CN103606138B - A kind of fusion method of the medical image based on texture region division - Google Patents

Abstract

The invention discloses a kind of fusion method of the medical image based on texture region division, the present invention is directed to the deficiency that existing integration technology exists, merge for CT and MR multi modal medical image, that cluster mode utilizes K-means with multicharacteristic information? Clustering splits the character pair point of extraction source image, by sorting out the unique point set merging and set up multi modal medical image, according to unique point distribution, image is divided into texture region and non-grain region, texture region coefficient of correspondence input PCNN obtains Fire mapping image, fusion coefficients is selected according to ignition times, the coefficient in non-grain region is merged by binary channels PCNN.Experimental result shows, the method can accurate partitioned image texture region, and then utilizes PCNN and binary channels PCNN to select respective advantage at image zones of different coefficient, fused images clean mark, and quality is improved.

Description

A kind of fusion method of the medical image based on texture region division

Technical field

The present invention relates to a kind of method of technical field of image processing, specifically a kind of Method of Medical Image Fusion divided based on multi-cluster center and texture region.

Background technology

Medical image fusion is an important branch in image co-registration field, is also difficult point and the focus of research at present.Medical image fusion carries out informix utilization by deriving from the dissimilar view data to same biorgan images such as () CT, MR and PET that multiclass Medical Devices obtain, image ratio single image after fusion contains abundanter useful information, for follow-up doctor's diagnosis and treatment are provided convenience, there is very strong using value.

According to the unusual of image procossing mode, image co-registration can be divided into Pixel-level, feature level and decision level 3 levels.Pixel-level image fusion is most widely used, the most direct to the process of pixel, and the relevance between pixel is considered less in fusion decision-making; Feature level image co-registration utilizes mathematical statistics amount characteristic information extraction from image, and carried out the process of comprehensive treatment and analysis; Decision level image co-registration be obtain image characteristic information basis on carry out abstract further, for next step decision-making provides foundation.

Medical image contrast is lower, and noise is serious, and image quality is poor, and these features have impact on the application of Pixel-level fusion method in Medical image fusion, reduce fused image quality.

Summary of the invention

The technical issues that need to address of the present invention are just the defect overcoming prior art; a kind of fusion method of the medical image based on texture region division is provided; it merges relative to Pixel-level; feature-based fusion based on area merges considers the correlativity between neighbor; highlight provincial characteristics; reduce the interference of noise to important informations such as textures, the texture information of available protecting image, more useful informations can be extracted.

For solving the problem, the present invention adopts following technical scheme:

The invention provides a kind of fusion method of the medical image based on texture region division, comprise the following steps:

1. calculate the average of source images respectively, standard deviation, entropy, the information such as greatest gradient value, as initial cluster centre, the foundation that two width image clustering centers are produced is consistent with the objective evaluation index of picture quality;

2. by K-meansClustering algorithm, two width source images are carried out cluster with cluster centre respectively, obtain feature space vector;

3. extract the feature distributed areas of every width image according to feature space vector, compare the corresponding region of two width images, setting threshold value T, extract coefficient in two width images and be all greater than the positional information of threshold value and extract respective regions according to this segmentation of class, be defined as texture region;

4. texture region pixel value input PCNN neural network is obtained respective Fire mapping image, get pixel that in two width images, ignition times is larger as the fusion coefficients of fused images, non-grain area pixel value is merged by binary channels PCNN;

5. obtain fused images by fusion coefficients.

The present invention is directed to the deficiency that existing integration technology exists, merge for CT and MR multi modal medical image, be that cluster mode utilizes K-meansClustering to split the character pair point of extraction source image with multicharacteristic information, by sorting out the unique point set merging and set up multi modal medical image, according to unique point distribution, image is divided into texture region and non-grain region, texture region coefficient of correspondence input PCNN obtains Fire mapping image, select fusion coefficients according to ignition times, the coefficient in non-grain region is merged by binary channels PCNN.Experimental result shows, the method can accurate partitioned image texture region, and then utilizes PCNN and binary channels PCNN to select respective advantage at image zones of different coefficient, fused images clean mark, and quality is improved.

Accompanying drawing explanation

Fig. 1 is embodiment of the present invention effect schematic diagram.

In figure: (a) is CT image, (b) is MR image, (c) is embodiment design sketch, (d) is be syncretizing effect figure based on wavelet transform (DWT) based on Laplacian Pyramid Transform (Lap) syncretizing effect figure, (e).

Embodiment

Elaborate to embodiments of the invention below, the present embodiment is implemented under premised on technical solution of the present invention, give detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

The present embodiment comprises the following steps:

The first step: the average calculating source images respectively, standard deviation, entropy, the information such as greatest gradient value, as initial cluster centre, the foundation that two width image clustering centers are produced is consistent with the objective evaluation index of picture quality;

Second step: by K-meansClustering algorithm, two width source images are carried out cluster with cluster centre respectively, obtain feature space vector;

N sample is divided into K bunch J={J by K-meansClustering algorithm ₁, J ₂,-J _k, in bunch, sample standard deviation has higher similarity, and between bunch, differences between samples is obvious.If Arg={Arg ₁, Arg ₂,-Arg _kbe class center corresponding to K class, wherein Arg _kj _kthe mean value of sample in individual bunch, each bunch can be represented by the class prototype of correspondence.K-meansClustering algorithm divides data by minimizing error sum of squares criterion function in class, and its objective function is defined as follows:

$T(M,J)=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}\underset{m_x\∈J_k}{\mathrm{\Σ}}{\left|\right|m_x-{\mathrm{Arg}}_k\left|\right|}^2---\left(1\right)$

J _k＝(m _x∈M)k＝argmin||m _x-Arg _j|| ²}（2）

${\mathrm{Arg}}_k=\frac{\mathrm{\Σ}{m}_x\∈J_k{m}_x}{\left|J_k\right|}---\left(3\right)$

For 2-D data (as image), K-meansClustering algorithm can be described as, and gets a sub-picture as training sample R, if any one pixel is I in a sub-picture _xy, sample R is clustered into K bunch.

K-meansClustering algorithm mainly comprises following step:

1. initialization: a random selecting K cluster centre;

2. sample is assigned: calculate the Euclidean distance of each pixel to each class center, sample is divided into the class nearest apart from it;

3. upgrade: recalculate each Xin Culei center;

4. repeat step 2 and 3, until criterion function convergence, cluster centre is no longer changed, thus obtains K cluster.

Here criterion function is defined as error sum of squares criterion function, and its physical meaning is that the similarity between pixel represents by the distance between them usually, and distance is less, represents that pixel diversity factor is less; Distance is larger, and pixel diversity factor is larger.

3. the feature distributed areas of every width image are extracted according to feature space vector, the relatively corresponding region of two width images, in setting threshold value T(the present embodiment, T is set as the half of gradation of image average), extract coefficient in two width images be all greater than the positional information of threshold value T and extract respective regions according to this segmentation of class, be defined as texture region;

4. texture region pixel value input PCNN neural network is obtained respective Fire mapping image, get pixel that in two width images, ignition times is larger as the fusion coefficients of fused images, non-grain area pixel value is merged by binary channels PCNN;

Described initialization refers to: time initial, each neuron is all in flameout state, F _xy(0)=0, U _xy(0)=0, Y _xy(0)=0, T _xy[n]=0.

Described iteration concrete steps comprise:

A) starting condition: each neuron is all in flameout state, F _xy(0)=0, U _xy(0)=0, Y _xy(0)=0, T _xy[n]=0;

B) iterative operation: coefficient of dissociation is inputted network, by the expression of acceptance domain, modulation domain and pulses generation territory, node-by-node algorithm U _xy(n) and T _xy(n-1), and compare both sizes to determine whether produce ignition event, specifically comprise:

Neuron in the corresponding iterative operation process of PCNN is made up of acceptance domain, modulation domain and pulses generation territory:

Acceptance domain:

$F_{\mathrm{xy}}\left[n\right]=e^{-{\mathrm{\α}}_F}{F}_{\mathrm{xy}}[n-1]+V_F\mathrm{\Σ}{W}_{\mathrm{xy}}{Y}_{\mathrm{xy}}[n-1]+S_{\mathrm{xy}}---\left(4\right)$

Modulation domain:

U _xy[n]＝1+βF _xy[n]（5）

Pulses generation territory:

$Y_{\mathrm{xy}}\left[n\right]=\left\{\begin{array}{c}1,U_{\mathrm{xy}}\left[n\right]>T_{\mathrm{xy}}\left[n\right]\\ 0,\mathrm{otherwise}\end{array}\right.---\left(6\right)$

$T_{\mathrm{xy}}\left[n\right]=e^{{-\mathrm{\α}}_T}{T}_{\mathrm{xy}}[n-1]+V_T{Y}_{\mathrm{xy}}\left[n\right]---\left(7\right)$

In formula, x and y represents image each pixel transverse and longitudinal coordinate figure.S _xyrepresenting input stimulus, can be generally Laplce's energy, gradient energy, spatial frequency domain etc. of coefficient after the grey scale pixel value after the normalization of (x, y) place, decomposition.N represents iterations, F _xyrepresent feedback channel input, w _xyfor cynapse connects power, V _tfor normaliztion constant, U _xyrepresent neuronic internal activity item.β represents strength of joint, Y _xyrepresent that neuronic pulse exports, its value is 0 or 1.T _xydynamic threshold, α _t, α _ffor regulating the constant of corresponding formula, n is iterations.If U _xy[n] > T _xy[n], then neuron produces a pulse, is called and once lights a fire.In fact, after n iteration, utilize the total ignition times of (x, y) neuron to represent the information of image corresponding point position.Through PCNN igniting, the Fire mapping image be made up of the ignition times that neuron is total is as the output of PCNN.

Binary channels PCNN is the improved form to PCNN, and the neuron in corresponding iterative operation process is made up of acceptance domain, modulation domain and pulses generation territory:

Acceptance domain:

$F_{\mathrm{xy}}^A\left[n\right]=S_{\mathrm{xy}}^A\left[n\right]---\left(8\right)$

$F_{\mathrm{xy}}^B\left[n\right]=S_{\mathrm{xy}}^B\left[n\right]---\left(9\right)$

Modulation domain:

$U_{\mathrm{xy}}\left[n\right]=\max \left(F_{\mathrm{xy}}^A\right[n](1+{\mathrm{\β}}_{\mathrm{xy}}^A[n\left]\right),F_{\mathrm{xy}}^B[n\left](1+{\mathrm{\β}}_{\mathrm{xy}}^B[n\left]\right)\right)---\left(10\right)$

Pulses generation territory:

$Y_{\mathrm{xy}}\left[n\right]=\left\{\begin{array}{c}1,U_{\mathrm{xy}}\left[n\right]>T_{\mathrm{xy}}\left[n\right]\\ 0,\mathrm{otherwise}\end{array}\right.---\left(11\right)$

$T_{\mathrm{xy}}\left[n\right]=e^{{-\mathrm{\α}}_T}{T}_{\mathrm{xy}}[n-1]+V_T{Y}_{\mathrm{xy}}\left[n\right]---\left(12\right)$

Wherein: two passage xth y neuronic feed back input amounts, for external drive input, T _xyfor neuron dynamic threshold, α _tfor time constant, V _tfor normaliztion constant, U _xyfor internal activity item, β _xy ^aand β _xy ^bbe respectively weight coefficient, Y _xyfor xth y neuronic output, n is iterations.

Described acceptance domain accepts to input from the outside of two passages, respectively the different focusing source figure of corresponding two width, and these two amounts are modulated in modulating part, produce internal activity item U _xy.U _xybe input to pulses generation part and produce neuronic pulse output valve Y _xy.In described pulses generation territory, work as U _xy[n] > T _xytime [n-1], neuron is activated, and exports a pulse, meanwhile, and T _xy[n] is promoted rapidly by feedback, proceeds next iteration.Work as U _xy[n]≤T _xytime [n-1], pulse generator is closed, and stops producing pulse.Afterwards, threshold value start index declines, and works as U _xy[n] > T _xytime [n-1], pulse generator is opened, and enters new iterative loop.

C) stopping criterion for iteration: after all coefficient of dissociation all calculate, complete current iteration.

Pulse producer determines ignition event according to current threshold value, all neuron firing quantity after recording each iteration.If when iterations reaches N, stop iteration.N refers to the iterations set in network.Determine fusion coefficients:

Make I _f(x, y)=U _xy(n), I in formula _f(x, y) represents the sub-band coefficients of fused images, U _xyn () represents internal activity item, (x, y) is the pixel being positioned at xth row, y row in image, and x=1,2 ,-P, y=1,2 ,-Q, P are the total line number of image, and Q is the total columns of image.

Normalized I _f(x, y) corresponding fusion coefficients.Due to I _fsome values of (x, y) may exceed dynamic range of images value, can not directly as output image data, therefore by I _fthe value of (x, y) normalizes to [0,1].

The fusion rule that the present invention relates to refers to:

A. fusion coefficients is selected by PCNN

Excite the size of produced ignition times as the preferred index of pixel according to the neuron that pixel maps, select the fusion coefficients of correspondence position in two width images;

B. fusion coefficients is selected by binary channels PCNN

Binary channels PCNN can improve the effect of PCNN inclined dark areas feature selecting in medical image, compared with traditional single channel PCNN, binary channels PCNN is walked abreast by two simplification PCNN and forms, first by calculating with pixel A (x, y) centered by position 3*3 neighborhood in any 3 points and with the difference of other any 3 points, obtain wherein minimum value and maximal value, by the poor H of maximal value and minimum value, pass through computing obtains the β value of A (x-1, y-1).By selecting neuronic internal activity item U in two passages _xycontrol the fired state of pixel.Thus according to U _xyselect pixel U in two width figure _xythe maximum pixel as fused images.

Initialization U _xy[0]=0, T _xy[0]=0, Y _xy[0]=0,

n＝200，α _T＝0.1，α _F＝0，V _T＝25。

The U in PCNN is calculated according to formula (5)-(7) _xy[n], T _xy[n], Y _xy[n], calculates the U in binary channels PCNN according to formula (10)-(12) _xy[n], T _xy[n], Y _xy[n].

The selection rule of fusion coefficients is as follows:

$\mathrm{Cog}{\mathrm{fF}}_{\mathrm{xy}}=\left\{\begin{array}{c}\mathrm{Cogf}{A}_{\mathrm{xy}},\mathrm{if}\left(U_{\mathrm{xy}}\right[n\left]{=F^A}_{\mathrm{xy}}\right[n](1+{{\mathrm{\β}}^A}_{\mathrm{xy}})\\ {\mathrm{cogfB}}_{\mathrm{xy}},\mathrm{if}\left(U_{\mathrm{xy}}\right[n]={F^B}_{\mathrm{xy}}[n](1+{{\mathrm{\β}}^B}_{\mathrm{xy}})\end{array}\right.---\left(13\right)$

CogfF _xyrepresent fusion coefficients, CogfA _xywith CogfB _xyrepresent source figure I respectively _aand I _bthe coefficient of middle correspondence.

5. obtain fused images by fusion coefficients.

Fig. 1 is embodiment of the present invention effect schematic diagram;

In figure: (a) is CT image, (b) is MR image, (c) is embodiment design sketch, (d) is be syncretizing effect figure based on wavelet transform (DWT) based on Laplacian Pyramid Transform (Lap) syncretizing effect figure, (e).

In sum, can be seen by the effectiveness comparison of Fig. 1, this method merges the respective information of multiple focussing image better, has not only effectively enriched the background information of image, and has protected the details in image to greatest extent, met human-eye visual characteristic.So in the figure real information of fused images faithful to source, the inventive method is significantly better than the syncretizing effect based on laplacian pyramid, wavelet transform, principal component analysis (PCA) and FSDPyramid.

Q is passed through in table 1 _aB/F, mutual information (MI) weigh different fusion method obtain fused image quality, Q _aB/Frepresent that in fused images, marginal information enriches degree, MI represents that fused images contains the degree of source image information, and can be seen by data in table 1, this method is at Q _aB/F, mutual information two indices compares with additive method and all has clear improvement, the fused images that display this method generates has larger partial gradient, and grey level distribution is disperseed more, and image texture is abundanter, and details is given prominence to, and syncretizing effect is better.

Table 1 fused images objective evaluation Indexes Comparison

Last it is noted that obviously, above-described embodiment is only for example of the present invention is clearly described, and the restriction not to embodiment.For those of ordinary skill in the field, can also make other changes in different forms on the basis of the above description.Here exhaustive without the need to also giving all embodiments.And thus the apparent change of amplifying out or variation be still among protection scope of the present invention.

Claims (1)

1., based on a fusion method for the medical image of texture region division, it is characterized in that, said method comprising the steps of:

1). calculate the average of source images, standard deviation, entropy, greatest gradient value respectively, as initial cluster centre, the foundation that two width image clustering centers are produced is consistent with the objective evaluation index of picture quality;

2). by K-meansClustering algorithm, two width source images are carried out cluster with cluster centre respectively, obtain feature space vector;

3). according to the vectorial feature distributed areas of extracting every width image of feature space, compare the corresponding region of two width images, setting threshold value T, extract coefficient in two width images and be all greater than the positional information of threshold value and split extraction respective regions accordingly, be defined as texture region;

4). texture region pixel value input PCNN neural network is obtained respective Fire mapping image, gets pixel that in two width images, ignition times is larger as the fusion coefficients of fused images, non-grain area pixel value is merged by binary channels PCNN;

5). obtain fused images by fusion coefficients.