CN103479398A - Method of detecting hepatic tissue microstructure based on ultrasonic radio frequency flow analysis - Google Patents

Abstract

The invention discloses a method of detecting a hepatic tissue microstructure. The method includes the steps of 1, acquiring continuous frame-P ultrasonic echo RF signals of a hepatic tissue area, and demodulating to obtain a pattern B graph of the RF signals; 2, selecting an ROI (region of interest) on the pattern B graph to obtain ultrasonic echo RF signals in the ROI; 3, acquiring FD feature of Higuchi in each radio frequency flow in the ROI; 4, calculating power spectrum of each radio frequency flow; 5, averaging the FD feature and the power spectrum respectively; 6, extracting features of the power spectrum; 7, using a pre-trained neural network model, inputting fractal features of the averaged Higuchi and features of the averaged power spectrum into the neural network, and detecting changes of the hepatic tissue microstructure according to the output. The method has the advantages that the detection method is simple and high in accuracy, and detection results and other parameters of hepatic tissue can be combined as quantitative reference information used for determining early hepatic fibrosis.

Description

A kind of detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis

Technical field

The present invention relates to a kind of ultrasound medicine technical field, particularly a kind of detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis.

Background technology

Liver organization is a non-uniform scattering body with elastic random distribution, its ultimate unit is lobules of liver, between lobules of liver, by a small amount of connective tissue, separated, after hepatic fibrosis, lobules of liver goes to pot and the ratio of connective tissue increases, cause the micro structure of normal liver tissue and Liver fibrosis tissue to have significant difference, the difference of this micro structure has caused the elastic difference of liver, therefore by the detection to the hepatic tissue micro structure, observe the hepatic tissue micro structure and whether change, a kind of foundation of usining while as the doctor, judging hepatic fibrosis.

It is the most promising technology of tissues observed micro structure that ultrasonic tissue is levied surely, and traditional ultrasonic tissue for hepatic tissue is surely levied and mainly contained following two class methods:

First method is to use the statistical method of B ultrasonic gradation of image, main B ultrasonic gradation of image texturing method and the statistical model method of adopting, in the method owing to having used the ultrasonoscopy gray scale, therefore be subject to the impact of the imaging parameters such as adjustment of model, TGC of diasonograph very large, cause the check result concordance of different instruments poor.

Second method is to use same frame back scattering echo RF signal Spectrum Analysis method, specific practice is to adopt the wideband ultrasonic probe, obtain a certain frame radiofrequency signal, choose region of interest, to the spectrum analysis one by one of the acoustic beam in region of interest, extract the spectrum parameter, experiment shows between different tissues, the same spectrum parameter of different lesions state of organizing there are differences.Be widely used in zoopery and preliminary human clinical's experiment (such as carcinoma of prostate, thyroid carcinoma, breast carcinoma, the identification of liver etc.), studied verified its spectrum parameter and can effectively reflect formation and the feature of organizing micro structure.But hepatic tissue is non-superficial organ, and it is larger that the path that ultrasonic propagation is experienced when health check-up is looked into is affected by the individual variation of organism, and the analysis of spectrum of its same echo acoustic beam needs the depth attenuation to compensate, and these factors have all affected levies precision surely.

Summary of the invention

The object of the invention is to overcome the shortcoming of prior art with not enough, a kind of detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis is provided, whether the method can detect the hepatic tissue micro structure and change, with the parameter information of other hepatic tissues carry out in conjunction with after quantitative reference information when can be used as the doctor and judging Early hepatic fibrosis.Have advantages of that detection method simply reaches accuracy high.

Purpose of the present invention is achieved through the following technical solutions: a kind of detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis comprises the following steps:

(1) use the hepatic tissue zone under ultrasonic probe scanning liver peplos, obtain the continuous P frame ultrasonic echo radio frequency rf signal in hepatic tissue zone, a certain frame of demodulation obtains the Type B figure of ultrasonic echo radio frequency rf signal;

(2) select region of interest ROI on Type B figure, obtain the continuous P frame ultrasonic echo radio frequency rf signal of each pixel the same space position in region of interest ROI, the RF seasonal effect in time series radio frequency stream that the formation data length is the P frame;

(3) obtain the Higuchi fractal dimension FD feature of each the radio frequency stream in region of interest ROI;

(4) each radio frequency stream is carried out to fast Fourier transform, calculate the power spectrum of each radio frequency stream;

(5) Higuchi fractal dimension FD feature and the power spectrum of all radio frequency streams of regional ROI are made respectively to average treatment;

(6), for the power spectrum after average treatment, extract power spectrum characteristic;

(7) use a neural network model that training in advance is good, Higuchi fractal characteristic after average treatment in step (5) and the power spectrum characteristic in step (6) are input in the neural network model that training in advance is good, according to the output of neural network model, determine whether the hepatic tissue micro structure changes and judge intensity of variation.

Preferably, big or small M1 * M2 of Type B graph region ROI in described step (2), wherein M1 is regional ROI pixel number transversely, M2 is the pixel number of regional ROI on vertical; Include the RF time series radio frequency stream that M1 * M2 data length is P on regional ROI.

Preferably, in described step (3), radio frequency stream Higuchi fractal dimension FD feature acquisition process is as follows:

(3-1) at first the time series { x (m) } of each radio frequency stream is carried out to phase space reconstruct, the value that it is postponed to point in some regular time is processed as new dimension, obtains a bit of certain hyperspace:

Wherein

for the dimension of phase space, it is an integer, and m is zero-time, and k is fixed delay, the overall length that N is this sequence; N=P wherein;

(3-2) computation time sequence average length

for:

Obtaining average length and L (k) is:

$L\left(k\right)=\underset{m=1}{\overset{k}{\mathrm{\Σ}}}{L}_m^k,L\left(k\right)\∝k^{-D};$

(3-3) draw f (k) image, wherein f (k) is:

$f\left(k\right)=\ln \left(\frac{1}{k}\right)-\ln \left(L\left(k\right)\right);$

Choose k _minand k _max, the slope that obtains f (k) image is the fractal dimension FD of radio frequency stream.

Preferably, in described step (4), the power spectrum of each radio frequency stream obtains in the following manner: at first each radio frequency stream is imposed to isometric Hanning window or Hamming window, then zero padding to length is 1024, finally does the fast Fourier variation and obtains F (f), calculates its power spectrum; Wherein the power spectrum R (f) of each radio frequency stream is:

R(f)＝F(f)·F*(f)/P。

The power spectrum R of the radio frequency stream after the average treatment of preferably, extracting in described step (6) _rOI(f) power spectrum characteristic comprises slope slope, intercept intercept and the intermediate frequency value midbandfit of low-frequency range energy integral S1, Low Medium Frequency section energy integral S2, the band energy integration S3 of senior middle school, high band energy integral S4, normalized power spectrum fitting a straight line;

Described low-frequency range energy integral S1 is:

$S1={\∫}_{f1}^{f2}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

Described Low Medium Frequency section energy integral S2 is:

$\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="HDA0000382891500000033.png"\; state="\{\{state\}\}">S</figure-callout>}2={\&Integral;}_{f3}^{f4}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

The described band energy integration S3 of senior middle school is:

$S3={\&Integral;}_{f4}^{f5}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

Described high band energy integral S4 is:

$S4={\&Integral;}_{f5}^{f6}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

Wherein [f1, f2] drops on the low-frequency range in frequency band, and [f3, f4] drops on the Low Medium Frequency section in frequency band, and [f4, f5] drops on the high Mid Frequency in frequency band, and [f5, f6] drops on the high band in frequency band;

Described normalized power spectrum is obtained in the following manner: at first try to achieve the maximum Max of each radio frequency stream power spectrum, according to following formula, each radio frequency stream power spectrum is made to normalized, normalized power spectrum R ₀(f) be:

$R_0\left(f\right)=10.\lg \frac{R_{\mathrm{ROI}}\left(f\right)}{\mathrm{Max}};$

With the fitting a straight line based on method of least square, obtain slope slope, intercept intercept and the intermediate frequency value midbandfit of normalized power spectrum fitting a straight line:

$\mathrm{midbandfit}=\frac{\underset{f1}{\overset{f6}{\mathrm{\&Sigma;}}}{R}_0\left(f\right)}{P}\&CenterDot;$

Preferably: the neural network model in described step (7) is three layers of BP neural network model, and this neutral net is comprised of node and 3 output nodes of n input layer, a H hidden layer; Described BP neural network model learning algorithm step is as follows:

(7-1) initializing all weights W is little random value;

(7-2) input sample value x _i, i=1,2...n, by action function f ⁽¹⁾calculate hidden layer output valve z _j, j=1,2...H, send into lower one deck using this hidden layer output as input, then by action function f ⁽²⁾obtain the output valve y of neural network model _l, l=1,2,3;

Hidden layer output valve z wherein _jfor:

$z_j=f^{\left(1\right)}(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{W}_{\mathrm{ij}}\&CenterDot;x_i+b^{\left(1\right)}),j=\mathrm{1,2}...H;$

The output valve y of neural network model _lfor:

$y_l=f^{\left(2\right)}(\underset{j=1}{\overset{H}{\mathrm{\&Sigma;}}}{W}_{\mathrm{jl}}\&CenterDot;z_j+b^{\left(2\right)}),j=\mathrm{1,2}...H,l=\mathrm{1,2,3};$

F wherein ⁽¹⁾for Sigmoid function, f ⁽²⁾adopt linear function, b ⁽¹⁾and b ⁽²⁾be respectively the biasing of hidden layer and output layer;

(7-3) set end condition;

(7-4), before reaching end condition, according to the LMS algorithm, upgrade weights;

(7-5) reach end condition, stop.

Further: as follows according to the step of LMS algorithm renewal weights in described step (7-4):

(7-4-1) for each output node of network, calculate its error ε _l=d _l-y _l, l=1,2,3, d _lfor output expected value, y _l, l=1,2,3 output valves that are the neural network model output node;

(7-4-2) for each concealed nodes of network, calculate its error ε _j=d _j-z _j, j=1,2...H; z _jfor j node output valve of neural network model hidden layer, d _joutput expected value for j node of neural network model hidden layer;

(7-4-3) upgrade the network weight of every one deck, W _ij← W _ij+ Δ W _ij, Δ W _ij=η ⁽¹⁾ε _jx _i, i=1,2...n, j=1,2...H; W _jl← W _jl+ Δ W _jl, Δ W _jl=η ⁽²⁾ε _lz _j, j=1,2...H, l=1,2,3; W wherein _ijfor input neuron i to the weights between hidden neuron j, W _jlfor hidden neuron j to the weights between output neuron l; Δ W _ijfor the adjusted value of the weights between hidden neuron j and input neuron i, Δ W _jladjusted value for the weights between output neuron l and hidden neuron j;

(7-4-4) revise the weights updating method, increase momentum term, the renewal when right value update while making the L time iteration partly depends on the L-1 time iteration: Δ W _ij(L)=α Δ W _ij(L-1)+η ⁽¹⁾ε _jx _i, i=1,2...n, j=1,2...H; Δ W _jl(L)=α Δ W _jl(L-1)+η ⁽²⁾ε _lz _j, j=1,2...H, l=1,2,3;

Δ W wherein _ij(L) adjusted value of the weights while meaning the L time iteration between hidden neuron j and input neuron i, Δ W _ij(L-1) adjusted value of the weights while meaning the L-1 time iteration between hidden neuron j and input neuron i; Δ W _jl(L) adjusted value of the weights while meaning the L time iteration between output neuron l and hidden neuron j, Δ W _jl(L-1) adjusted value of the weights while meaning the L-1 time iteration between output neuron l and hidden neuron j; α is factor of momentum, η ⁽¹⁾for learning rate between input layer and hidden layer, η ⁽²⁾for learning rate between hidden layer and output layer.

Further: described BP neutral net includes 8 input layers and 8 hidden layer nodes; Be that n is that 8, H is 8, described input sample x _ibe respectively slope slope, intercept intercept and the intermediate frequency value midbandfit of Higuchi fractal characteristic, low-frequency range energy integral S1, Low Medium Frequency section energy integral S2, the band energy integration S3 of senior middle school, high band energy integral S4 and normalized power spectrum fitting a straight line after average treatment.

Further: in described step (7), the neutral net output state is y1=1, y2=0, and during y3=0, judgement hepatic tissue micro structure is normal; Output state is y1=0, y2=1, and during y3=0, the slight micro structure of judgement hepatic tissue changes; Output state is y1=0, y2=0, and during y3=1, judgement hepatic tissue micro structure severe changes.

Preferably, adopt the hepatic tissue zone under 128 array element wideband linear array ultrasonic probe scanning liver peplos in described step (1), hyperacoustic propagation is subject to the effect of hepatic tissue micro structure, echo is received by same linear array probe, through radio frequency, amplify and time gain compensation, synthesize, and then obtain the ultrasonic echo radio frequency rf signal through A/D conversion, time delay again.

The present invention has following advantage and effect with respect to prior art:

(1) the inventive method is obtained the ultrasonic echo radio frequency rf signal in hepatic tissue zone and the Type B figure of certain frame by ultrasonic probe, on Type B figure, chooses region of interest ROI, and each frame signal of same point in area-of-interest is formed to a radio frequency stream; Then obtain fractal dimension FD feature and the power spectrum of each radio frequency stream in region of interest ROI and make average treatment, extracting power spectrum characteristic; Finally the power spectrum characteristic of extraction and the Higuchi fractal characteristic after average treatment are input in neural network model as the input sample value, then the output valve by neural network model judges whether the hepatic tissue micro structure changes and intensity of variation; Output rusults is combined to rear quantitative reference information when the doctor judges Early hepatic fibrosis with other parameter informations of hepatic tissue, the dynamic observation that the inventive method is the hepatic tissue micro structure simultaneously provides noninvasive iconography means, has advantages of that detection method simply reaches accuracy high.

(2) the wideband linear array ultrasonic probe in the inventive method is at the scanning hepatic tissue, and the ultrasonic echo radio frequency rf signal that uses same probe to obtain, as signal source, can complete data acquisition, not extra spending when ultrasound diagnosis simultaneously.In the time of can avoiding checking in addition because the impact that the individual variation of organism causes ultrasonic wave propagation path (skin, fat, muscle etc.) difference to bring, because radio frequency stream derives from the same degree of depth, thereby do not need the depth attenuation to compensate, further improved and surely levied precision.

The accompanying drawing explanation

Fig. 1 is the detection method flow chart of hepatic tissue micro structure of the present invention.

Fig. 2 be in the inventive method the ultrasonic echo radio frequency rf signal obtain flow chart.

Fig. 3 is the Type B figure of the inventive method ultrasonic echo radio frequency rf signal.

Fig. 4 is the Type B figure of the continuous P frame of certain some the same space position on Type B figure region of interest ROI.

Fig. 5 is the RF seasonal effect in time series radio frequency stream that the data length of certain point of intercepting in Fig. 4 is 128 frames.

Fig. 6 is the normalized power spectrum signature figure of the inventive method.

Fig. 7 is three layers of BP neural network model figure in the inventive method.

The specific embodiment

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not limited to this.

Embodiment

As shown in Figure 1, the present embodiment discloses a kind of detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis, comprises the following steps:

(1) use the hepatic tissue zone under 128 array element wideband linear array ultrasonic probe scanning liver peplos, hyperacoustic propagation is subject to the effect of hepatic tissue micro structure, as shown in Figure 2, echo is received by same linear array probe, through radio frequency, amplify and time gain compensation, synthesize through A/D conversion, time delay, obtain the continuous 128 frame ultrasonic echo radio frequency rf signals in hepatic tissue zone, demodulation the first frame obtains the Type B figure of ultrasonic echo radio frequency rf signal as shown in Figure 3.The vibration-generating because the hepatic tissue micro structure is subject to hyperacoustic effect, it is different causing the 128 frame ultrasonic echo radio frequency rf signals that obtain, this time variation has comprised tissue characterization information.

(2) select region of interest ROI (in as Fig. 3 in the black box scope) on Type B figure as shown in Figure 3, as shown in Figure 4, obtain the continuous 128 frame ultrasonic echo radio frequency rf signals of each pixel the same space position in region of interest ROI, form the RF seasonal effect in time series radio frequency stream that data length as shown in Figure 5 is 128; Wherein the size of region of interest ROI is M1 * M2, and M1 is regional ROI pixel number transversely, and M2 is the pixel number of regional ROI on vertical; M1 * M2 is 20 * 50 in the present embodiment.Total M1 * M2 radio frequency stream on regional ROI, each radio frequency stream is the formed radio frequency time series of the ultrasonic RF signal of continuous 128 frame of the corresponding the same space of each pixel position.

(3) obtain the Higuchi fractal dimension FD feature of selecting each radio frequency stream of region of interest ROI; Wherein radio frequency stream Higuchi fractal dimension FD feature acquisition process is as follows:

(3-1) at first the time series { x (m) } of each radio frequency stream is carried out to phase space reconstruct, the value that it is postponed to point in some regular time is processed as new dimension, obtains a bit of certain hyperspace:

Wherein

for the dimension of phase space, it is an integer, and m is zero-time, and k is fixed delay, the overall length that N is this sequence;

(3-2) computation time sequence

average length

for:

Obtaining average length and L (k) is:

$L\left(k\right)=\underset{m=1}{\overset{k}{\mathrm{\&Sigma;}}}{L}_m^k,L\left(k\right)\&infin;k^{-D};$

(3-3) draw f (k) image, wherein f (k) is:

$f\left(k\right)=\ln \left(\frac{1}{k}\right)-\ln \left(L\left(k\right)\right);$

Choose k _minand k _max, the slope that obtains f (k) image is the fractal dimension FD feature of radio frequency stream, and in the present embodiment, the FD of a certain radio frequency stream is characterized as 1.4841 as calculated.

(4) each radio frequency stream is imposed to Hanning window or the Hamming window that length is 128, then zero padding to length is 1024, finally does fast Fourier transform and obtains F (f), calculates its power spectrum; Wherein the power spectrum R (f) of each radio frequency stream is:

R(f)=F(f)·F*(f)／128。

(5) Higuchi fractal dimension FD feature and the power spectrum of the M1 * M2 in regional ROI radio frequency stream are made respectively to average treatment; The Higuchi fractal dimension FD feature of M1 * M2 radio frequency stream on average obtains the meansigma methods of the Higuchi fractal dimension FD feature of regional ROI; Power spectrum R (f) the spectrum value addition of each radio frequency stream of same frequency, again divided by M1 * M2, is obtained to the average power spectra R of regional ROI _rOI(f).

(6), for the power spectrum after average treatment, extract power spectrum characteristic; The power spectrum characteristic that the present embodiment extracts comprises slope slope, intercept intercept and the intermediate frequency value midbandfit of low-frequency range energy integral S1, Low Medium Frequency section energy integral S2, the band energy integration S3 of senior middle school, high band energy integral S4 and normalized power spectrum fitting a straight line;

Low-frequency range energy integral S1 is:

$S1={\&Integral;}_{f1}^{f2}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

Low Medium Frequency section energy integral S2 is:

$\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="HDA0000382891500000033.png"\; state="\{\{state\}\}">S</figure-callout>}2={\&Integral;}_{f3}^{f4}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

The band energy integration S3 of senior middle school is:

$S3={\&Integral;}_{f4}^{f5}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

High band energy integral S4 is:

$S4={\&Integral;}_{f5}^{f6}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

Wherein [f1, f2] drops on the low-frequency range in frequency band, and [f3, f4] drops on the Low Medium Frequency section in frequency band, and [f4, f5] drops on the high Mid Frequency in frequency band, and [f5, f6] drops on the high band in frequency band;

Described normalized power spectrum is obtained in the following manner: at first try to achieve the maximum Max of regional ROI radio frequency stream power spectrum, according to following formula, ROI zone radio frequency stream power spectrum is made to normalized, normalized power spectrum R ₀(f) be:

$R_0\left(f\right)=10.\lg \frac{R_{\mathrm{ROI}}\left(f\right)}{\mathrm{Max}};$

With the fitting a straight line based on method of least square, slope slope, the intercept intercept and the intermediate frequency value midbandfit that obtain normalized power spectrum fitting a straight line are respectively:

$\mathrm{midbandfit}=\frac{\underset{f1}{\overset{f6}{\mathrm{\&Sigma;}}}{R}_0\left(t\right)}{128};$

Each spectrum signature of a certain ROI extracted in the present embodiment as shown in Figure 6, the low-frequency range energy integral S1=0.4082 0 to 0.25, the Low Medium Frequency section energy integral S2=0.1210 0.25 to 0.5, the band energy integration S3=0.0464 of senior middle school 0.5 to 0.75, the high band energy integral S4=0.0315 0.75 to 1; Slope slope=-0.1294, intercept intercept=0.1185 and the intermediate frequency value midbandfit=0.1612 of normalized power spectrum fitting a straight line in the present embodiment;

(7) use three layers of BP neural network model that training in advance is good, the BP neutral net of the present embodiment as shown in Figure 7, by n input layer, wherein n is that input feature vector number, a H hidden layer node and 3 output nodes form; Wherein in the present embodiment, n is that input feature vector is several 8, and the number H of hidden layer node is also 8; Higuchi fractal characteristic and the power spectrum characteristic in step (6) after average treatment in step (5) are input to 8 input layers in above-mentioned BP neural network model; Wherein three layers of BP neural network model learning algorithm step in the present embodiment are as follows:

(7-1) initializing all weights W is little random value;

(7-2) input sample value x _ii=1,2...n wherein inputting sample is slope slope, intercept intercept and the intermediate frequency value midbandfit of Higuchi fractal characteristic, low-frequency range energy integral S1, Low Medium Frequency section energy integral S2, the band energy integration S3 of senior middle school, high band energy integral S4 and normalized power spectrum fitting a straight line after input feature vector is respectively average treatment; By action function f ⁽¹⁾calculate hidden layer output valve z _j, j=1,2...H, send into lower one deck using this hidden layer output as input, by action function f ⁽²⁾obtain the output valve y of neural network model _l, l=1,2,3;

Hidden layer output valve z wherein _jfor:

$z_j=f^{\left(1\right)}(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{W}_{\mathrm{ij}}\&CenterDot;x_i+b^{\left(1\right)}),j=\mathrm{1,2}...H;$

The output valve y of neural network model _lfor:

$y_l=f^{\left(2\right)}(\underset{j=1}{\overset{H}{\mathrm{\&Sigma;}}}{W}_{\mathrm{jl}}\&CenterDot;z_j+b^{\left(2\right)}),j=\mathrm{1,2}...H,l=\mathrm{1,2,3};$

F wherein ⁽¹⁾for Sigmoid function, f ⁽²⁾adopt linear function, b ⁽¹⁾and b ⁽²⁾be respectively the biasing of hidden layer and output layer;

(7-3) set end condition; Wherein, in this enforcement, end condition is the minimum threshold value 10 that the nerve network system error is less than setting ^-5; End condition also can be for reaching the iterations of setting in addition.

(7-4), before reaching end condition, according to the LMS algorithm, upgrade weights; Wherein as follows according to the step of LMS algorithm renewal weights:

(7-4-1) for each output node of network, calculate its error ε _l=d _l-y _l, l=1,2,3, d _lfor output expected value, y _l, l=1,2,3 output valves that are the neural network model output node;

(7-4-2) for each concealed nodes of network, calculate its error ε _j=d _j-z _j, j=1,2...H; z _jfor j hidden layer output valve of neural network model, d _jthe output expected value that means j hidden layer;

(7-4-3) upgrade the network weight of every one deck, W _ij← W _ij+ Δ W _ij, Δ W _ij=η ⁽¹⁾ε _jx _j, i=1,2...n, j=1,2...H; W _j1← W _j1+ Δ W _j1, Δ W _jl=η ⁽²⁾ε _lz _j, j=1,2...H, l=1,2,3; W wherein _ijfor input neuron i to the weights between hidden neuron j, W _j1for hidden neuron j to the weights between output neuron l; Δ W _ijfor the adjusted value of the weights between hidden neuron j and input neuron i, Δ W _j1adjusted value for the weights between output neuron l and hidden neuron j;

(7-4-4) revise the weights updating method, increase momentum term, the renewal when right value update while making the L time iteration partly depends on the L-1 time iteration: Δ W _ij(L)=α Δ W _ij(L-1)+η ⁽¹⁾ε _jx _i, i=1,2...n, j=1,2...H; Δ W _j1(L)=α Δ W _j1(L-1)+η ⁽²⁾ε _lz _j, j=1,2...H, l=1,2,3;

Δ W wherein _ij(L) adjusted value of the weights while meaning the L time iteration between hidden neuron j and input neuron i, Δ W _ij(L-1) adjusted value of the weights while meaning the L-1 time iteration between hidden neuron j and input neuron i; Δ W _j1(L) adjusted value of the weights while meaning the L time iteration between output neuron l and hidden neuron j, Δ W _j1(L-1) adjusted value of the weights while meaning the L-1 time iteration between output neuron l and hidden neuron j.Wherein α is factor of momentum, η ⁽¹⁾for input layer-hidden layer learning rate, η ⁽²⁾for hidden layer-output layer learning rate.

(7-5) reach end condition, stop.

According to above-mentioned three layers of BP neural network model that train, Higuchi fractal characteristic after average treatment in step (5) and the power spectrum characteristic in step (6) are input in the neural network model that above-mentioned training in advance is good, then according to the output of BP neural network model, determine whether the hepatic tissue micro structure changes and judge intensity of variation.Under the present embodiment initial conditions, when output state is y1=1, y2=0, during y3=0, judgement hepatic tissue micro structure is normal; When output state is y1=0, y2=1, during y3=0, the slight micro structure of judgement hepatic tissue changes; Output state is y1=0, y2=0, and during y3=1, judgement hepatic tissue micro structure severe changes.

Above-described embodiment is preferably embodiment of the present invention; but embodiments of the present invention are not restricted to the described embodiments; other any do not deviate from change, the modification done under spirit of the present invention and principle, substitutes, combination, simplify; all should be equivalent substitute mode, within being included in protection scope of the present invention.

Claims (10)

1. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis, is characterized in that, comprises the following steps:

(1) use the hepatic tissue zone under ultrasonic probe scanning liver peplos, obtain the continuous P frame ultrasonic echo radio frequency rf signal in hepatic tissue zone, a certain frame of demodulation obtains the Type B figure of ultrasonic echo radio frequency rf signal;

(2) select region of interest ROI on Type B figure, obtain the continuous P frame ultrasonic echo radio frequency rf signal of each pixel the same space position in region of interest ROI, the RF seasonal effect in time series radio frequency stream that the formation data length is the P frame;

(3) obtain the Higuchi fractal dimension FD feature of each the radio frequency stream in region of interest ROI;

(4) each radio frequency stream is carried out to fast Fourier transform, calculate the power spectrum of each radio frequency stream;

(5) Higuchi fractal dimension FD feature and the power spectrum of all radio frequency streams of regional ROI are made respectively to average treatment;

(6), for the power spectrum after average treatment, extract power spectrum characteristic;

(7) use a neural network model that training in advance is good, Higuchi fractal characteristic after average treatment in step (5) and the power spectrum characteristic in step (6) are input in the neural network model that training in advance is good, according to the output of neural network model, determine whether the hepatic tissue micro structure changes and judge intensity of variation.

2. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis according to claim 1, it is characterized in that, big or small M1 * M2 of Type B graph region ROI in described step (2), wherein M1 is regional ROI pixel number transversely, and M2 is the pixel number of regional ROI on vertical; Include the RF time series radio frequency stream that M1 * M2 data length is P on regional ROI.

3. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis according to claim 1, is characterized in that, in described step (3), radio frequency stream Higuchi fractal dimension FD feature acquisition process is as follows:

(3-1) at first the time series { x (m) } of each radio frequency stream is carried out to phase space reconstruct, the value that it is postponed to point in some regular time is processed as new dimension, obtains a bit of certain hyperspace:

Wherein

for the dimension of phase space, it is an integer, and m is zero-time, and k is fixed delay, the overall length that N is this sequence; N=P wherein;

(3-2) computation time sequence

average length

for:

Obtaining average length and L (k) is:

$L\left(k\right)=\underset{m=1}{\overset{k}{\mathrm{\&Sigma;}}}{L}_m^k,L\left(k\right)\&Proportional;k^{-D};$

(3-3) draw f (k) image, wherein f (k) is:

$f\left(k\right)=\ln \left(\frac{1}{k}\right)-\ln \left(L\left(k\right)\right);$

Choose k _minand k _max, the slope that obtains f (k) image is the fractal dimension FD of radio frequency stream.

4. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis according to claim 1, it is characterized in that, in described step (4), the power spectrum of each radio frequency stream obtains in the following manner: at first each radio frequency stream is imposed to isometric Hanning window or Hamming window, then zero padding to length is 1024, finally do the fast Fourier variation and obtain F (f), calculate its power spectrum; Wherein the power spectrum R (f) of each radio frequency stream is:

R(f)＝F(f)·F*(f)/P。

5. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis according to claim 1, is characterized in that, the power spectrum R of the radio frequency stream after the average treatment of extracting in described step (6) _rOI(f) power spectrum characteristic comprises slope slope, intercept intercept and the intermediate frequency value midbandfit of low-frequency range energy integral S1, Low Medium Frequency section energy integral S2, the band energy integration S3 of senior middle school, high band energy integral S4, normalized power spectrum fitting a straight line;

Described low-frequency range energy integral S1 is:

$S1={\&Integral;}_{f1}^{f2}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

Described Low Medium Frequency section energy integral S2 is:

$S2={\&Integral;}_{f3}^{f4}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

The described band energy integration S3 of senior middle school is:

$S3={\&Integral;}_{f4}^{f5}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

Described high band energy integral S4 is:

$S4={\&Integral;}_{f5}^{f6}{R}_{\mathrm{ROI}}\left(f\right)\mathrm{dt};$

Wherein [f1, f2] drops on the low-frequency range in frequency band, and [f3, f4] drops on the Low Medium Frequency section in frequency band, and [f4, f5] drops on the high Mid Frequency in frequency band, and [f5, f6] drops on the high band in frequency band;

Described normalized power spectrum is obtained in the following manner: at first try to achieve the maximum Max of each radio frequency stream power spectrum, according to following formula, each radio frequency stream power spectrum is made to normalized, normalized power spectrum R ₀(f) be:

$R_0\left(f\right)=10.\lg \frac{R_{\mathrm{ROI}}\left(f\right)}{\mathrm{Max}};$

With the fitting a straight line based on method of least square, obtain slope slope, intercept intercept and the intermediate frequency value midbandfit of normalized power spectrum fitting a straight line:

$\mathrm{midbandfit}=\frac{\underset{f1}{\overset{f6}{\mathrm{\&Sigma;}}}{R}_0\left(f\right)}{P}.$

6. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis according to claim 1, it is characterized in that: the neural network model in described step (7) is three layers of BP neural network model, and this neutral net is comprised of node and 3 output nodes of n input layer, a H hidden layer; Described BP neural network model learning algorithm step is as follows:

(7-1) initializing all weights W is little random value;

(7-2) input sample value x _i, i=1,2...n, by action function f ⁽¹⁾calculate hidden layer output valve z _j, j=1,2...H, send into lower one deck using this hidden layer output as input, then by action function f ⁽²⁾obtain the output valve y of neural network model _l, l=1,2,3;

Hidden layer output valve z wherein _jfor:

$z_j=f^{\left(1\right)}(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{W}_{\mathrm{ij}}\&CenterDot;x_i+b^{\left(1\right)}),j=\mathrm{1,2},...H;$

The output valve y of neural network model _lfor:

$y_l=f^{\left(2\right)}(\underset{j=1}{\overset{H}{\mathrm{\&Sigma;}}}{W}_{\mathrm{jl}}\&CenterDot;z_j+b^{\left(2\right)}),j=\mathrm{1,2}...H,l=\mathrm{1,2,3};$

F wherein ⁽¹⁾for Sigmoid function, f ⁽²⁾adopt linear function, b ⁽¹⁾and b ⁽²⁾be respectively the biasing of hidden layer and output layer;

(7-3) set end condition;

(7-4), before reaching end condition, according to the LMS algorithm, upgrade weights;

(7-5) reach end condition, stop.

7. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis according to claim 6 is characterized in that: upgrade the step of weights according to the LMS algorithm in described step (7-4) as follows:

(7-4-1) for each output node of network, calculate its error ε _l=d _l-y _l, l=1,2,3, d _lfor output expected value, y _l, l=1,2,3 output valves that are the neural network model output node;

(7-4-2) for each concealed nodes of network, calculate its error ε _j=d _j-z _j, j=1,2...H; z _jfor j node output valve of neural network model hidden layer, d _joutput expected value for j node of neural network model hidden layer;

(7-4-3) upgrade the network weight of every one deck, W _ij← W _ij+ Δ W _ij, Δ W _ij=η ⁽¹⁾ε _jx _i, i=1,2...n, j=1,2...H; W _j1← W _j1+ Δ W _j1, Δ W _jl=η ⁽²⁾ε _lz _j, j=1,2...H, l=1,2,3; W wherein _ijfor input neuron i to the weights between hidden neuron j, W _j1for hidden neuron j to the weights between output neuron l; Δ W _ijfor the adjusted value of the weights between hidden neuron j and input neuron i, Δ W _j1adjusted value for the weights between output neuron l and hidden neuron j;

(7-4-4) revise the weights updating method, increase momentum term, the renewal when right value update while making the L time iteration partly depends on the L-1 time iteration: Δ W _ij(L)=α Δ W _ij(L-1)+η ⁽¹⁾ε _jx _i, i=1,2...n, j=1,2...H; Δ W _j1(L)=α Δ W _j1(L-1)+η ⁽²⁾ε _lz _j, j=1,2...H, l=1,2,3;

Δ W wherein _ij(L) adjusted value of the weights while meaning the L time iteration between hidden neuron j and input neuron i, Δ W _ij(L-1) adjusted value of the weights while meaning the L-1 time iteration between hidden neuron j and input neuron i; Δ W _j1(L) adjusted value of the weights while meaning the L time iteration between output neuron l and hidden neuron j, Δ W _j1(L-1) adjusted value of the weights while meaning the L-1 time iteration between output neuron l and hidden neuron j; α is factor of momentum, η ⁽¹⁾for learning rate between input layer and hidden layer, η ⁽²⁾for learning rate between hidden layer and output layer.

8. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis according to claim 7, it is characterized in that: described BP neutral net includes 8 input layers and 8 hidden layer nodes; Be that n is that 8, H is 8, described input sample x _ibe respectively slope slope, intercept intercept and the intermediate frequency value midbandfit of Higuchi fractal characteristic, low-frequency range energy integral S1, Low Medium Frequency section energy integral S2, the band energy integration S3 of senior middle school, high band energy integral S4 and normalized power spectrum fitting a straight line after average treatment.

9. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis according to claim 8, it is characterized in that: in described step (7), the neutral net output state is y1=1, y2=0, during y3=0, judgement hepatic tissue micro structure is normal; Output state is y1=0, y2=1, and during y3=0, the slight micro structure of judgement hepatic tissue changes; Output state is y1=0, y2=0, and during y3=1, judgement hepatic tissue micro structure severe changes.

10. the detection method of the hepatic tissue micro structure based on ultrasonic radio frequency flow analysis according to claim 1, it is characterized in that, adopt the hepatic tissue zone under 128 array element wideband linear array ultrasonic probe scanning liver peplos in described step (1), hyperacoustic propagation is subject to the effect of hepatic tissue micro structure, echo is received by same linear array probe, amplify and time gain compensation through radio frequency, then synthesize, and then obtain the ultrasonic echo radio frequency rf signal through A/D conversion, time delay.

Images