, spleen and kidney using convolutional neural networks automatically, and
And study obtains the prior probability image of accurate each organ, it then, can simultaneously accurately using region-competitive model
It is partitioned into the profile of liver, spleen and kidney.The present invention does not need man-machine interactively, and algorithm speed is fast, and segmentation result is accurate.
Specific embodiment
Present invention is further described in detail with specific embodiment with reference to the accompanying drawing:
The following examples can make the professional technician of this profession that the present invention be more fully understood, but not with any side
The formula limitation present invention.
As shown in Figure 1, using the multiple organ dividing method based on convolutional neural networks and region-competitive model, to computer
Liver, spleen and kidney in Tomography Angiography image carry out while dividing, specific steps are as follows:
The process one specifically include the following steps:
Step A: 140 sizes are collected and are the abdomen Hepatic CT A volume data of 512 × 512 × N, and these are provided 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 the data number of plies is reduced to 286 layers;For the data of N < 286, increase in its last layer several
Layer without liver organization, so that the data number of plies increases to 286 layers.
Step B, C: selecting size with machine frame from 512 × 512 × 286 image is the image of 496 × 496 × 279 sizes
Block inputs in network.Random initializtion network parameter, the probability graph then exported using propagated forward.According to what is obtained
Cost function between probability graph and the label of Standard Segmentation, backpropagation update network parameter.After 90 step of iteration, stop network
Training.
The process two specifically: input three dimensional CT A sweep image I size is 512 × 512 × 245, adjusts window width and window level
So that liver intensity range is between 0 to 255.Image is tested in trained convolutional neural networks, obtains four
With original image with the probability graph L of sizei, i=0,1,2,3.Probability graph given image every belongs to the probability L of class ii(x), value
Range is in [0.1,0.9].
The process three specifically: selected threshold value is 0.5, to probability graph L1Take truncation, it is believed that probability is greater than 0.5 point category
In liver;To probability graph L2Take truncation, it is believed that point of the probability greater than 0.5 belongs to spleen;To probability graph L3Take truncation, it is believed that probability
Point greater than 0.5 belongs to kidney;To probability graph L0Take truncation, it is believed that point of the probability greater than 0.5 belongs to background.In this way given four
Histioid initial segmentation result.
The process four specifically: to the every bit pixel x in the original image I (x) of input, for background, liver, spleen
Dirty and kidney prime area Si, i=0,1,2,3, count the grey level histogram H in each regioni, i=0,1,2,3.Then root
According to the gray value of pixel x, by this gray value in grey level histogram HiIn shared ratio belong to i class as this pixel
Probability, be denoted as pi(x)。
The process five specifically: image is calculated in each pixel according to the formula that five step B, C of process provides in explanation
The C of pointi(x) and the value of g (x).In this example, it takes and determines parameter alpha1=40, α2=24, λ=12.
The process six specifically: utilize convex relaxation and that primal-dual interior pointmethod, introduce variable ps, pt, p, by energy functional
Conversion are as follows:
s.t.pi(x)≤Ci(x), qi(x)≤λg(x);I=0,1,2,3
To solve above-mentioned new model, using the technology of staggeredly iteration, respectively to i=0,1,2,3, iteratively calculate following formula
Son:
1, fixed variableOptimization Solution
This step can use Chambolle
Projection algorithm solves;
2, fixed variable updates
3, fixed variableIt updates
4, fixed variableIt updates
Iterative steps k=k+1 repeats the above 1-4 iteration until stopping after convergence.
Wherein, c is the step-length in iteration, is set as 0.35 in this example.G is that this illustrates that boundary defined in step 4 is visited
It surveys function g (x).After finally solving above-mentioned model, optimum segmentation label function
5, to obtained label functionTwo-value words, obtain 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. indicates " constrained in ";Max indicates maximizing, and min expression is minimized;K and k+1
Indicate iteration kth step and+1 step of kth;The subscript k of equal symbols indicates iteration kth step;Arg max expression, which is asked, to be made
Obtain the maximum variate-value of energy function;|| ||∞Indicate Infinite Norm;Div indicates divergence operator;:=indicate " being defined as ";∫Ω
Indicate the integral in the Ω of region;Dx indicates domain integral member;Set omega0, Ω1, Ω2, Ω3Respectively correspond background, liver, spleen
Dirty and kidney region.
The process seven specifically: to the binary segmentation result that process seven obtains, fill out the processing of hole operator with two dimension.Specifically
, to every tomographic image of volume data, detect the hole in liver area, it is believed that liver is also belonged to by the hole that liver surrounds,
The value of its corresponding binary segmentation result is set as 1.Similarly, same processing is also done to spleen, kidney region.
The the 118th, 189 layer of the respectively exemplary three-dimensional volume data of Fig. 2,5.Fig. 3,6 is by Three dimensional convolution Processing with Neural Network
Initial segmentation afterwards, Fig. 4,7 are the technology of the present invention treated final segmentation result.As can be seen that Three dimensional convolution network can be with
The position of liver, kidney and spleen is accurately navigated to, but segmentation result is inaccurate;By point of region-competitive model
It cuts, available each accurate segmentation result of organ.
Finally it should be noted that the above enumerated are only specific embodiments of the present invention.It is clear that the invention is not restricted to
Above embodiments can also have many variations.Those skilled in the art can directly lead from present disclosure
Out or all deformations for associating, it is considered as protection scope of the present invention.