CN106340000A - Bone age assessment method - Google Patents

Abstract

The invention discloses a bone age assessment method, which is characterized by comprising the steps of A, acquiring X-ray images of a plurality of target bone parts, and grading the acquired X-ray images according to the age group; B, performing feature extraction on a plurality of X-ray images in each grade, converting each grade of X-ray images into data by using a directional gradient histogram feature method, and building a mathematical model; and C, classifying the X-ray images in which the bone age is required to be assessed into the above grades by using a support vector machine classification algorithm. According to the bone age assessment method disclosed by the invention, an image processing technology and a computer vision technology are combined, and judgment for the bone maturity is realized through an interactive or automatic method.

Description

Bone age assessment method

Technical field

The present invention relates to bone age assessment method.

Background technology

Both at home and abroad it is believed that judging that developing individual maturity is a kind of effective method by the stone age, can be more comprehensive Ground reflection body physiological state, it represents the biological development age to a great extent.Skeleton development Maturity illustrates with certain One age was the level of the upgrowth and development of children of feature, was to evaluate individual growth to develop the most reliable index.

Currently, stone age researcher typically using shoot the single joint of body (as carpal joint) or each large joint of body (shoulder, elbow, Wrist, hip, SCID Mice etc.) method of x line piece carries out skeletal maturation degree evaluation.Sum up, being applied to the stone age at present pushes away Disconnected method has Atlas Method, scoring method, measurement method and mathematical model method etc..Traditional bone age assessment method is to be read by expert Piece result and standard diagram are compared or are carried out what tax point completed according to epiphysis development degree, but score value is subjective, It is unfavorable for the objective evaluation of stone age.And due to will being contrasted to the skeleton at multi-joint position or being scored, artificial treatment is taken Between longer, to scoring people Professional knowledge require higher.Therefore, some read tablet errors are more or less had by artificial read tablet, These errors can produce the impact of " bigger than normal " or " less than normal " to bone age assessment result.

Content of the invention

An object of the present invention is to overcome deficiency of the prior art, providing a kind of stone age reducing interference from human factor Appraisal procedure.

For realizing object above, the present invention is achieved through the following technical solutions:

Bone age assessment method is it is characterised in that include step:

A, the x light picture of the multiple targeted bone region of collection, and be classified according to age bracket；

B, feature extraction is carried out to the multiple x light pictures in every one-level；Utilization orientation histogram of gradients characterization method is by every one-level Described x light picture switchs to data, founding mathematical models；

C, using support vector cassification algorithm by need assess the stone age x light picture be divided in above rank.

Preferably, in described step a, targeted bone region includes shoulder joint, elbow joint, carpal joint, hip joint, knee joint The epiphysis position of one or more of joint, ankle joint and sternoclavicular joint.

Preferably, in described step a, targeted bone region includes one of ulna far-end epiphysis, distal radiuses epiphysis Or two.

Preferably, described image is divided into multiple cell cell；By multiple cell blocking block；Obtain 1764 Individual characteristic vector.

Preferably, in described step b,

First, gradient gx, gy, the gradient modular matrix g of image and the gradient direction square in image x direction and y direction are calculated Battle array θ then can be calculated according to following equation:

Then calculate the gradient orientation histogram of image in units of cell cell, computing formula is:

$dim(x,y)=cell\left(\frac{\left(\arctan \right(\frac{g_y(x,y)}{g_x(x,y)})+\frac{\mathrm{\π}}{2})\×190/\mathrm{\π}}{\mathrm{int}erval}\right)$

$weight\[i\]=weight\[i\]+\sqrt{g_x{(x,y)}^2+g_y{(x,y)}^2}$

(i=dim (x, y)).

Preferably, need are assessed by the vector in the x light picture of stone age using algorithm of support vector machine and substitute into formula y_i(w^tx_i+ b) -1, the w in each rank^t, b value be the numerical value of determination；When need assess the stone age x light picture in Amount makes y_i(w^tx_i+ b) -1=0 when, then need assess the stone age x light picture belong to this rank.

Preferably, choose the other x light image of known bone age grade, seek w in accordance with the previously described process^t, the value of b.

Preferably, gather the teenager skeleton x light image of 11.0～20.0 one full year of life.

After bone age assessment method according to claim 1 is it is characterised in that be classified the x light picture of stone age to be assessed, Accurately assess the stone age again.

Bone age assessment method in the present invention, image processing techniquess and computer vision technique are combined, by interaction or automatically Method achieves skeletal maturation degree decision method.Using computer carry out automated image analysis can reach accurately, the effect of high speed, The individual variation brought during read tablet can be overcome simultaneously, the interference of anthropic factor can be reduced, easier, quick, objective Ground assessment the stone age.With subject knowledges such as image procossing, image recognition and computer visions, develop prudence stone age mirror Determining computer system has very big using value and directive significance to actual appraisal.

Brief description

Fig. 1 is the original image of hand jnjuries x line piece；

Fig. 2 is the region of interest area image obtaining after the x line piece to skeleton intercepts；

Fig. 3 is the flow chart of teenager bone age assessment；

Fig. 4 is Optimal Separating Hyperplane schematic diagram；

Fig. 5 is that initial data uses kernel function to higher-dimension perspective view；

Specific embodiment

As shown in Figures 1 to 5, bone age assessment method is it is characterised in that include step:

A, the x light picture of the multiple targeted bone region of collection, and be classified according to age bracket；Targeted bone region include shoulder joint, The epiphysis position of one or more of elbow joint, carpal joint, hip joint, knee joint, ankle joint and sternoclavicular joint.

Collect 1897 ages between 11.0-20.0 one full year of life Han nationality man, female adolescent seven large joint totally 13279 x line pieces As research data.

Data acquisition is carried out to body 24 epiphysis of seven large joints, data is collected with x line sheet form and is converted into jpeg by scanning Image is stored in file system.

According to age bracket, sample is classified, both can be according to unified age grading, both all samples had been pressed every two years old is one-level Uniformly be classified, or uneven classification, such as 11-12 year be one-level, 12-14 is one-level.

Can also be classified according to different joints, such as elbow joint was uniformly classified for one-level by every two years old；Carpal joint is pressed every three years old Uniformly it is classified for one-level.

B, feature extraction is carried out to the multiple x light pictures in every one-level；Utilization orientation histogram of gradients characterization method is by every one-level Described x light picture switchs to data, founding mathematical models.

Described image is divided into multiple cell cell；By multiple cell blocking block；Obtain 1764 characteristic vectors.

First, gradient gx, gy, the gradient modular matrix g of image and the gradient direction matrix in image x direction and y direction are calculated

θ then can be calculated according to following equation:

Then calculate the gradient orientation histogram of image in units of cell cell, computing formula is:

$dim(x,y)=cell\left(\frac{\left(\arctan \right(\frac{g_y(x,y)}{g_x(x,y)})+\frac{\mathrm{\π}}{2})\×190/\mathrm{\π}}{\mathrm{int}erval}\right)$

$weight\[i\]=weight\[i\]+\sqrt{g_x{(x,y)}^2+g_y{(x,y)}^2}$

(i=dim (x, y)).

C, using support vector cassification algorithm by need assess the stone age x light picture be divided in above rank.

Need are assessed by the vector in the x light picture of stone age using algorithm of support vector machine and substitutes into formula y_i(w^tx_i+ b) -1, often W in individual rank^t, b value be the numerical value of determination；Assessing the vector in the x light picture of stone age when need makes y_i(w^tx_i+ b) -1=0 when, then need assess the stone age x light picture belong to this rank.

Choose the other x light image of known bone age grade, seek w in accordance with the previously described process^t, the value of b.

After the x light picture of stone age to be assessed is classified, more accurately assess the stone age.

A kind of method that Land use models technology of identification assesses China's Adolescents of Han Nationality skeletal age, using following steps:

A, to 140 (man, each 70 of women) 11.0～20.0 one full year of life Adolescents of Han Nationality chis, distal radiuses epiphysis application modes Technology of identification carries out image recognition；

B, choose five development stageses of chi, distal radiuses epiphysis respectively as research index, wherein chi, distal radiuses epiphysis Each classification all comprises 28 samples, builds chi, the pattern recognition classifier model of five development stageses of distal radiuses epiphysis；

C, the outside of the cross-validation and 35 separately checked sample mode identification models being carried out model using leaving-one method are tested Card, the accuracy rate (p of computation model identification respectively_a).

D, shot using x line camera system 1757 11.0～20.0 one full year of life Han nationality men, female adolescent body shoulder, elbow, wrist, Hip, SCID Mice and sternoclavicular joint totally seven large joint x line piece.

E, the x line piece to skeleton carry out image interception, image characteristics extraction, obtain image of interest region；

F, the image feature information to acquisition carry out dimension-reduction treatment；

G, using the characteristic information after dimensionality reduction, set up the osteoarticular disaggregated model of teenager；

H, the skeleton classification information of disaggregated model is substituted in multiple regression mathematical model and stone age standard of perfection collection of illustrative plates, forecast Teen-age skeletal age.

Details are as follows for concrete steps:

1st, x line piece sample collection

Collect east China, middle part and 1897 ages of southern areas between 11.0～20.0 one full year of life Han nationality men, female adolescent seven Totally 13279 x line pieces, as research data, carry out data acquisition to body 24 epiphysis of seven large joints, data is with x to large joint The collection of line sheet form is converted into jpeg image by scanning and is stored in file system.

2nd, image cropping and pretreatment

For setting up the accuracy of qualitative model, cutting area-of-interest is particularly important, using different cutting methods to qualitative classification After model foundation, the accuracy of model and robustness have strong influence.By changing area-of-interest sectional drawing scope, we taste Try the impact to modeling accuracy rate for the different sectional drawing modes, for the purpose of reducing noise as far as possible in the case of not loss information, will Its metaphysis divides intercepting as sample.Before carrying out the research of seven large joint stone ages, we are with teenager chi, distal radiuses epiphysis And as a example metaphysis, choose 140 samples as preliminary experiment, chi interested, distal radiuses region entirety are selected, it two It is part between dry epiphysis that the main epiphysis of person develops area-of-interest, and its epiphysis is integrally cut out as not losing its information as far as possible, Reduce noise jamming as far as possible.

3rd, image characteristics extraction is carried out to cutting and pretreated image

View data feature extraction is by the basis of image steganalysis, and the purpose of feature extraction algorithm is by the feature letter of image Cease to extracting, the characteristic information of image includes the color of image, texture, characteristic point (as especially bright point) etc..Direction Histogram of gradients (histogram of oriented gradient, hog) method is local shape factor algorithm, is extracted special by it There are after levying totally 1764 variables the feature of picture is described, image is converted to numeric form with graphic form.

When carrying out hog feature extraction, first image is divided into cell cell little one by one according to certain specification, Then calculate the gradient of entire image, we count the gradient orientation histogram of each cell cell afterwards.It is over institute in statistics After having the histogram of gradients of cell cell, with the block block of several continuous cell cell compositions for unit normalization Gradient orientation histogram, finally collects the gradient orientation histogram of block block with cell cell for step-length slip whole image, The gradient orientation histogram of all pieces of block is combined and can be formed by hog profiler.Hog feature pair The gradient of image local area has carried out being based on statistical description from intensity and direction.

Its concrete calculating process is as follows:

Image border is extracted using the method for canny rim detection and obtains image border matrix e.Then sobel operator is adopted to calculate Image x direction and gradient gx in y direction, gy, the gradient modular matrix g and gradient direction matrix θ of image then can according under Row formula is calculated.

$g=\sqrt{g_x^2+g_y^2}$

$\mathrm{\θ}=\arctan \frac{g_y}{g_x}$

By being calculated the information of image above, then substitute into equation below and calculate weight histogram of gradients.

3. calculate weight histogram of gradients

The gradient orientation histogram of whole image is counted in units of cell cell.Computing formula is as follows:

$dim(x,y)=cell\left(\frac{\left(\arctan \right(\frac{g_y(x,y)}{g_x(x,y)})+\frac{\mathrm{\π}}{2})\×190/\mathrm{\π}}{\mathrm{int}erval}\right)$

$weight\[i\]=weight\[i\]+\sqrt{g_x{(x,y)}^2+g_y{(x,y)}^2}$

(i=dim (x, y))

Dim (x, y) represents the classification that the gradient direction of pixel (x, y) should belong in rectangular histogram, and this classification is by gradient direction The group such as angle divide away from (interval) and to obtain, the scope of angle is 0 degree of classification to 180 degree.weight[i] Represent the gradient weight that classification is i.With block block for unit normalized gradient direction histogram, then the hog calculating entire image Feature descriptor.Count the gradient orientation histogram of whole image in units of cell cell, obtain the characteristic information of image, The characteristic variable modeling for back as gray value, characteristic information.

4th, build qualitative classification model prediction epiphysis development stages

By using algorithm for pattern recognition, great amount of samples data is carried out with the foundation of mathematical model, thus reaching when acquisition unknown sample When unknown sample can be carried out forecast detection.

The algorithm for pattern recognition that the present invention adopts is support vector cassification algorithm (support vector classification, svc). Svc starts with from the situation of the simplest linear separability first.

During linear separability, in d dimension space, the general type of linear discriminant function is g (x)=w^tX+b, (note: herein w with X is vector；, classifying face equation is w^tx+b；Discriminant function is normalized by we, so that all samples of two classes is all met |g(x)|≥1；Now from | the g (x) |=1 of the nearest sample of classifying face；And require classifying face that all samples can correctly be classified, It is exactly required to meet: y_i(w^tx_i+ b) -1 >=0, i=1,2 ..., n ...；N is that sample number " makes those that equal sign is set up in above formula Sample is called supporting vector (support vectors).

Cause is exactly it as the sample point nearest from classifying face, and they finally establish significant role, so being claimed to model For supporting vector " so, gap size in the classification space (margin) of two class samples:

$m\arg in=\frac{2}{\left|\right|w\left|\right||}$

Therefore, the constrained optimization problems that optimal classification surface problem can be expressed as, that is, under the constraint of conditional, find a function:

$\mathrm{\φ}\left(w\right)=\frac{1}{2}\left|\right|w|{|}^2\left(w^{t}w\right)$ Minima.

Simultaneously it can also be seen that supporting vector is necessarily in hyperplane (w^tx_i+ b)=1 or (w^tx_i+ b)=- 1 on.

For this reason, the lagrange function that can be defined as follows:

$l(w,b,\mathrm{\α})=\frac{1}{2}\left(w^{t}w\right)-\underset{i=1}{\overset{n}{\σ}}{\mathrm{\α}}_i\[y_i(w^t{x}_i+b)-1\]$

Wherein, α_iFor lagrange coefficient or multiplier.

Based on the research of above-mentioned qualitative classification model, we are directed to China Adolescents of Han Nationality man, womens body 24 bones of seven large joints Epiphysis development stages carries out qualitative classification model and explores and by different method validation result accuracys rate, be shown in Table 1.

Table 1 teenager man, 24 epiphysis development stages modeling results of womens body seven large joint

5th, multiple regression mathematical model and the Skeletal Age Standards collection of illustrative plates assessment stone age

Intercepted by Image semantic classification, image characteristics extraction, after support vector machine qualitative classification modeling, you can obtain seven large joints Gained classification can be passed through multiple regression mathematical model and stone age standard of perfection collection of illustrative plates to sample to be tested by the classification of 24 indexs Stone age is estimated.

Embodiment in the present invention is only used for that the present invention will be described, does not constitute the restriction to right, this area Other substantially equivalent replacements that interior technical staff is contemplated that, all in the scope of the present invention.

Claims (9)

1. bone age assessment method is it is characterised in that include step:

A, the x light picture of the multiple targeted bone region of collection, and be classified according to age bracket；

B, feature extraction is carried out to the multiple x light pictures in every one-level；Utilization orientation histogram of gradients characterization method is by every one-level Described x light picture switchs to data, founding mathematical models；

C, using support vector cassification algorithm by need assess the stone age x light picture be divided in above rank.

2. bone age assessment method according to claim 1 is it is characterised in that in described step a, targeted bone region includes The osteoepiphyseal portion of one or more of shoulder joint, elbow joint, carpal joint, hip joint, knee joint, ankle joint and sternoclavicular joint Position.

3. bone age assessment method according to claim 1 is it is characterised in that in described step a, targeted bone region includes One of ulna far-end epiphysis, distal radiuses epiphysis or two.

4. bone age assessment method according to claim 1 is it is characterised in that be divided into multiple cell cell by described image； By multiple cell blocking block；Obtain 1764 characteristic vectors.

5. bone age assessment method according to claim 1 is it is characterised in that in described step b,

First, gradient gx, gy, the gradient modular matrix g of image and the gradient direction matrix in image x direction and y direction are calculated θ then can be calculated according to following equation: $g=\sqrt{g_x^2+g_y^2};\mathrm{\θ}=\arctan \frac{g_y}{g_x};$

Then calculate the gradient orientation histogram of image in units of cell cell, computing formula is:

$dim(x,y)=cell\left(\frac{\left(\arctan \right(\frac{g_y(x,y)}{g_x(x,y)})+\frac{\mathrm{\π}}{2})\×190/\mathrm{\π}}{\mathrm{int}erval}\right)$

$weight\[i\]=weight\[i\]+\sqrt{g_x{(x,y)}^2+g_y{(x,y)}^2}$

(i=dim (x, y)).

6. bone age assessment method according to claim 5 is it is characterised in that assess the stone age using algorithm of support vector machine by need Vector in x light picture substitutes into formula y_i(w^tx_i+ b) -1, the w in each rank^t, b value be the numerical value of determination；When The vector that need to assess in the x light picture of stone age makes y_i(w^tx_i+ b) -1=0 when, then need assess the stone age x light picture belong to This rank.

7. bone age assessment method according to claim 6 it is characterised in that choose the other x light image of known bone age grade, according to Step described in claim 5,6 seeks w^t, the value of b.

8. bone age assessment method according to claim 1 is it is characterised in that gather 11.0～20.0 one full year of life teenager skeleton x Light image.

9. after bone age assessment method according to claim 1 is it is characterised in that be classified the x light picture of stone age to be assessed, then The accurately assessment stone age.