CN100349548C - Intelligence enhanced system and method for early testing idiopathetic scoliosis - Google Patents

Abstract

The present invention relates to an intelligence enhanced system used for early detecting idiopathic scoliosis by using chest radiation imaging. The intelligence enhanced system comprises an image processing module, a vectoring module, a cost function calculating module and a judging module, wherein the image processing module is used for obtaining chest image raster data; the vectoring module is used for vectoring a spine in an image and obtaining the estimating centre point of the spine; the cost function calculating module is used for calculating the degree of a vector fitting straight line of the spine of the image and obtaining a mean square error; the judging module is used for judging the magnitude of the mean square error and the threshold value R, so that the image is classified into normal spines and scoliosis. The present invention also relates to a method for early detecting idiopathic scoliosis by using the intelligence enhanced system.

Description

A kind of intelligence enhanced system of early testing idiopathetic scoliosis

Technical field

The present invention relates to a kind of intelligence enhanced system and method thereof of utilizing chest radiophotography early testing idiopathetic scoliosis.

Background technology

Along with the appearance of medical image management with communication system (PACS), Ji Lei a large amount of digital image databases and numerical map have more been given prominence to the problem of obtaining picture material gradually.In traditional medical image databases, most of index and search method all based on text message as research identity, patient's name, date etc.But in a lot of image data bases, information itself is not letter and number, but a large amount of characteristics of image such as color, structure and shape facility, and these features all belong to picture material and can not find based on traditional data base's search method.

In fact, in a lot of researchs, clinical and teaching practice, if view data can retrieve based on picture material, thereby then more information can be utilized and promotes the retrospective of some methodological image and prospect Journal of Sex Research.If well developing, the searching system of these picture materials will help clinical doctor and educator and researcher to find image with similar characteristics by the image that belongs to close classification or similar features in automatic coupling institute's query image and the PACS image data base.

Medical image is the difficult problem that research worker is generally acknowledged as the classification of chest imaged image, because these medical images are extremely similar, they only have difference usually on some small details.Present carry out content-based index technology based on specific image features such as structure, color, shape etc., it is accurate not enough to be applied on the medical image.

Since 2000, ophthalmology optometry and radiological science have been independent research a kind of Image Management and communication system.Since calendar year 2001, this PACS is applied in clinically the radiophotography clinic.From then on, this image data base has obtained fast development.Existing directory system among the PACS is that the operation for daily system designs.Yet when our focus was concrete characteristics of image or shape, this directory system was difficult to satisfy research and teaching purpose.

In Hong Kong, student's scoliotic sickness rate is about 2.7%.In radiophotography is clinical, also detect the special skoliosis of sending out by undergraduate chest X ray.As if the special detection of sending out deformation of spinal column be the problem that the healthy doctor of plastic surgeon and community is concerned about.In most of the cases, scoliotic image can just not told when going to a doctor for the first time.This will make teen-age scoliotic subsequently research and retrospective study complicated.

Existing PACS basic structure and scoliotic existing case have promoted the exploitation of intelligent Image Retrieval and categorizing system.The exploitation of this system helps community medicine, because still lack a kind of Image Retrieval system that has integrated PACS preferably at present.Therefore, a kind of brand-new Image Retrieval system based on image matching algorithm and computer intelligence method has shown the good prospect of exploitation.

Summary of the invention

The object of the present invention is to provide a kind of intelligence enhanced system that utilizes chest radiophotography early testing idiopathetic scoliosis, this intelligence enhanced system comprises:

A. image processing module is in order to obtain chest image raster data;

B. the vectorization module in order to the spinal column vectorization in the image, obtains spinal column and estimates central point;

C. the cost function calculation module in order to the vectorial collinear degree of coincideing of the spinal column of computed image, obtains mean square error;

D. determination module in order to judging the size of mean square error and threshold value R, thereby is normal spinal column and skoliosis with image classification.

Wherein, the vectorization module can comprise correlation technique unit, blur estimation unit and selected cell, and described selected cell is positioned at before correlation technique unit and the blur estimation unit, in order to select to carry out correlation technique or blur estimation.

The present invention also provides and has utilized above-mentioned intelligence enhanced system to carry out the method for early testing idiopathetic scoliosis, and this method comprises the steps:

(1) handles image, obtain the raster data of chest image;

(2), obtain spinal column and estimate central point with the spinal column vectorization of image;

(3) the calculation cost function obtains mean square error;

(4) with image classification: image is divided into normal spinal column and skoliosis by the size that compares mean square error and threshold value R.

Wherein, the vectorization of step (2) can be adopted correlation technique and blur estimation, and adopts before correlation technique and blur estimation and select step, and this step is to adopt correlation technique or blur estimation in order to judgement.

Intelligence enhanced system of the present invention is to be used to detect the scoliotic morbidity and the order of severity, and it is according to digitized chest irradiation image, and this irradiation image is based on fuzzy logic [1,2,3,5], neutral net [4,6] and method of least square [7].Unlike adopting moving shape model (Active ShapeModeling (ASM)) method, the profile of spinal column is studied bit by bit, when adopting system of the present invention, only represent spinal column, find the spinal column that every straight line mated of chest irradiation image to estimate central point with a vertical line.This method is not only fast and convenient but also be easy to locate automatically spinal column in the chest irradiation image.Defined cost function in addition, so that quantize estimated spinal column site of skoliosis index assessment and the deviation between the vertical line.All cases all are considered to normally at first, and cost function all converges to zero.During the training mode of system, import data by the expert and obtain error level R value.During operational mode, when the cost functional value was lower than error level, case was classified as normal spinal column; On the contrary, when the cost functional value was higher than error level, case was classified as undesired class, and promptly skoliosis is seen Fig. 1 intelligence system work sketch plan.The enforcement of system comprises two computational processes: (a) vectorization of spinal column central point; (b) calculating of cost function.

Description of drawings

Fig. 1 is an intelligence system work sketch plan;

Fig. 2 is the flow chart of Intelligent Composite method;

Fig. 3 represents that template is the pixel value curve chart of a line segment in the chest irradiation image,

3a is selected line segment, and 3b is the curve chart of this line segment correspondence;

Fig. 4 is the structure of traditional fuzzy neuron network;

Fig. 5 is the basic function of expression single order B-batten (B-Spline) function;

Fig. 6 is the operation full figure of intelligent method.

The specific embodiment

Below in conjunction with the drawings and specific embodiments the present invention is described in further detail.

(a) vectorization of spinal column central point

When detecting scoliotic morbidity and estimating the scoliotic order of severity, do not need to describe the profile of spine.Find the spinal column central point of every line correspondence of digital chest irradiation image just to be enough to satisfy the requirement of implementing.But the chest irradiation image is with raster format organ and tissue to be carried out superposition to form, and the shape and the position of the organ or tissue that exists with vector format are not provided itself.The present invention utilizes a kind of Intelligent Composite method, and the spinal column shape of chest image is carried out vectorization, and this Intelligent Composite method is convenient to change correlation technique into blur estimation based on predefined standard.Before introducing this method, the symbol in the table 1 once need be described earlier.

The definition of table 1 symbol

Symbol	Definition
		H	Picture altitude
W	Picture traverse
		j	Row number (1 on the top, and h is in the bottom)
x c j	Utilize the estimation central point of the capable spinal column of j that correlation technique obtains
		x f j	Utilize the estimation central point of the capable spinal column of j that blur estimation obtains
x j	The final estimation central point of the capable spinal column of j
		C j	The maximum correlation coefficient that j is capable

Fig. 2 is the flow chart of Intelligent Composite method.This method has been handled in the chest image each the bar line that removes bottom margin, this with bottom margin be defined as be positioned at the image bottom, highly be the part of total height 1/10th.Second decision box of flow chart included this bottom margin in this method in.This edge comprises the vertebral region that hides liver, whether belongs to skoliosis and does not determine according to this part spinal column.In first decision box, by judging which should be final estimation spinal column central point to three conditions with the dateout of determining to come autocorrelation method or blur estimation.First condition is the reliability that is used for confirming correlation technique, i.e. C _j＞0.94.Second condition is the seriality that is used to guarantee spinal column, promptly | and X ^c _j-X _J-1|＜0.0284*W.The 3rd condition is j＞0.341*H, and when estimating upward vertical diaphragm place spinal column, do not meet this condition this moment, then uses blur estimation, because blur estimation effect in this zone is fine.And when estimating down vertical diaphragm place spinal column, this moment, condition was set up, because the stack of heart shadow (X, large artery trunks and spinal column needs more complicated blur estimation, and computation time is also longer, is more suitable for utilizing at this moment correlation technique.0.94,0.0284 and 0.341 be the parameter that in the experiment test of this method, draws by counting carefully of some.

(i) correlation technique:

Obtain the spinal level profile board at vertical diaphragm place down by sample image.This template is a series of n pixel numbers, P _T=[P (1), P (2) ..., P (n)], as shown in Figure 3.

If image j is capable and the pixel value of i row be P (i, j).To each bar line at vertical diaphragm place down (as j＞0.341*H), consider the contiguous pixels window, that is,

P _w(x，j)＝[P(x+1，j)，P(x+2，j)，...，P(x+n，j)] (1)

Wherein, x=0,1 ..., W-n.(x j) is defined as correlation coefficient C

Wherein, (a b) is the covariance of vectorial a and b to cov, and var (a) is the variance of vectorial a.(x, j) maximum is the C that j is capable

C _j＝max C(x，j)，x＝0，1，...，W-n (3)

Therefore, the spinal column central point of estimation is

x ^c _j＝x _max+n/2 (4)

X wherein _MaxBe the meansigma methods of all x, thus C (x, j)=C _jLike this, this correlation technique draws two numerical value (x for every line of image ^c _j, C _j), as shown in table 1.

(ii) blur estimation

Different organs, tissue or material according to its separately density and thickness exposure has in various degree decay to X ray.Skeleton, soft tissue and air should be represented by the pixel value of different range in the irradiation image.We can regard pixel data as and only represent a monochrome image plane, and wherein sample value is intended to show after carrying out volume of interest district (VOI) gray scale conversion and is black, and promptly the photometric interpretation in the dicom standard equals ' monochromatic 2 '.Skeleton, soft tissue and air should be respectively with the gray-scale pixels value representation of high, medium and low scope.Yet the central point that finds spinal column is not so directly, because this point not necessarily has max pixel value in high scope.Because the fuzzy neuron network is a kind of model of mind, it comprises location (localized) basic function and training weight [1,2], and the pattern of spine image can be incorporated in this model at an easy rate.Blur estimation based on the fuzzy neuron network structure that is widely used in system identification can be used to seek the likelihood of the pixel that becomes the spinal column central point, estimate based on these likelihoods then.

The structure of tradition fuzzy neuron network is shown among Fig. 4.Suppose that the fuzzy neuron network has single output, the input space is described by grid, and the center of basic function is on the each point of grid.B-batten (B-Spline) function is usually as the basic function [1-6] of fuzzy neuron network.

The output of fuzzy neuron network is calculated as follows:

$y\left(t\right)=Q\left(x\left(t\right)\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\θ}}_i^{,}{\mathrm{\μ}}_i\left(x\left(t\right)\right)---\left(5\right)$

Wherein, μ _iAnd θ ' _iBe respectively i basic function and i weight.Data can by N to input-output to the expression, as [x (1), y (1)], [x (2), y (2)] ..., [x (N), y (N)].The mapping relations of input-output are expressed as y=Q (x).For m many output, mapping relations are:

$Q_k\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\θ}}_{\mathrm{ik}}^{,}{\mathrm{\μ}}_{\mathrm{ik}}\left(x\right)$ K=1 wherein, 2 ..., m (6)

In order to estimate the central point of spinal column, will import x and be defined as a variable, its representative is exported Q along the horizontal level of a line on the chest image _k(x) the expression conduct is based on the likelihood of the horizontal level (k+n/2) of the spinal column central point of input x.Suppose that n is an even number, m=W-n.Weight θ ' _IkFor coordinate (k-n/2+i-1, the j) pixel value on, promptly P (k-n/2+i-1, j), basic function μ _IkBe single order B-batten (B-Spline) function:

Fig. 5 is the diagram of basic function.

As can be seen from Figure 5, this basic function has a key property,

$\underset{0}{\overset{W}{\∫}}{\mathrm{\μ}}_{\mathrm{ik}}\left(x\right)\mathrm{dx}=1---\left(8\right)$

By along image j bar line to Q _k(x) carry out integration, and with in (8) substitution integration, (k-n/2+i-1, total likelihood L j) _KjFor,

$L_{\mathrm{kj}}=\underset{0}{\overset{W}{\∫}}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}P(k-\frac{n}{2}+i-1,j){\mathrm{\μ}}_{\mathrm{ik}}\left(x\right)\mathrm{dx}$

$=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}P(k-\frac{n}{2}+i-1,j)\underset{0}{\overset{W}{\∫}}{\mathrm{\μ}}_{\mathrm{ik}}\left(x\right)\mathrm{dx}$

$=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}P(k-\frac{n}{2}+i-1,j)$ K=1 wherein, 2 ..., W-n (9)

Note,, can utilize high-order B-batten (B-Spline) function according to the rare degree of feature to be detected.Yet if take high-order B-batten (B-Spline) function, total likelihood will be more complicated, and therefore, computation time is longer.Because the pixel value of spinal column is in the high scope of gray scale, blur estimation is output as the meansigma methods of horizontal level k, and wherein total likelihood is a maximum, promptly

$x_j^f=\mathrm{ave}\left(\underset{k}{\arg \max }{L}_{\mathrm{jk}}\right)$

Note, total for the operator of selecting likelihood, as select for use here ' arg max ' wants the feature that detects and decides according to us, also can change with the different target object.

Therefore, the vector (u (j), v (j)) of expression spinal column j bar line central point is (x ^f _j, j) or (x ^c _j, j), the latter looks the C that correlation technique obtains _jThe value and decide.

(b) calculating of cost function

Scoliotic morbidity can be determined afterwards at the vector (u (j), v (j)) of the every line that obtains the chest image.Determine by reduction to absurdity, suppose that wherein spinal column is normal spinal column, spinal column is for being close to straight line.Under this hypothesis, vector (u (j), v (j)) can meet the straight line of following formula:

u＝m·v+c (11)

Following formula provides zero model error:

e(j)＝u(j)-m·v(j)-c≈0 (12)

Yet data inevitably comprise noise content to a certain degree, therefore, have adopted method of least square [7] to obtain the most identical straight line here.If the chest image has N bar line.According to formula (12), model error is:

$\mathrm{\ϵ}=\left[\begin{array}{c}e\left(1\right)\\ e\left(2\right)\\ \·\\ \·\\ \·\\ e\left(N\right)\end{array}\right]=\left[\begin{array}{c}u\left(1\right)\\ u\left(2\right)\\ \·\\ \·\\ \·\\ u\left(N\right)\end{array}\right]-\left[\begin{array}{cc}v\left(1\right)& 1\\ v\left(2\right)& 1\\ \·& \·\\ \·& \·\\ \·& \·\\ v\left(N\right)& 1\end{array}\right]\left[\begin{array}{c}m\\ c\end{array}\right]=U-\mathrm{Vw}---\left(13\right)$

Wherein, ε=[e (1), e (2) ..., e (N)] ^T, U=[u (1), u (2) ..., u (N)] ^T, V=[v (1), v (2) ..., v (N)] ^T, and w=[m, c] ^TAccording to method of least square, the best estimate of w is:

$\hat{w}={\left(V^{T}V\right)}^{-1}{V}^{T}U---\left(14\right)$

Can draw from (13) and (14), the corresponding model error is:

$\widehat{\mathrm{\ϵ}}=U-V\hat{w}$

For the identical collinear degree of evaluating data, adopt mean square error (M.S.E):

$M.S.E=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\hat{e}{\left(j\right)}^2=\frac{1}{N}{\widehat{\mathrm{\ϵ}}}^T\widehat{\mathrm{\ϵ}}---\left(16\right)$

Therefore, by cost function the skoliosis exponential quantity is turned to mean square error.

M.S.E. value can be divided into S and R two parts, i.e. M.S.E.=S+R.R represents the acceptable degree of skoliosis and noise, and S represents scoliotic significance degree.The R value is that error level is determined by the expert.Definite method of R value is as follows: utilize interactive user interface to receive the data of radiologist's input.This intelligence system is in training mode at this moment, and the error level that system will present after set of chest irradiation image and the corresponding calculated thereof is given the radiology expert.Radiology expert then estimates this case and whether belongs to skoliosis, and therefore the radiology expert estimates this system that trains with it in definite R value.The radiologist utilizes the training subclass of existing chest irradiation image to determine the R value.For example, if a patient does not suffer from skoliosis, the R value that obtains from system is 70, and the radiologist chooses " no " in user interface, and then R=70 will be considered to normal spinal column.Perhaps, if a chest films showed skoliosis, this moment R=90, the expert determines to exist skoliosis, and chooses " yes ", therefore, when R=90, rabat is classified as skoliosis.Then, the meansigma methods and the standard deviation of R value calculate according to a series of cases in system, thereby obtain the R value, and this moment, the R value was a meansigma methods.Threshold value R, promptly meansigma methods R is exactly an indicator, needs only the image greater than this value, is classified as skoliosis.

If M.S.E.＜R supposes that effectively then spinal column is normal.If M.S.E.＞R, spinal column is kept off straight line, and according to reduction to absurdity, spinal column is then unusual.The whole service of this intelligence system as shown in Figure 6.

The feature of intelligence system of the present invention has:

1. the earlier detection spy sends out skoliosis

2. monitor the protruding degree of scoliotic state of development and side

3. based on X-ray chest image, detect the scoliotic order of severity, need not further whole spinal column to be carried out the X-irradiation, reduced radiation dose

4. do not have human intervention, from image data base, extract possible skoliosis image automatically

The range of application of intelligence system of the present invention:

Use native system and be in order to form a kind of picture material indexing means, thereby realize with retrieving images from PACS:

1. develop content-based intelligent image search method

2. provide reference example to help clinical diagnosis based on characteristics of image

3. the view data research for graphical analysis provides a kind of method

Application Example

Select 100 routine cases in the PACS database from the clinical roentgenology of The Hong Kong Polytechnic University, wherein half is normal, and second half is a skoliosis.Intelligence enhanced system can detect exactly automatically has scoliotic case.Result and Cobb ' s angular dependence (-dance), Cobb ' s angle are a kind of indexs of representing the skoliosis flexibility, and it gets by directly measuring the irradiation image rabat, as what the anatomist measured.

Intelligence enhanced system can directly be read the DICOM image from the PACS image data base.The chest image can be sent on the application server, screens by our system, and perhaps the chest image in the image server can directly be read by our system, thereby detects skoliosis.

In the device of The Hong Kong Polytechnic University, the chest image is sent on the application server in the intelligence enhanced system, thereby detects whether rabat is skoliosis.

Claims (2)

1. the intelligence enhanced system of an early testing idiopathetic scoliosis is characterized in that, described system comprises:

A. image processing module is in order to obtain chest image raster data;

B. vectorization module is in order to turn to (X with the spinal column vector in the image _j, j), obtain each row spinal column and estimate central point, wherein said X _jBe the estimation central point of the capable spinal column of j, j represents the row number of spinal column;

C. the cost function calculation module in order to the vectorial collinear degree of coincideing of the spinal column of computed image, obtains mean square error;

D. determination module in order to judging the size of mean square error and threshold value R, thereby is normal spinal column and skoliosis with image classification.

2. intelligence enhanced system according to claim 1, it is characterized in that, described vectorization module comprises correlation technique unit, blur estimation unit and selected cell, described selected cell is positioned at before correlation technique unit and the blur estimation unit, in order to select to carry out correlation technique or blur estimation.

Images