CN105184799A - Modified non-supervision brain tumour MRI (Magnetic Resonance Imaging) image segmentation method - Google Patents

Abstract

The invention relates to the technical field of image processing, and specifically relates to a modified non-supervision brain tumour MRI (Magnetic Resonance Imaging) image segmentation method. For the modified non-supervision brain tumour MRI image segmentation method, a non-linear deflection resonance image optimization model is added to perform optimization of segmented images through a non-linear deflection resonance image optimization model. Therefore, the segmented images obtained through the method are high in accuracy so that the identification of brain tumour is higher in accuracy.

Description

A kind of nothing supervision brain tumor MRI image partition method of improvement

Technical field

The present invention relates to technical field of image processing, particularly relate to a kind of nothing supervision brain tumor MRI image partition method of improvement.

Background technology

Medical image segmentation extracts the indispensable means of quantitative information of particular tissues in imaged image, is also 3-dimensional reconstruction and visual prerequisite simultaneously.Image after segmentation is widely used in various occasion, as location and the diagnosis of pathological tissues, and the study of anatomical structure, computer guidance operation and three-dimensional visualization etc.

MRI (Magnetic resonance imaging) image is one of important component part of medical imaging, but because MRI image exists noise to a certain degree, so we need to carry out pre-service to original MRI image, to obtain better picture element, improve the precision of process, obtain desirable segmentation effect.

Prior art mainly adopts based on support vector machine (SupportVectorMachines, SVMs) edge local method solves problem of false edges during rim detection, utilize the region-expanding technique of adaptive automatic selected seed point to Image Segmentation Using simultaneously, improve the accuracy of Iamge Segmentation.But adopt the accuracy of the segmentation image obtained in this way or not high enough.

Summary of the invention

Technical matters to be solved by this invention is: the nothing supervision brain tumor MRI image partition method providing a kind of improvement, adopts the segmentation imaging accuracy obtained in this way higher, and then makes the recognition accuracy of brain tumor higher.

The technical solution adopted in the present invention is: a kind of nothing supervision brain tumor MRI image partition method of improvement, and it comprises the following steps:

(1) the standard smnr data intersection of known brain MRI image, is set up;

(2), obtain brain MRI image to be split, and be converted into gray level image;

(3) the edge local method of SVM, is utilized to carry out rim detection to the gray level image that step (2) obtains;

(4), to each bar edge that step (3) detects carry out feature extraction, then build SVM classifier and edge local is carried out to the characteristic parameter extracted, sort out brain different tissues edge;

(5), from the 3*3 neighborhood at the brain different tissues edge that step (4) obtains, the point of the multiple pixel value of random selecting within the scope of 230-250 as growing point, and sets a predetermined threshold value T;

(6) growing point, adopting step (5) to obtain is core, judge whether the non-growing point in the neighborhood of its 3*3 meets principle of similarity, if meet, this non-growing point is joined in growing point intersection, if do not meet, cast out this non-growing point, until after all non-growing points all judge, growth district is split and obtains brain MRI and split image;

(7) the segmentation image, step (6) obtained is with the pattern of 3*3 neighborhood, piecemeal is input to non-linear deflection resonance image Optimized model successively, obtain output signal-to-noise ratio aggregated data, wherein said non-linear deflection resonance image Optimized model is as follows:

$\frac{\∂s}{\∂t}=A\×cos(2{\mathrm{\πf}}_{0}t+\mathrm{\ψ})+img\left(t\right)+ns-{\mathrm{ms}}^3+\sqrt{2\mathrm{\α}}\mathrm{\ξ}\left(t\right);$

In formula, A is signal amplitude; f ₀for periodic signal frequency; ψ is periodic signal initial phase; S is the Brownian Particles coordinates of motion; T is Brownian Particles run duration; A × cos (2 π f ₀t+ ψ) be deflection resonance cycle input signal function; Img (t) is 3*3 neighborhood input picture; M and n is bistable state potential barrier real parameter; α is noise intensity; ξ (t) for average be the white Gaussian noise of 0, its autocorrelation function E [ξ (t) ξ (0)]=2 α δ (t), its intensity is α, δ (t) is unit impulse function;

(8) compared with the standard smnr data intersection that the output signal-to-noise ratio aggregated data, by step (7) obtained and step (1) obtain, if similarity >=90%, optimize successfully in judgement; If similarity < 90%, then judge that optimization is unsuccessful, return step (2) and proceed segmentation and optimize.

The standard smnr data intersection setting up known brain MRI image in step (1) comprises following concrete steps:

A, split multiple brain MRI image by classic method;

B, split by doctor's visual inspection determining step A after image whether accurate, if accurately, be included into correct images and concentrate;

In C, calculation procedure B, the smnr data intersection of all segmentation images that correct images is concentrated, then averages and obtains standard smnr data intersection.

Adopt above method compared with prior art, the present invention has the following advantages: problem of false edges when adopting the edge local method based on support vector machine to solve rim detection, utilize the region-expanding technique of adaptive automatic selected seed point to Image Segmentation Using simultaneously, the last region adopting non-linear deflection resonance image Optimized model to prune overexpansion again, expansion area is owed in supplement, and split image with the standard brain MRI that prior art obtains and contrast, brain MRI after being finally optimized splits image, obtain segmentation imaging accuracy higher, and adopt the recognition accuracy of this segmentation image to brain tumor higher.

Accompanying drawing explanation

Fig. 1 is original-gray image.

Fig. 2 is the segmentation image adopting the inventive method to obtain.

Fig. 3 is the graph of a relation without noise intensity and signal to noise ratio (S/N ratio) in supervision brain tumor MRI image partition method of a kind of improvement of the present invention.

Embodiment

Below in conjunction with accompanying drawing and embodiment, the present invention is described further, but the present invention is not limited only to following embodiment.

A nothing supervision brain tumor MRI image partition method for improvement, it comprises the following steps:

(1) the standard smnr data intersection of known brain MRI image, is set up;

A, split multiple brain MRI image by classic method; Described classic method is mark watershed method, also can be the image partition method of other routines;

B, split by doctor's visual inspection determining step A after image whether accurate, if accurately, be included into correct images and concentrate;

Whether the image after described determining step A segmentation accurately refers to that doctor judges whether each tissue of brain to separate by experience, if complete parttion has come, then judge it is correct, if be destroyed during some tissue segmentation or two tissue segmentation do not come, then judge it is incorrect;

In C, calculation procedure B, the smnr data intersection of all segmentation images that correct images is concentrated, then averages and obtains standard smnr data intersection.

Signal to noise ratio (S/N ratio) is compared by the gray-scale value of pixel useful in computed segmentation image and the gray-scale value of useless pixel to obtain.

(2), obtain brain MRI image to be split, and be converted into gray level image;

(3) the edge local method of SVM, is utilized to carry out rim detection to the gray level image that step (2) obtains;

Tumor tissues has certain profile, and the gray-scale value of this profile is consistent substantially, and therefore we are just using the edge of the profile of this gray-scale value as gray level image, are equal to area-of-interest initial profile.

Calculate gradient magnitude and the histogram of gradients of every width image sequence, determine the optimal segmenting threshold t=45 of high gradient regions and low gradient region according to the iterative step of Gonzalez and Woods proposition, then calculate average and the variance μ of high and low gradient region ₁(t)=2.7, μ ₀(t)=0.6, σ ₁ ²=8.4 and σ ₀ ²=14.4, thus calculate high threshold τ _h=194 and Low threshold τ _l=166.Canny operator is finally used to detect every width image border.

(4), to each bar edge that step (3) detects carry out feature extraction, then build SVM classifier and edge local is carried out to the characteristic parameter extracted, sort out brain different tissues edge;

Brain tissue mainly comprises brain, brain stem, cerebellum, pons and oblongata;

First separately get the brain MRI sequence image of certain patient in the present embodiment, the edge carried out after rim detection marks, and be 1 by oblongata edge labelling, non-oblongata edge labelling is 0.This patient obtains altogether 512 sample edge as training set.Test set selects the MRI image of other 20 these patients, often opens 5 samples, and totally 100 samples are as test set.Then the edge feature of training sample and test sample book is normalized, the Selection of kernel function Radial basis kernel function K (x of SVM _i, x _j)=exp (-γ || x _i-x _j|| ²), and use particle cluster algorithm to be optimized the punishment parameter C of SVM and nuclear parameter γ, then build two classification SVM classifier with training sample.

Finally classify with test set sample, draw oblongata edge and non-oblongata edge, and only select display oblongata edge.

(5), from the 3*3 neighborhood at the brain different tissues edge that step (4) obtains, the point of the multiple pixel value of random selecting within the scope of 230-250 as growing point, and sets a predetermined threshold value T;

The span of described predetermined threshold value T is 200-280;

(6) growing point, adopting step (5) to obtain is core, judge whether the non-growing point in the neighborhood of its 3*3 meets principle of similarity, if meet, this non-growing point is joined in growing point intersection, if do not meet, cast out this non-growing point, until after all non-growing points all judge, growth district is split and obtains brain MRI and split image;

The judgment formula of principle of similarity is max|f _xy-m| (x, y ∈ R) < T, wherein f _xydenotation coordination position is the pixel value of the non-seed point of (x, y), and x is x coordinate figure a little, and y is y coordinate figure a little, and m represents the pixel average of all Seed Points, and R represents the coordinate set of each point in growth district Seed Points 3*3 neighborhood;

(7) the segmentation image, step (6) obtained is with the pattern of 3*3 neighborhood, piecemeal is input to non-linear deflection resonance image Optimized model successively, obtain output signal-to-noise ratio aggregated data, wherein said non-linear deflection resonance image Optimized model is as follows:

$\frac{\&part;s}{\&part;t}=A\&times;cos(2{\mathrm{\&pi;f}}_{0}t+\mathrm{\&psi;})+img\left(t\right)+ns-{\mathrm{ms}}^3+\sqrt{2\mathrm{\&alpha;}}\mathrm{\&xi;}\left(t\right);$

In formula, A is signal amplitude; f ₀for periodic signal frequency; ψ is periodic signal initial phase; S is the Brownian Particles coordinates of motion; T is Brownian Particles run duration; A × cos (2 π f ₀t+ ψ) be deflection resonance cycle input signal function; Img (t) is 3*3 neighborhood input picture; M and n is bistable state potential barrier real parameter; α is noise intensity; ξ (t) for average be the white Gaussian noise of 0, its autocorrelation function E [ξ (t) ξ (0)]=2 α δ (t), its intensity is α, δ (t) is unit impulse function;

Non-linear deflection resonance is noise-induced signal resonance effect, in signal analysis, has widespread use at present, generally describes this model with the movement locus of the power-actuated Brownian movement particle of one-period in bistable potential well:

$\frac{\&part;s}{\&part;t}=A\&times;cos(2{\mathrm{\&pi;f}}_{0}t+\mathrm{\&psi;})-\frac{\&part;V\left(s\right)}{\&part;s}+\sqrt{2\mathrm{\&alpha;}}\mathrm{\&xi;}\left(t\right)---\left(1\right)$

In formula, A is signal amplitude; f ₀for periodic signal frequency; ψ is periodic signal initial phase; S is the Brownian Particles coordinates of motion; T is Brownian Particles run duration; A × cos (2 π f ₀t+ ψ) be deflection resonance cycle input signal function; V (s) is Symmetric Double Well-potential model, and shown in (2), m, n are real parameter; ξ (t) for average be the white Gaussian noise of 0, its autocorrelation function E [ξ (t) ξ (0)]=2 α δ (t), its intensity is α, δ (t) is unit impulse function.

$V\left(s\right)=\frac{1}{4}{\mathrm{ms}}^4-\frac{1}{2}{\mathrm{ns}}^2+m\&times;n---\left(2\right)$

Therefore formula (1) can be rewritten as:

$\frac{\&part;s}{\&part;t}=A\&times;cos(2{\mathrm{\&pi;f}}_{0}t+\mathrm{\&psi;})+ns-{\mathrm{ms}}^3+\sqrt{2\mathrm{\&alpha;}}\mathrm{\&xi;}\left(t\right)---\left(3\right)$

Output signal-to-noise ratio parameter is usually used to characterize the resonance of nonlinear bistability deflection, and it is defined as:

Wherein s is the Brownian Particles coordinates of motion; A is signal amplitude; α is noise intensity;

View data img (t) and A × cos (2 π f ₀t+ ψ) coupling after input deflection resonating member, formula (3) is rewritten as:

$\frac{\&part;s}{\&part;t}=A\&times;cos(2{\mathrm{\&pi;f}}_{0}t+\mathrm{\&psi;})+img\left(t\right)+ns-{\mathrm{ms}}^3+\sqrt{2\mathrm{\&alpha;}}\mathrm{\&xi;}\left(t\right)---\left(5\right)$

In formula, deflection resonator system parameter comprises noise intensity α, fixed cycle signal intensity A, fixed cycle signal frequency f ₀, the parameters such as bistable state potential barrier real parameter m and n, fixed cycle signal initial phase ψ.In actual analysis, keep fixed cycle signal parameter A=6.5, f ₀=1kHz, ψ=0.5 are constant, and make noise intensity α span be [0,40], this seasonal bistable state barrier parameters n=1, and make m carry out the change that stepping is 1 within [1,100], supervisory system output signal-to-noise ratio simultaneously, when output signal-to-noise ratio curve produces characteristic peak and peak value is maximal value, namely can determine m=8.9, now parameters is optimization selection.

This object calculated is marginal information after optimized image is split, be conducive to the accuracy improving segmentation, as shown in Figure 3: under the excitation of certain noise intensity, system output signal-to-noise ratio reaches optimal condition, namely there is maximum value in output signal-to-noise ratio, in maximum value situation, be that the optimization meeting image is split, then can obtain optimized Iamge Segmentation according to output signal-to-noise ratio now.

(8) compared with the standard smnr data intersection that the output signal-to-noise ratio aggregated data, by step (7) obtained and step (1) obtain, if similarity >=90%, optimize successfully in judgement; If similarity < 90%, then judge that optimization is unsuccessful, return step (2) and proceed segmentation and optimize.

This compares is first inside output signal-to-noise ratio intersection, take out a snr value, then in standard signal to noise ratio (S/N ratio) intersection, a snr value is taken out with the former opposite position, both are subtracted each other and takes absolute value, then by the absolute value of difference divided by the snr value taken out in standard signal to noise ratio (S/N ratio) intersection before, be multiplied by 100% again, finally all values obtained are averaged, judge whether to be less than 10%.

Claims (2)

1. the nothing supervision brain tumor MRI image partition method improved, it is characterized in that, it comprises the following steps:

(1) the standard smnr data intersection of known brain MRI image, is set up;

(2), obtain brain MRI image to be split, and be converted into gray level image;

(3) the edge local method of SVM, is utilized to carry out rim detection to the gray level image that step (2) obtains;

(4), to each bar edge that step (3) detects carry out feature extraction, then build SVM classifier and edge local is carried out to the characteristic parameter extracted, sort out brain different tissues edge;

(5), from the 3*3 neighborhood at the brain different tissues edge that step (4) obtains, the point of the multiple pixel value of random selecting within the scope of 230-250 as growing point, and sets a predetermined threshold value T;

(6) growing point, adopting step (5) to obtain is core, judge whether the non-growing point in the neighborhood of its 3*3 meets principle of similarity, if meet, this non-growing point is joined in growing point intersection, if do not meet, cast out this non-growing point, until after all non-growing points all judge, growth district is split and obtains brain MRI and split image;

(7) the segmentation image, step (6) obtained is with the pattern of 3*3 neighborhood, piecemeal is input to non-linear deflection resonance image Optimized model successively, obtain output signal-to-noise ratio aggregated data, wherein said non-linear deflection resonance image Optimized model is as follows:

$\frac{\&part;s}{\&part;t}=A\&times;cos(2{\mathrm{\&pi;f}}_{0}t+\mathrm{\&psi;})+img\left(t\right)+ns-{\mathrm{ms}}^3+\sqrt{2\mathrm{\&alpha;}}\mathrm{\&xi;}\left(t\right);$

In formula, A is signal amplitude; f ₀for periodic signal frequency; ψ is periodic signal initial phase; S is the Brownian Particles coordinates of motion; T is Brownian Particles run duration; A × cos (2 π f ₀t+ ψ) be deflection resonance cycle input signal function; Img (t) is 3*3 neighborhood input picture; M and n is bistable state potential barrier real parameter; α is noise intensity; ξ (t) for average be the white Gaussian noise of 0, its autocorrelation function E [ξ (t) ξ (0)]=2 α δ (t), its intensity is α, δ (t) is unit impulse function;

(8) compared with the standard smnr data intersection that the output signal-to-noise ratio aggregated data, by step (7) obtained and step (1) obtain, if similarity >=90%, optimize successfully in judgement; If similarity < 90%, then judge that optimization is unsuccessful, return step (2) and proceed segmentation and optimize.

2. the nothing supervision brain tumor MRI image partition method of a kind of improvement according to claim 1, is characterized in that: the standard smnr data intersection setting up known brain MRI image in step (1) comprises following concrete steps:

A, split multiple brain MRI image by classic method;

B, split by doctor's visual inspection determining step A after image whether accurate, if accurately, be included into correct images and concentrate;

In C, calculation procedure B, the smnr data intersection of all segmentation images that correct images is concentrated, then averages and obtains standard smnr data intersection.