CN101908891A - Medical image ROI (Region of Interest) compression method based on lifting wavelet and PCNN (Pulse Coupled Neural Network) - Google Patents

Abstract

The invention publishes a medical image ROI compression method based on a lifting wavelet and a PCNN, which comprises the following steps of: circling a region of interest by a doctor, and separating the region of interest from a region of no interest by a difference image method; adopting lossless compression in the region of interest, constructing compactly supported biorthogonal wavelet transformation through a lifting scheme, and then carrying out Huffman encoding; adopting lossy compression in the region of no interest, segmenting gray-value pixel approximate points through the PCNN, carrying out ignition operation, and then carrying out run-length encoding; and finally carrying out inverse transformation restoration, merging the region of interest and the region of no interest, and eliminating a boundary discontinuity problem through linear interpolation. Experimental results prove that the region of interest can be flexibly selected and controlled by the compression method, the used information for doctor diagnosis can be completely reserved, and the compression ratio is higher. Meanwhile, the computation for an ROI mask and the computation and the encoding for a wavelet coefficient difference value are omitted, the compression and decompression time and the algorithm complexity are reduced, and the image processing and transmitting efficiency is improved.

Description

Medical image ROI compression method based on Lifting Wavelet and PCNN

Technical field

The present invention designs a kind of medical image ROI compression method based on Lifting Wavelet and PCNN.

Background technology

Medical image is than normal image resolution height, quantized level is many, data volume is big, development along with PACS (Picture Archiving andCommunication Systems) image communication system and remote live medical treatment, be badly in need of Medical Image Compression, promptly under the prerequisite that guarantees the image service quality, the message bit pattern of medical image is converted into a kind of array form of data volume reduction.Image compression roughly is divided into lossless compress and lossy compression method at present.Lossless compress can recover original image fully, and detailed information is not lost, but its compression ratio is generally between 50% to 80%, and compression back data volume is still very big.Lossy compression method can not recover former figure fully, and its compression ratio can reach 5%, but some material particular loss can influence the auxiliary diagnosis of doctor to disease.

As seen, infeasible to the entire image lossy compression method, the compression ratio of lossless compress is very low again.In order to resolve contradiction, many experts propose area-of-interest (ROI, Regions of Interest) compressed encoding, (see document [1]: JoelAskelof as Maxshift, Mathias Larsson Carlander, Charilaos Christopoulos.Region of interest coding inJPEG2000[J] .Signal Processing:Image Communication, 2002,17:105～111.), PSBshift (sees document [2]: Liu Jie, Fan Guoliang.A new JPEG2000 region-of-interest image coding method:Partial significant bitplanes shift[J] .IEEE Signal Processing Letters, 2003,10 (2): 35～38), (see document [3]: ZHUO Li based on spiht algorithm, SHEN Lansun, Lam Kinman.Region of Interest BasedImage Coding and Progressive Transmission[J] .Chinese Journal of Electronics, 2004,32 (3): 411～415.), these algorithm basic principle all are to image wavelet transform, point out to be arranged in the wavelet coefficient part of ROI with the ROI mask, on move wavelet coefficient in the ROI, allow it be positioned at the higher bit plane, be encoded prior to non-ROI.The lossless compress when wavelet coefficient in the ROI is moved on all, lossy compression method when moving on the part is partly encoded or is directly omitted non-ROI.In wavelet transformation, the traditional double orthogonal wavelet can only carry out lossy compression method and (see document [4]: WangXiangyang, Yang Hongying, Gao Cunchen.The Image Compression Based on WaveletTransform[J] .Computer Engineering and Applications, 2001,15:82～84+159; Document [5]: ZHANG Guo-yun, PENG Shi-yu.Image compression based on anti-symmetrical biorthogonalwavelets[J] .Computer Engineering, 2008,34 (6): 205-209), and integer lifting wavelet transform (is seen document [6]: LI Shi-peng, LI Wei-ping.Shape-adaptive Discrete Wavelet Transform for ArbitrarilyShaped Visual Object Coding[J] .IEEE Transactions on Circuits and Systems for VideoTechnology, 2000,10 (5): 725-743; Document [7]: Chen Hongxin, et al.Memory efficient imagecompressing using lifting scheme[A] .International Conference on Signal ProcessingProceedings (ICSP ' 04) [C] .2004.), the advantage of inheriting tradition wavelet multiresolution rate, do not rely on Fourier transform, set up the algorithm identical with the Mallat function, implementation (in-place) computing of ascending the throne, operand reduces about 30%, process is simple, realize the conversion of set of integers, can realize lossless compress to set of integers.Aspect lossy compression method, Eckhorn proposes the connection of pulse granting and the basic model (PCNN of Pulse Coupled Neural Network, Pulse Coupled Neural Network) (sees document [8]: Ranganath H S, Kuntimad G.Iterative Segmentation Using Pulse Coupled NeuralNetworks[C] .Proc.SPIE, 1996, (2760): 543～554), can well cut apart to image.PCNN was cut apart in the past, the image encoding compression algorithm of rebuilding with orthogonal basis (is seen document [9]: MA Yi-de again, LI Lian, DAIRuo-lan.Automated image segmentation using PCNN and entropy of image[J] .Journal ofChina Institute of Communicationgs, 2002,29 (3): 49-51) effective, but the process of the setting up complexity of orthogonal basis when rebuilding.

Summary of the invention

The present invention seeks to provides a kind of medical image ROI compression method based on Lifting Wavelet and PCNN at the defective that prior art exists.

The present invention adopts following technical scheme for achieving the above object:

The present invention is based on the medical image ROI compression method of Lifting Wavelet and PCNN, it is characterized in that comprising the steps: preliminary treatment

Adopt man-machine interactively formula difference shadow dividing method to carry out medical image segmentation, again with area-of-interest in the difference shadow method Medical Image Segmentation and non-area-of-interest and produce the two-value mask, view picture medical image and two-value mask being multiplied each other obtains the difference shadow again, thereby picture is divided into area-of-interest and non-area-of-interest;

Lossy compression method

Non-area-of-interest is cut apart differentiation grey scale pixel value discrete point, the zones of different of dividing non-region of interest area image through PCNN; The close scope of gray value is set according to actual needs, carries out Run-Length Coding after the close pixel process igniting computing;

Lossless compress

Area-of-interest after adapting to conversion, integer 5/3 small echo is carried out the Ha Fuman coding through promoting;

Recover

Non-area-of-interest after the lossy compression method and the area-of-interest after the lossless compress are recovered medical image through merging, linear interpolation successively, realize medical image ROI compression.

Preferably, it is as follows that described lifting integer 5/3 small echo adapts to transform method:

With low pass synthesis filter h (z), high pass synthesis filter g (z), lowpass analysis filter

With the high pass analysis filter

All resolve into even number and odd number two parts:

h(z)＝h _e(z ²)+z ^-1h _o(z ²)

g(z)＝g _e(z ²)+z ^-1g _o(z ²)

(1)

$h^{\%}\left(z\right)=h_e^{\%}\left(z^2\right)+z^{-1}{h}_o^{\%}\left(z^2\right)$

$g^{\%}\left(z\right)=g_e^{\%}\left(z^2\right)+z^{-1}{g}_o^{\%}\left(z^2\right)$

Wherein, subscript e idol coefficient multinomial, subscript o represents strange coefficient multinomial, promptly

$h_e\left(z\right)=\underset{k}{\mathrm{\Σ}}{h}_{2k}{z}^{-k},$ $h_o\left(z\right)=\underset{k}{\mathrm{\Σ}}{h}_{2k+1}{z}^{-k}---\left(2\right)$

In the formula, h _2k, h _2k+1Be respectively h _e(z) and h _o(z) press z ^-1The even number sequence number of launching, the coefficient of odd indexed.

And multinomial is defined as: $P\left(z\right)=\left[\begin{array}{cc}{h}_e\left(z\right)& g_e\left(z\right)\\ h_o\left(z\right)& g_o\left(z\right)\end{array}\right]---\left(3\right)$

Input signal carries out odd even too and decomposes, output low, and the z conversion of high fdrequency component represents to be respectively s ₁(z), d ₁(z), then decomposable process is:

$\left[\begin{array}{c}{s}_1\left(z\right)\\ d_1\left(z\right)\end{array}\right]=P^T\left(z\right)\left[\begin{array}{c}{x}_e\left(z\right)\\ z^{-1}{x}_o\left(z\right)\end{array}\right]---\left(4\right)$

(z) reaches Via Lifting Scheme below by decomposed P:

By $h_e\left(z^2\right)=\frac{h\left(z\right)+h(-z)}{2},$ $h_o\left(z^2\right)=\frac{h\left(z\right)-h(-z)}{2}z---\left(5\right)$

Push away $P{\left(z^2\right)}^T=\frac{1}{2}M\left(z\right)\left[\begin{array}{cc}1& z\\ 1& -z\end{array}\right]---\left(6\right)$

$P^{\%}{\left(z^{-1}\right)}^T=\left[\begin{array}{cc}{h}_e^{\%}\left(z^{-1}\right)& h_o^{\%}\left(z^{-1}\right)\\ g_e^{\%}\left(z^{-1}\right)& g_o^{\%}\left(z^{-1}\right)\end{array}\right]---\left(7\right)$

Then desirable reconstruction condition When (h, g) constitute complementary filter to the time, detP (z)=1 (det show determinant), after doing an antithesis and promoting, newly polynomial matrix is:

$P^{\mathrm{new}}\left(z\right)=P\left(z\right)\left[\begin{array}{cc}1& 0\\ t\left(z\right)& 1\end{array}\right]---\left(8\right)$

By the Euclid algorithm, $\left[\begin{array}{c}{h}_e\left(z\right)\\ h_o\left(z\right)\end{array}\right]=\underset{i=1}{\overset{n}{\mathrm{\Π}}}\left[\begin{array}{cc}{q}_i\left(z\right)& 1\\ 1& 0\end{array}\right]\left[\begin{array}{c}K\\ 0\end{array}\right]---\left(9\right)$

Therefore $P\left(z\right)=\underset{i=1}{\overset{\mathrm{<figure-callout\; id="1"\; label="m"\; filenames="HSA00000240662500013.png"\; state="\{\{state\}\}">m</figure-callout>}}{\mathrm{\&Pi;}}}\left[\begin{array}{cc}1& 0\\ -s_i\left(z^{-1}\right)& 1\end{array}\right]\left[\begin{array}{cc}1& -t_i\left(z^{-1}\right)\\ 0& 1\end{array}\right]\left[\begin{array}{cc}{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="HSA00000240662500013.png"\; state="\{\{state\}\}">K</figure-callout>}}_1& 0\\ 0& K_2\end{array}\right]---\left(10\right)$

K ₁And K ₂Be constant, s _i(i=1,2, L, m is integer) and be the filter of prediction lifting step, t _iBe the filter that upgrades lifting step, promptly corresponding to the wavelet transformation of P (z) forward, the wavelet transformation of formation comprises: (Update) and 4 steps of normalization (Scaling) are upgraded in division (Split), prediction (Predict):

1. divide original signal sequence x (n) is split into two mutually disjoint even number sequence number subclass x _e(n) and odd indexed subclass x _o(n), i.e. inertia (lazy) dividing method:

x(n)＝{x _e(n)，x _o(n)} (11)

2. predict according to the correlation between data, use x _e(n) dope x _o(n), concrete grammar is that the average with adjacent 2 even numbers is used as the odd number predicted value between them, goes to substitute this odd bits with the difference of odd bits and predicted value again, and expression formula is:

$x_o^{\%}\left(n\right)=x_o\left(n\right)-\frac{x_e(n-1)+x_e\left(n\right)}{2}---\left(12\right)$

3. upgrade and use x _o(n) and s _i(z) multiply each other and add x _e(n) realize:

Concrete method be with the predicted value of last position odd bits and back one digit pair digit value and 1/4th as the degree of deviation values of adjusting, remove alternative this even bit with the difference of this even bit and degree of deviation value again, expression formula is:

$x_e^{\%}\left(n\right)=x_e\left(n\right)-\frac{x_o^{\%}(n-1)+x_e(n+1)}{4}---\left(13\right)$

4. normalization realization coefficient is unified conversion, and promptly the even number with output partly multiply by scale factor K ₂, odd number partly multiply by scale factor K ₁

After repeating said process, obtain after n the decomposition

Represented the low frequency part of signal; Represented the HFS of signal.

Preferably, described PCNN dividing method is as follows:

The neuronal structure model of PCNN model is divided into three parts: the input area is input variable I _IjWith adjacent neurons output pulse Y _IjConnecting the input area is neuronic coupling part; Pulse produces district's passing threshold adjustment and produces pulse Y; Each discrete mathematical variable iterative equation of this model is as follows:

$F_{\mathrm{ij}}\left(n\right)=\exp (-{\mathrm{\&alpha;}}_F)F_{\mathrm{ij}}(n-1)+V_F\underset{\mathrm{kl}}{\mathrm{\&Sigma;}}{M}_{\mathrm{ijkl}}{Y}_{\mathrm{kl}}(n-1)+I_{\mathrm{ij}}---\left(14\right)$

$L_{\mathrm{ij}}\left(n\right)=\exp (-{\mathrm{\&alpha;}}_L)L_{\mathrm{ij}}(n-1)+V_L\underset{\mathrm{kl}}{\mathrm{\&Sigma;}}{W}_{\mathrm{ijkl}}{Y}_{\mathrm{kl}}(n-1)---\left(15\right)$

U _ij(n)＝F _ij(n)(1+βL _ij(n)) (16)

$Y_{\mathrm{ij}}\left(n\right)=\left\{\begin{array}{cc}1,& U_{\mathrm{ij}}\left(n\right)>E_{\mathrm{ij}}(n-1)\\ 0,& U_{\mathrm{ij}}\left(n\right)\&le;E_{\mathrm{ij}}(n-1)\end{array}\right.---\left(17\right)$

E _ij(n)＝exp(-α _E)E _ij(n-1)+V _E∑Y _kl(n-1) (18)

(13) I in the formula _IjBe that (i, the grey scale pixel value of j) locating are neuron and force the external drive excite, Y at point for the picture element matrix of image _KlBe that adjacent neuron is at point (k, the output of l) locating, F _IjBe the linear, additive result of input, M _IjklBe for feedback input domain mid point (i, j) and point (k, weight matrix l), α _FBe damping time constant, V _FIt is the amplification coefficient in the feedback input domain; (14) L in the formula _IjBe connect the input area be of coupled connections α _LAnd V _LBe respectively its damping time constant and amplification coefficient, W _IjklBe point (i, j) and point (k, the weight matrix of connection matrix l); (15) U in the formula _IjBe the internal activity item, β is its coefficient of connection, E in (16) formula _IjBe the dynamic moving thresholding of excitation pulse, α in (17) formula _EAnd V _EBe respectively its damping time constant and amplification coefficient;

Passing threshold adjustment decision dynamic moving thresholding, the picture element matrix point close to the adjacent area gray-scale pixel values produces pulse, and is designated as 1, by Y _IjMatrix output, thus the grey scale pixel value discrete point distinguished, and the partitioned image zones of different realizes image segmentation.

Preferably, work as α _F＜α _L＜α _EThe time, each neuron all is in the square connection matrix center of a n * n among the PCNN, and general n gets 3, and it connects weights and gets W=M.

The medical image ROI compression method that the present invention is based on Lifting Wavelet and PCNN can be issued to the undistorted of pathological area in the situation that guarantees the good compression rate, realizes simple.Obtain the region of interest area image with difference shadow method, need not to calculate the just correctly amalgamation of ROI mask and location parameter, consuming time low, real-time, recovery effects is good.Can decide the non-ROI distortion factor and compression ratio by adjustment, adapt to the different situations needs flexibly PCNN internal activity item.Satisfy the transmission quality and the transmission speed of medical image, kept efficient diagnosis and efficient storage.Along with the lifting of hardware system, the real-time of compression has bigger raising, has widely to use.

Description of drawings

Fig. 1: compression algorithm overall flow figure;

Fig. 2: ROI schematic diagram;

Fig. 3: pulse coupled neural meta-model figure;

Fig. 4: comparison diagram before and after the area-of-interest compression;

Fig. 5: comparison diagram before and after the non-area-of-interest compression;

Fig. 6: former figure;

Fig. 7: the figure after the integer lifting wavelet lossless compress is restored;

Fig. 8: β of the present invention is the figure behind 0.3 compression recovery;

Fig. 9: β of the present invention is the figure behind 3 compression recoveries.

Embodiment

As shown in Figure 1, this method is divided into preprocessing process on the whole, lossy compression method process, lossless compress process, and recovery process.In preprocessing process, cutting apart for area-of-interest has automatically and artificial two kinds of ways.Automatically cutting apart is double-hump characteristics according to the grey level histogram of pathological area in the medical image, utilizes single or many threshold values are cut apart, but many threshold values automatic division method is also incomplete, more consuming time, and with doctor's judgement bigger gap is arranged.So this paper adopts man-machine interactively formula difference shadow dividing method to carry out image segmentation, this method is to iris out interested pathological area (as shown in Figure 2) by the doctor with mouse, with difference shadow method area-of-interest and non-area-of-interest are separated again, difference shadow method is meant to choose with mouse and produces the two-value mask behind the area-of-interest, again view picture medical picture and two-value mask are multiplied each other, get its difference shadow, thereby picture is divided into area-of-interest and non-area-of-interest, just can compress it respectively.

First generation wavelet transformation process, in the decomposition and reconstruct of sub-band transforms coding, decomposition is that input signal x (n) is passed through high pass respectively

And low pass

Two analysis filters are made 2 times of down-samplings, obtain the high and low frequency coefficient of signal, and reconstruction is that the high and low frequency coefficient is made 2 times of up-samplings earlier, more respectively by high pass g (z) and two synthesis filters of low pass h (z).The complete reconstruction condition that constitutes above-mentioned bank of filters is:

$h\left(z\right)h^{\%}\left(z^{-1}\right)+g\left(z\right)g^{\%}\left(z^{-1}\right)=2,$

$h\left(z\right)h^{\%}(-z^{-1})+g\left(z\right)g^{\%}(-z^{-1})=0---\left(1\right)$

Formula (1) is expressed as with the polynomial matrix form

$M^{\%T}\left(z^{-1}\right)M\left(z\right)=2I---\left(2\right)$

Definition amplitude modulation matrix M (z) is in the formula (2)

$M\left(z\right)=\left[\begin{array}{cc}h\left(z\right)& h(-z)\\ g\left(z\right)& g(-z)\end{array}\right]---\left(3\right)$

Antithesis amplitude modulation matrix for M (z).

Wavelet coefficient is a floating type after the conversion like this, rebuilds after the quantification to be certain to that distortion is arranged again, and can not be used for lossless compress, and inefficiency, if still up-sampling is before bank of filters, efficient will be enhanced.Its basic principle is the diagonal matrix that the polynomial matrix of wavelet filter is decomposed into a series of upper triangular matrix, lower triangular matrix and a constant, makes wavelet transformation realize by these matrix multiples.

This method for improving with filter h (z), g (z), With

All resolve into even number and odd number two parts:

h(z)＝h _e(z ²)+z ^-1h _o(z ²) (4)

Wherein, h _e(z) be even coefficient multinomial, h _o(z) be strange coefficient multinomial, promptly

$h_e\left(z\right)=\underset{k}{\mathrm{\&Sigma;}}{h}_{2k}{z}^{-k},$ $h_o\left(z\right)=\underset{k}{\mathrm{\&Sigma;}}{h}_{2k+1}{z}^{-k}---\left(5\right)$

And multinomial is defined as: $P\left(z\right)=\left[\begin{array}{cc}{h}_e\left(z\right)& g_e\left(z\right)\\ h_o\left(z\right)& g_o\left(z\right)\end{array}\right]---\left(6\right)$

Input signal carries out odd even too and decomposes, output low, and the z conversion of high fdrequency component represents to be respectively s ₁(z), d ₁(z), then decomposable process is

$\left[\begin{array}{c}{s}_1\left(z\right)\\ d_1\left(z\right)\end{array}\right]=P^T\left(z\right)\left[\begin{array}{c}{x}_e\left(z\right)\\ z^{-1}{x}_o\left(z\right)\end{array}\right]---\left(7\right)$

(z) reaches Via Lifting Scheme below by decomposed P:

By $h_e\left(z^2\right)=\frac{h\left(z\right)+h(-z)}{2},$ $h_o\left(z^2\right)=\frac{h\left(z\right)-h(-z)}{2}z---\left(8\right)$

Push away $P{\left(z^2\right)}^T=\frac{1}{2}M\left(z\right)\left[\begin{array}{cc}1& z\\ 1& -z\end{array}\right]---\left(9\right)$

Similar

Definition

Then

When (h, g) constitute complementary filter to the time, detP (z)=1, after doing an antithesis and promoting, new polynomial matrix is

$P^{\mathrm{new}}\left(z\right)=P\left(z\right)\left[\begin{array}{cc}1& 0\\ t\left(z\right)& 1\end{array}\right]---\left(10\right)$

By the Euclid algorithm, $\left[\begin{array}{c}{h}_e\left(z\right)\\ h_o\left(z\right)\end{array}\right]=\underset{i=1}{\overset{n}{\mathrm{\&Pi;}}}\left[\begin{array}{cc}{q}_i\left(z\right)& 1\\ 1& 0\end{array}\right]\left[\begin{array}{c}K\\ 0\end{array}\right]---\left(11\right)$

Therefore $P\left(z\right)=\underset{i=1}{\overset{\mathrm{<figure-callout\; id="1"\; label="m"\; filenames="HSA00000240662500013.png"\; state="\{\{state\}\}">m</figure-callout>}}{\mathrm{\&Pi;}}}\left[\begin{array}{cc}1& 0\\ -s_i\left(z^{-1}\right)& 1\end{array}\right]\left[\begin{array}{cc}1& -t_i\left(z^{-1}\right)\\ 0& 1\end{array}\right]\left[\begin{array}{cc}{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="HSA00000240662500013.png"\; state="\{\{state\}\}">K</figure-callout>}}_1& 0\\ 0& K_2\end{array}\right]---\left(12\right)$

K ₁And K ₂Be constant, s _iBe the filter of prediction lifting step, t _iBe the filter that upgrades lifting step, promptly corresponding to the wavelet transformation of P (z) forward, the wavelet transformation of formation comprises: (Update) and 4 steps of normalization (Scaling) are upgraded in division (Split), prediction (Predict).

1. divide original signal sequence x (n) is split into two mutually disjoint even number sequence number subclass x _e(n) and odd indexed subclass x _o(n), i.e. inertia (lazy) dividing method.

x(n)＝{x _e(n)，x _o(n)} (13)

2. predict and to use x according to the correlation between data _e(n) dope x _o(n), promptly use x _e(n) and t _i(z) multiply each other and add x _o(n) realize.

Concrete grammar is that the average with adjacent 2 even numbers is used as the odd number predicted value between them, go to substitute this odd bits with the difference of odd bits and predicted value again, this difference has reflected the degree of approximation of the two, and it is reasonable more to predict, the amount of information that comprises is just lacked more than original amount of information, expression formula is:

$x_o^{\%}\left(n\right)=x_o\left(n\right)-\frac{x_e(n-1)+x_e\left(n\right)}{2}---\left(14\right)$

3. after being updated in the step of above-mentioned two processes, the coefficient subclass of generation generally can not keep the consistency of some overall permanence in the original image, therefore, adopt renewal process, promptly uses x _o(n) and s _i(z) multiply each other and add x _e(n) realize.

Concrete method be with the predicted value of last position odd bits and back one digit pair digit value and 1/4th as the degree of deviation values of adjusting, remove to substitute this even bit with the difference of this even bit and degree of deviation value again, so that it keeps the overall permanence of legacy data collection, expression formula is:

$x_e^{\%}\left(n\right)=x_e\left(n\right)-\frac{x_o^{\%}(n-1)+x_e(n+1)}{4}---\left(15\right)$

4. normalization realization coefficient is unified conversion, and promptly the even number with output partly multiply by scale factor K ₂, odd number partly multiply by scale factor K ₁

After repeating said process, obtain after n the decomposition

Represented the low frequency part of signal;

Represented the HFS of signal.

Lifting step during reconstruct is its inverse transformation process.What this paper selected for use is that 5 grades of integer 5/3 Lifting Wavelet are decomposed, after using integer lifting wavelet transform, pixel value to image itself has carried out conversion, and kept the consistency of the integer before and after the conversion, inverse transformation process can be able to recover the pixel value before decomposing fully, can realize the lossless compress of area-of-interest.

Coefficient after carrying out after 5 small echos promote conversion, utilizing Huffman encoding to conversion is again encoded, and can well remove the correlation between data, has reduced the amount of information of original image greatly.

Huffman encoding is constructed the shortest elongated coding of average length according to the probability that data occur, and is an application of Huffman tree, and detailed process is:

1. with the coefficient behind the wavelet transformation by the probability that the occurs sequence arrangement of successively decreasing.

2. with 2 probability combination additions of minimum, obtain and be worth again and repeat this process with remaining probability, always by the probability arrangement of successively decreasing, with probability big be placed on above, till probability is 1.

3. each is carried out code assignment to combination, be encoded to 1 as if bigger, then another volume is 0, and oppositely distribution also can.

4. each coefficient is arranged from right to left, the distribution codeword sequence obtains elongated Huffman encoding.

The generation of Huffman encoding is not a compression process, in order to realize compression, for the symbol that produces code, must carry out conversion or mapping to it according to the code that generates.The Huffman method utilizes the unit width histogram between maximum and the minimum value that input matrix is encoded.As long as utilize minimum value and histogram during decoding and rebuilding, substitute the Huffman code itself that keeps, thereby can well reduce its amount of information.

The neuronal structure of PCNN model as shown in Figure 3.It is the model of individual layer, does not need training process, utilizes the distinctive linear, additive of neuron, and the non-linear modulation coupling of multiplying each other can be used for image segmentation and target classification etc.

This structural model is divided into three parts: the input area is input variable I _IjWith adjacent neurons output pulse Y _IjConnecting the input area is neuronic coupling part; Pulse produces district's passing threshold adjustment and produces pulse Y.Each discrete mathematical variable iterative equation of this model is as follows:

$F_{\mathrm{ij}}\left(n\right)=\exp (-{\mathrm{\&alpha;}}_F)F_{\mathrm{ij}}(n-1)+V_F\underset{\mathrm{kl}}{\mathrm{\&Sigma;}}{M}_{\mathrm{ijkl}}{Y}_{\mathrm{kl}}(n-1)+I_{\mathrm{ij}}---\left(16\right)$

$L_{\mathrm{ij}}\left(n\right)=\exp (-{\mathrm{\&alpha;}}_L)L_{\mathrm{ij}}(n-1)+V_L\underset{\mathrm{kl}}{\mathrm{\&Sigma;}}{W}_{\mathrm{ijkl}}{Y}_{\mathrm{kl}}(n-1)---\left(17\right)$

U _ij(n)＝F _ij(n)(1+βL _ij(n)) (18)

$Y_{\mathrm{ij}}\left(n\right)=\left\{\begin{array}{cc}1,& U_{\mathrm{ij}}\left(n\right)>E_{\mathrm{ij}}(n-1)\\ 0,& U_{\mathrm{ij}}\left(n\right)\&le;E_{\mathrm{ij}}(n-1)\end{array}\right.---\left(19\right)$

E _ij(n)＝exp(-α _E)E _ij(n-1)+V _E∑Y _kl(n-1) (20)

(16) I in the formula _IjBe that (i, the grey scale pixel value of j) locating are neuron and force the external drive excite, Y at point for the picture element matrix of image _KlBe that adjacent neuron is at point (k, the output of l) locating, F _IjBe the linear, additive result of input, M _IjklBe for feedback input domain mid point (i, j) and point (k, weight matrix l), α _FBe damping time constant, V _FIt is the amplification coefficient in the feedback input domain; (17) L in the formula _IjBe connect the input area be of coupled connections α _LAnd V _LBe respectively its damping time constant and amplification coefficient, W _IjklBe point (i, j) and point (k, the weight matrix of connection matrix l); (18) U in the formula _IjBe the internal activity item, β is its coefficient of connection, E in (19) formula _IjBe the dynamic moving thresholding of excitation pulse, α in (20) formula _EAnd V _EBe respectively its damping time constant and amplification coefficient.

Work as α _F＜α _L＜α _EThe time, each neuron all is in the connection matrix center of a n * n among the PCNN, is generally 3 * 3 or 5 * 5, and it connects weights a variety of selections, generally gets W=M.

Passing threshold adjustment decision dynamic moving thresholding, the picture element matrix point close to the adjacent area gray-scale pixel values produces pulse, and is designated as 1, by Y _IjMatrix output, thus the grey scale pixel value discrete point distinguished, and the partitioned image zones of different realizes image segmentation.

Run-Length Coding is a kind of of entropy coding, and the thinking of coding is simple relatively, mainly is meant a string continuous identical number is represented with its value and its string length.Be used on the image is exactly representing with its pixel value and continuous length that with the identical pixel of the Continuous Gray Scale value of delegation its length is exactly the distance of swimming.For example, certain delegation has a string A continuous pixel to have identical gray scale B, then behind the Run-Length Coding, with (B, A) such two numbers are represented the value of A pixel.

It is simple that Run-Length Coding has a coding, is easy to storage behind the coding, and encoding and decoding speed is fast, the advantage that is well suited for for real-time application.Single words of image being compressed with Run-Length Coding are exactly lossless compress.

It is not good that but Run-Length Coding is directly used in the compression effectiveness of multi-grey image, and the tangible picture of gray-value variation is also inapplicable, and the poorest situation is gray value different with on every side all of every, and compression back memory space not only can not reduce to increase on the contrary like this.For the more zone of the continuous color lump of background area, fine such as the large tracts of land black background zone effect that is used for ultrasonic picture.Run-Length Coding all is mixed together use with other coding method generally speaking.

The effect of directly using Run-Length Coding in the multi-grey image is bad, and this is the factor decision of Run-Length Coding itself.For the unconspicuous zone of grey scale change as the ultrasonic medical image background area, its details is little for the use of medical science diagnosis and treatment, but keeping its integrality judges pathology and plays the position with reference to effect, we can be suitable the pixel that its gray value is close become the identical point of gray value, so just can well use Run-Length Coding.

PCNN is used in good effect on the image segmentation, can well split the close zone of gray value.Close and the adjacent pixels point of gray value by PCNN, can produce same pulse, and remaining point does not produce pulse.Passing threshold adjustment like this, we can control the gray values of pixel points excursion that produces pulse, thus the distortion factor of lossy compression method after can controlling.The pulse that PCNN handles the back generation is exactly a two-value mask matrix, and the corresponding position of the similar adjacent pixels point of gray value is 1, and all the other are 0.Programming realizes ignition process again, and every row is that the pixel of 1 position correspondence is averaged gray value and composes back the value that substitutes former pixel more continuously in the two values matrix with this generation.When carrying out Run-Length Coding so again, identical value will increase greatly, and compression ratio also can increase substantially, and Run-Length Coding just can well be used, and the distortion factor can be controlled very easily by threshold value adjustment among the PCNN.

Test used tumor of bladder image and derive from DingXiangYuan website (http://www.dxy.cn/cms), at Matlab7.1, windows XP, programming realizes under the 1G memory environment, after area-of-interest is selected, earlier area-of-interest is carried out lossless compress, contrast as shown in Figure 4 before and after the compression.

Shown in Fig. 4 (a), the compression back is shown in Fig. 4 (b) before the area-of-interest compression, and the pathological area that we select as can be seen from figure (tumor of bladder) is without any the distortion on the details.The area-of-interest lossless compress can be compressed to 70.13% of original image.The Wavelet Transformation Algorithm that experiment showed, lifting scheme reduces by half than Mallat algorithm computing time.

Non-area-of-interest is compressed, contrast as shown in Figure 5 before and after the compression again.

Shown in Fig. 5 (a), can control compression artefacts degree and compression ratio by the adjustment to the coefficient of connection β of PCNN internal activity item before the compression of non-area-of-interest, β obtained recovery effect after the compression shown in Fig. 5 (b) at 0.3 o'clock; β obtained recovery effect after the compression shown in Fig. 5 (c) at 3 o'clock.And can find that by contrasting us the high more details distortion factor of compression ratio is big more, but it is not a pathological area, can not influence doctor's diagnosis, its existence is a location positioning to pathological area.

Original image as shown in Figure 6, after the integer lifting wavelet lossless compress is restored as shown in Figure 7, the inventive method area-of-interest and non-area-of-interest merge eliminate boundary after, β be after restoring in 0.3 o'clock as shown in Figure 8, β be after restoring in 3 o'clock as shown in Figure 9.

The Y-PSNR PSNR of the whole bag of tricks is as shown in table 1:

The PSNR of table 1 the whole bag of tricks relatively

By relatively we can find do not have distortion fully after the integer lifting wavelet lossless compress is restored, signal to noise ratio is very high, but its compression ratio is bad, has only 70.13%.In the methods of the invention, though its signal to noise ratio is less than the lossless compress height, but only be that non-area-of-interest has distortion, at β is can obtain 37.56% compression ratio at 0.3 o'clock, β can reach 28.45% good compression rate at 3 o'clock, important area-of-interest does not have distortion, obtains the good compression rate with the distortion of unessential non-area-of-interest.

Claims (4)

1. the medical image ROI compression method based on Lifting Wavelet and PCNN is characterized in that comprising the steps:

Preliminary treatment

Adopt man-machine interactively formula difference shadow dividing method to carry out medical image segmentation, again with area-of-interest in the difference shadow method Medical Image Segmentation and non-area-of-interest and produce the two-value mask, view picture medical image and two-value mask being multiplied each other obtains the difference shadow again, thereby image is divided into area-of-interest and non-area-of-interest;

Lossy compression method

Non-area-of-interest is cut apart differentiation grey scale pixel value discrete point, the zones of different of dividing non-region of interest area image through PCNN; The close scope of gray value is set according to actual needs, carries out Run-Length Coding after the close pixel process igniting computing;

Lossless compress

Area-of-interest after adapting to conversion, integer 5/3 small echo is carried out the Ha Fuman coding through promoting;

Recover

Non-area-of-interest after the lossy compression method and the area-of-interest after the lossless compress are recovered medical image through merging, linear interpolation successively, realize medical image ROI compression.

2. the medical image ROI compression method based on Lifting Wavelet and PCNN according to claim 1, it is as follows to it is characterized in that described lifting integer 5/3 small echo adapts to transform method:

With low pass synthesis filter h (z), high pass synthesis filter g (z), lowpass analysis filter

With the high pass analysis filter

All resolve into even number and odd number two parts:

h(z)＝h _e(z ²)+z ^-1h _o(z ²)

g(z)＝g _e(z ²)+z ^-1g _o(z ²)

(1)

$h^{\%}\left(z\right)=h_e^{\%}\left(z^2\right)+z^{-1}{h}_o^{\%}\left(z^2\right)$

$g^{\%}\left(z\right)=g_e^{\%}\left(z^2\right)+z^{-1}{g}_o^{\%}\left(z^2\right)$

Wherein, subscript e idol coefficient multinomial, subscript o represents strange coefficient multinomial, promptly

$h_e\left(z\right)=\underset{k}{\mathrm{\&Sigma;}}{h}_{2k}{z}^{-k},$ $h_o\left(z\right)=\underset{k}{\mathrm{\&Sigma;}}{h}_{2k+1}{z}^{-k}---\left(2\right)$

In the formula, h _2k, h _2k+1Be respectively h _e(z) and h _o(z) press z ^-1The even number sequence number of launching, the coefficient of odd indexed.

And multinomial is defined as: $P\left(z\right)=\left[\begin{array}{cc}{h}_e\left(z\right)& g_e\left(z\right)\\ h_o\left(z\right)& g_o\left(z\right)\end{array}\right]---\left(3\right)$

Input signal carries out odd even too and decomposes, output low, and the z conversion of high fdrequency component represents to be respectively s ₁(z), d ₁(z), then

Decomposable process is:

$\left[\begin{array}{c}{s}_1\left(z\right)\\ d_1\left(z\right)\end{array}\right]=P^T\left(z\right)\left[\begin{array}{c}{x}_e\left(z\right)\\ z^{-1}{x}_o\left(z\right)\end{array}\right]---\left(4\right)$

(z) reaches Via Lifting Scheme below by decomposed P:

By $h_e\left(z^2\right)=\frac{h\left(z\right)+h(-z)}{2},$ $h_o\left(z^2\right)=\frac{h\left(z\right)-h(-z)}{2}z---\left(5\right)$

Push away $P{\left(z^2\right)}^T=\frac{1}{2}M\left(z\right)\left[\begin{array}{cc}1& z\\ 1& -z\end{array}\right]---\left(6\right)$

$P^{\%}{\left(z^{-1}\right)}^T=\left[\begin{array}{cc}{h}_e^{\%}\left(z^{-1}\right)& h_o^{\%}\left(z^{-1}\right)\\ g_e^{\%}\left(z^{-1}\right)& g_o^{\%}\left(z^{-1}\right)\end{array}\right]---\left(7\right)$

Then desirable reconstruction condition

When (h, g) constitute complementary filter to the time, detP (z)=1 (det represents determinant), after doing an antithesis and promoting, new polynomial matrix is:

$P^{\mathrm{new}}\left(z\right)=P\left(z\right)\left[\begin{array}{cc}1& 0\\ t\left(z\right)& 1\end{array}\right]---\left(8\right)$

By the Euclid algorithm, $\left[\begin{array}{c}{h}_e\left(z\right)\\ h_o\left(z\right)\end{array}\right]=\underset{i=1}{\overset{n}{\mathrm{\&Pi;}}}\left[\begin{array}{cc}{q}_i\left(z\right)& 1\\ 1& 0\end{array}\right]\left[\begin{array}{c}K\\ 0\end{array}\right]---\left(9\right)$

Therefore $P\left(z\right)=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}\left[\begin{array}{cc}1& 0\\ -s_i\left(z^{-1}\right)& 1\end{array}\right]\left[\begin{array}{cc}1& -t_i\left(z^{-1}\right)\\ 0& 1\end{array}\right]\left[\begin{array}{cc}{K}_1& 0\\ 0& K_2\end{array}\right]---\left(10\right)$

K ₁And K ₂Be constant, s _iBe the filter of prediction lifting step, t _iBe the filter that upgrades lifting step, promptly corresponding to the wavelet transformation of P (z) forward, the wavelet transformation of formation comprises: (Update) and 4 steps of normalization (Scaling) are upgraded in division (Split), prediction (Predict):

1. divide original signal sequence x (n) (n is a time series) is split into two mutually disjoint even number sequence number subclass x _e(n) and odd indexed subclass x _o(n), i.e. inertia (lazy) dividing method:

x(n)＝{x _e(n)，x _o(n)} (11)

2. predict according to the correlation between data, use x _e(n) dope x _o(n), concrete grammar is that the average with adjacent 2 even numbers is used as the odd number predicted value between them, goes to substitute this odd bits with the difference of odd bits and predicted value again, and expression formula is:

$x_o^{\%}\left(n\right)=x_o\left(n\right)-\frac{x_e(n-1)+x_e\left(n\right)}{2}---\left(12\right)$

3. upgrade and use x _o(n) and s _i(z) multiply each other and add x _e(n) realize:

Concrete method be with the predicted value of last position odd bits and back one digit pair digit value and 1/4th as the degree of deviation values of adjusting, remove alternative this even bit with the difference of this even bit and degree of deviation value again, expression formula is:

$x_e^{\%}\left(n\right)=x_e\left(n\right)-\frac{x_o^{\%}(n-1)+x_e(n+1)}{4}---\left(13\right)$

4. normalization realization coefficient is unified conversion, and promptly the even number with output partly multiply by scale factor K ₂, odd number partly multiply by scale factor K ₁

After repeating said process, obtain after n the decomposition

Represented the low frequency part of signal; Represented the HFS of signal.

3. the medical image ROI compression method based on Lifting Wavelet and PCNN according to claim 1 is characterized in that described PCNN dividing method is as follows:

The neuronal structure model of PCNN model is divided into three parts: the input area is input variable I _IjWith adjacent neurons output pulse Y _IjConnecting the input area is neuronic coupling part; Pulse produces district's passing threshold adjustment and produces pulse Y; Each discrete mathematical variable iterative equation of this model is as follows:

$F_{\mathrm{ij}}\left(n\right)=\exp (-{\mathrm{\&alpha;}}_F)F_{\mathrm{ij}}(n-1)+V_F\underset{\mathrm{kl}}{\mathrm{\&Sigma;}}{M}_{\mathrm{ijkl}}{Y}_{\mathrm{kl}}(n-1)+I_{\mathrm{ij}}---\left(14\right)$

$L_{\mathrm{ij}}\left(n\right)=\exp (-{\mathrm{\&alpha;}}_L)L_{\mathrm{ij}}(n-1)+V_L\underset{\mathrm{kl}}{\mathrm{\&Sigma;}}{W}_{\mathrm{ijkl}}{Y}_{\mathrm{kl}}(n-1)---\left(15\right)$

U _ij(n)＝F _ij(n)(1+βL _ij(n)) (16)

$Y_{\mathrm{ij}}\left(n\right)=\left\{\begin{array}{cc}1,& U_{\mathrm{ij}}\left(n\right)>E_{\mathrm{ij}}(n-1)\\ 0,& U_{\mathrm{ij}}\left(n\right)\&le;E_{\mathrm{ij}}(n-1)\end{array}\right.---\left(17\right)$

E _ij(n)＝exp(-α _E)E _ij(n-1)+V _E∑Y _kl(n-1) (18)

(14) I in the formula _IjBe that (i, the grey scale pixel value of j) locating are neuron and force the external drive excite, Y at point for the picture element matrix of image _KlBe that adjacent neuron is at point (k, the output of l) locating, F _IjBe the linear, additive result of input, M _IjklBe for feedback input domain mid point (i, j) and point (k, weight matrix l), α _FBe damping time constant, VF is the amplification coefficient in the feedback input domain; (15) L in the formula _IjBe connect the input area be of coupled connections α _LAnd V _LBe respectively its damping time constant and amplification coefficient, W _IjklBe point (i, j) and point (k, the weight matrix of connection matrix l); (16) U in the formula _IjBe the internal activity item, β is its coefficient of connection, E in (17) formula _IjBe the dynamic moving thresholding of excitation pulse, α in (18) formula _EAnd V _EBe respectively its damping time constant and amplification coefficient;

Passing threshold adjustment decision dynamic moving thresholding, the picture element matrix point close to the adjacent area gray-scale pixel values produces pulse, and is designated as 1, by Y _IjMatrix output, thus the grey scale pixel value discrete point distinguished, and the partitioned image zones of different realizes image segmentation.

4. the medical image ROI compression method based on Lifting Wavelet and PCNN according to claim 3 is characterized in that working as α _F＜α _L＜α _EThe time, each neuron all is in the square connection matrix center of a n * n among the PCNN, and general n gets 3, and it connects weights and gets W=M.

Images