CN105389813B - The recognition methods of organ and dividing method in medical image - Google Patents
The recognition methods of organ and dividing method in medical image Download PDFInfo
- Publication number
- CN105389813B CN105389813B CN201510729150.0A CN201510729150A CN105389813B CN 105389813 B CN105389813 B CN 105389813B CN 201510729150 A CN201510729150 A CN 201510729150A CN 105389813 B CN105389813 B CN 105389813B
- Authority
- CN
- China
- Prior art keywords
- organ
- image
- target organ
- medical image
- detection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 210000000056 organs Anatomy 0.000 title claims abstract description 133
- 238000001514 detection method Methods 0.000 claims abstract description 72
- 230000004927 fusion Effects 0.000 claims abstract description 7
- 230000000717 retained Effects 0.000 claims abstract description 7
- 238000000034 method Methods 0.000 claims description 21
- 238000005070 sampling Methods 0.000 claims description 19
- 230000011218 segmentation Effects 0.000 claims description 14
- 230000000875 corresponding Effects 0.000 claims description 9
- 238000007689 inspection Methods 0.000 claims description 3
- 230000003044 adaptive Effects 0.000 abstract description 5
- 210000003734 Kidney Anatomy 0.000 description 24
- 210000003128 Head Anatomy 0.000 description 15
- 210000000038 chest Anatomy 0.000 description 12
- 210000001015 Abdomen Anatomy 0.000 description 10
- 238000010586 diagram Methods 0.000 description 6
- 210000001519 tissues Anatomy 0.000 description 6
- 238000005194 fractionation Methods 0.000 description 5
- 238000010801 machine learning Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 4
- 238000002059 diagnostic imaging Methods 0.000 description 4
- 238000000638 solvent extraction Methods 0.000 description 4
- 210000002391 Femur Head Anatomy 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000003384 imaging method Methods 0.000 description 3
- 230000003902 lesions Effects 0.000 description 3
- 238000001356 surgical procedure Methods 0.000 description 3
- 241001270131 Agaricus moelleri Species 0.000 description 2
- 238000010521 absorption reaction Methods 0.000 description 2
- 238000002591 computed tomography Methods 0.000 description 2
- 239000003814 drug Substances 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000003709 image segmentation Methods 0.000 description 2
- 239000004615 ingredient Substances 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000002600 positron emission tomography Methods 0.000 description 2
- 210000000709 Aorta Anatomy 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- 210000004185 Liver Anatomy 0.000 description 1
- 210000000214 Mouth Anatomy 0.000 description 1
- 210000004197 Pelvis Anatomy 0.000 description 1
- 210000003491 Skin Anatomy 0.000 description 1
- 210000000952 Spleen Anatomy 0.000 description 1
- 210000002784 Stomach Anatomy 0.000 description 1
- 210000001138 Tears Anatomy 0.000 description 1
- 210000003462 Veins Anatomy 0.000 description 1
- 230000003187 abdominal Effects 0.000 description 1
- 230000001058 adult Effects 0.000 description 1
- 230000001174 ascending Effects 0.000 description 1
- 210000000746 body regions Anatomy 0.000 description 1
- 239000002775 capsule Substances 0.000 description 1
- 238000003759 clinical diagnosis Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000004195 computer-aided diagnosis Methods 0.000 description 1
- 239000000470 constituent Substances 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 229940079593 drugs Drugs 0.000 description 1
- 239000000686 essence Substances 0.000 description 1
- 230000014509 gene expression Effects 0.000 description 1
- 238000002675 image-guided surgery Methods 0.000 description 1
- 238000007912 intraperitoneal administration Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000006011 modification reaction Methods 0.000 description 1
- 244000144985 peep Species 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 230000003595 spectral Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000003325 tomography Methods 0.000 description 1
- 210000000689 upper leg Anatomy 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 230000003313 weakening Effects 0.000 description 1
Abstract
The invention discloses a kind of recognition methods of organ in medical image, including：Pending medical image is obtained, the medical image is split into several two dimensional images in X, Y and Z-direction respectively, and detection window is set according to the size of target organ；Traversal detection is carried out to the two dimensional image respectively according to the detection step-length of setting using the detection window, obtains the testing result in X, Y and Z-direction；The testing result is subjected to result fusion, all pixels of test positive are retained on X, Y and three directions of Z axis, so that it is determined that the target organ boundary.In medical image of the present invention the recognition methods of organ can accurately, quickly identify target organ element it is even even region, determine target organ boundary, and adaptive ability is strong.In addition, the present invention also provides a kind of dividing methods of organ in medical image.
Description
【Technical field】
The present invention relates to field of medical image processing, the recognition methods more particularly to organ in medical image and segmentation side
Method.
【Background technology】
It, can be with the increasingly mature and various medical imaging device extensive use within the hospital of Medical Imaging Technology
It is convenient nondestructively to get inside of human body organizational information image, these information are effectively treated by image processing techniques,
Diagnosis even surgery planning etc. for assisting doctor has great social benefit and is widely applied foreground.Such as computer
Tomoscan (Computed Tomography, CT) image is by certain amount by the black pixel arrangement group to white different gray scales
At matrix, the X-ray absorption coefficient of corresponding voxel can be reflected by pixel, and different gray scales then reflects organ or tissue
To the degree of absorption of X-ray.It is emerging to can extract out sense by carrying out correct, rational segmentation to CT images in the image procossing later stage
Interesting organ, tissue or lesion body, and then can realize the three-dimensional visualization of the organ being extracted to these, tissue or lesion, with
Achieve the purpose that auxiliary treatment and surgery planning.However, for complicated human organ, as abdomen is (intraperitoneal to include liver, courage
Capsule, spleen, stomach, inferior caval vein and aorta etc.), the segmentation for executing each organ takes at 3 minutes or more, if indefinite
In the case of the corresponding body part (head, chest, abdomen and pelvic cavity) of currently processed CT images, all organs are blindly called
Partitioning algorithm then needs to consume a large amount of time.In addition, limitation and tissue wriggling of the abdominal CT images due to imaging device,
Some artifacts and noise are will produce, causing partial organ to organize, fuzzy, lesion body edge is indefinite, these all bring to segmentation
Sizable difficulty.Therefore, it calls and needs to prejudge the position at currently processed CT images essential body position before partitioning algorithm
It sets.
By doctor's manual identified and the body part for marking current CT images to be located at, doctor is needed largely to be repeated
Work, it is less efficient.And existing CT images body part automatic identifying method can be mainly divided into three kinds：(1) it is based on medicine
Digital imagery and communication (Digital Imaging and Communications in Medicine, DICOM) file header letter
The body part identification [1] of breath, usual DICOM file head include the label information of CT image scannings, but due to various cultural languages
Say that difference, the label of different language record can increase the difficulty for accurately identifying DICOM file header, wronger DICOM
File header information can cause to identify mistake；(2) method based on grey value characteristics, mainly according to X-ray by body not
Different with attenuation degree when structural constituent, grey value profile of the body different tissues ingredient in CT images is different, based on ash
The method of angle value feature is according to the priori of body Main Tissues ingredient grey value profile rule in CT images, to body
Position is divided, however this method is relatively low [2] to head, pelvic cavity discrimination；(3) method based on machine learning, this method
It is broadly divided into training and two stages of test, the corresponding Haar characteristics of image of body part critical organ is extracted in the training stage,
The a large amount of positive negative samples of structure extract organ effective Haar characteristic sequences and its phase accordingly by training AdaBoost graders
Weight is answered, in test phase, inputs image to be measured, calculates the Haar characteristic values of the image, by itself and existing training result
Compared, judge whether the image is positive sample [3], this method needed when extracting image Haar features to characteristic window into
Row repeatedly up-sampling or down-sampling, and the Haar feature quantities used are more, exist and compute repeatedly and calculation amount is larger, fortune
Calculate the low problem of efficiency.Based on this, it is necessary to be improved to the recognition methods of organ in existing medical image.
[1] Gueld M O, Kohnen M, Keysers D, et al.Quality of DICOM header
information for image categorization[C].Medical Imaging.International Society
For Optics and Photonics, 2002：280-287.
[2] Dicken V, Lindow B, Bornemann L, et al.Rapid image recognition of
body parts scanned in computed tomography datasets[J].International journal
Of computer assisted radiology and surgery, 2010,5 (5)：527-535.
[3] Nakamura K, Li Y, Ito W, et al.A machine learning approach for body
part recognition based on CT images[C].Medical Imaging.International Society
For Optics and Photonics, 2008：69141U-69141U-9.
【Invention content】
The medical image that technical problem to be solved by the invention is to provide a kind of adaptive abilities is strong, recognition accuracy is high
The recognition methods of middle organ.
Technical solution is used by the present invention solves above-mentioned technical problem：The identification side of organ in a kind of medical image
Method includes the following steps：
Pending medical image is obtained, the medical image is split into several X-Y schemes in X, Y and Z-direction respectively
Picture, and detection window is set according to the size of target organ；
Traversal detection is carried out to the two dimensional image respectively according to the detection step-length of setting using the detection window, is obtained
In the testing result of X, Y and Z-direction；
The testing result is subjected to result fusion, is retained on X, Y and three directions of Z axis all pixels of test positive
Point, so that it is determined that target organ boundary described in the medical image.
Further, further include the training process that AdaBoost cascade classifiers are generated using AdaBoost algorithms, specifically
For：
A) training data is built, and positive sample region and negative sample region, the positive sample are chosen from the training data
Region is the sampling window for including the target organ, and the negative sample region is the sampling for not including the target organ completely
Window；
B) the Haar characteristic values for calculating the positive sample and negative sample, are selected using AdaBoost algorithms from the Haar features
Effective Haar features are taken, each effectively Haar features form single Weak Classifier；
C) several described Weak Classifiers form single strong classifier, several described strong classifier cascades are constituted
AdaBoost cascade classifiers.
Further, the Haar features are calculated by integrogram, and the integrogram numerical value is a coordinate on image
The sum of the pixel value all put of the upper surface of point and the left side.
Further, it deletes and is less than the Haar features that are sized and delete position is adjacent and the identical Haar of size
Feature.
Further, the two dimensional image is traversed respectively according to the detection step-length of setting using the detection window
Detection, obtains the testing result in X, Y and Z-direction, and detailed process is：
It is detected respectively according to the detection step-length of setting using the AdaBoost cascade classifiers in the detection window
In the two dimensional image that X, Y and Z-direction are split, and preserve respectively in three directions by the AdaBoost cascade classifiers
Testing result；
Judge whether the traversal of two dimensional image is completed, if it is not, then continuing above-mentioned detection process；If it is not, then detection knot
Beam.
Further, the detection step-length is the distance between adjacent three pixels.
Further, further include to result merge after image split into two dimensional image in X, Y and Z-direction respectively, and
The pixel number that testing result on the two dimensional image is positive is counted respectively, and institute is further determined that according to the distribution of pixel
State the boundary of target organ.
Further, further determine that the boundary of the target organ is specially according to the distribution of pixel：Using Gauss
The method of fitting of distribution determines described image in the boundary maximum value and minimum value of X, Y and Z-direction, the boundary maximum respectively
The region that value and minimum value surround is the target organ range.
The present invention also proposes a kind of dividing method of organ in medical image, includes the following steps：
Image collection module is provided, described image acquisition module obtains pending medical image；
Target organ identification module is provided, the target organ identification module carries out relevant treatment to the medical image,
To obtain the target organ boundary；
Target organ is provided and divides module, the target organ segmentation module objectives organ boundaries are interior to the target organ
It is split；
The relevant treatment is：
The medical image is split into several two dimensional images along at least two reference directions respectively in three-dimensional system of coordinate,
And detection window is set according to the size of target organ；
Traversal detection is carried out to the two dimensional image respectively according to the detection step-length of setting using the detection window, is obtained
Along the testing result of corresponding reference direction；
The testing result is merged in three-dimensional system of coordinate, it is flat to be retained in the two dimension split along different reference directions
All pixel of test positive on face, so that it is determined that the target organ boundary.
Further, the reference direction is X-axis, Y-axis, Z axis or arbitrary combinations of directions between the two.
The present invention is generated compared with the existing technology to be had the beneficial effect that：First by three-dimensional medical image in X, Y and the side Z
To the multiple two dimensional images for splitting into sagittal plane, coronal-plane and cross section, then the image of traversal detection is merged again, not only
It solves the problems, such as lack of training samples, and can be applied to different target organ identification, it is adaptive adaptable；According to device
The fixed image resolution ratio of actual physical size selection of official, avoids carrying out repeatedly up-sampling and down-sampling to image, and choosing
During Haar features, ignore the Haar features and adjacent position size of the undersized stronger representative difference of randomness
Similar Haar features reduce operand on the basis of ensureing that Haar features are representative, improve recognition efficiency；It is chosen at three
Testing result is all positive pixel on a direction, excludes false positive point, and the Gauss according to pixel in all directions
Fitting distribution peaks determine the boundary of target image, can effectively remove noise at the boundary, and recognition accuracy is high.
【Description of the drawings】
Fig. 1 is the recognition methods flow chart of organ in medical image of the present invention；
Fig. 2 is that 3 d medical images of the present invention split into two dimensional image schematic diagram；
Fig. 3 is the original rectangular feature schematic diagram for image detection；
Fig. 4 is image-region integrogram；
Fig. 5 is to utilize AdaBoost cascade classifier detection image schematic diagrames；
Fig. 6 is the left kidney cross section testing result of the embodiment of the present invention；
Fig. 7 a are kidney testing result front view of the embodiment of the present invention；
Fig. 7 b are kidney testing result left view of the embodiment of the present invention；
Fig. 7 c are kidney testing result vertical view of the embodiment of the present invention；
Fig. 8 is left kidney cross section pixel Gauss curve fitting distribution schematic diagram；
Fig. 9 is the boundary schematic diagram of the kidney organ determined using the method for the present invention；
Figure 10 is the dividing method flow chart of organ in medical image of the present invention.
【Specific implementation mode】
To make the above purposes, features and advantages of the invention more obvious and understandable, with reference to the accompanying drawings and examples
The specific implementation mode of the present invention is described in detail.
In clinical diagnosis, medical image plays key player, medical image segmentation be medical image data analysis and
Visual first stage and computer-aided diagnosis, medical image three-dimensional visualization, image-guided surgery, it is virtual in peep
The primary premise and committed step of numerous medical image applications such as mirror.The people in accurate judgement medical image before medical image segmentation
The position of body organ sites plays an important roll for improving segmentation accuracy.It is as shown in Figure 1 device in medical image of the present invention
The flow chart of the recognition methods of official, mainly includes the following steps that：
S10, pending medical image is obtained, the medical image is split into several two in X, Y and Z-direction respectively
Image is tieed up, and detection window is set according to the size of target organ.For adult, the actual physical size of major body organs
Difference is simultaneously little, and the present invention carries out resampling to pending medical image, and the resolution ratio of medical image is made to be fixed as setting value.
Therefore it is based on actual physical size in the present invention, the resolution ratio of medical image is uniformly set as by X-resolution by resampling
=Y-resolution=Z resolution ratio=3mm.Medical image is split into two dimensional image in X, Y and Z-direction respectively, and according to target
The size of the size setting detection window of organ.It selects CT images to handle in the present invention, is generated for current CT equipment
Image is mostly 3-D view, two-dimensional Haar features is directly realized complex problem applied to 3-D view, such as Fig. 2 institutes
Show, three-dimensional CT image is split into sagittal plane, coronal-plane and three, cross section plane by the present invention respectively along X, Y and Z axis first
Two dimensional image is handled.In this specific implementation, 3 d medical images size is 117 × 117 × 107, i.e., is with X-direction
With reference to when share 117 layers of two dimensional image, each layer of size is 117 × 107 pixels；It is to be shared when referring to Y direction
117 layers of two dimensional image, each layer of size is 107 × 117 pixels；It is to share 107 layers of X-Y scheme when referring to Z-direction
Picture, each layer of size is 117 × 117 pixels.By aforesaid operations can solve the problems, such as follow-up lack of training samples with
And variability issues of the organ on three-dimensional planar, it is applicable to the different organ of pattern, improves adaptive ability.To organ
The actually available both ends apex coordinate P of positioningmin(xmin, ymin, zmin) and Pmax(xmax, ymax, zmax) determine.The present invention is by three
Victoria C T images split into the two dimensional image of three change in coordinate axis direction, train the Haar in three directions to scheme using the method for machine learning
As feature, by identifying that the sagittal plane two dimensional image split in X-direction can determine xminAnd xmax, by identifying in Y-axis side
Y can be determined to the coronal-plane two dimensional image of fractionationminAnd Ymax, by identifying the cross section two dimensional image split in Z-direction
It can determine ZminAnd Zmax, can thus convert three-dimensional problem to the problem of two dimension.
Flow is trained according to subsequent AdaBoost algorithms, it is thus necessary to determine that the size of detection window, and the CT of different model
Equipment shoots the image come, and often resolution ratio is different, changes so as to cause the pixel window size of organ.Preceding
It states in step, pending image resolution ratio has been unified value, and the selection principle of detection window is：It can be by entire target organ
In window.It should be noted that target organ is different, the size of the detection window of selection is also different.The present invention is by body
Body region is mainly divided into head, chest, abdomen and pelvic cavity, and wherein head is apparent because being characterized, and area is smaller, and identification holds
Easily, complicated algorithm process need not be passed through.The positioning of other body parts can be realized by positioning critical organ, of the invention
There is the organ (study and training that utilize Haar features) of feature in following embodiments on selected shape and gray value, chest is found
Critical organ be heart, abdomen is kidney, and pelvic cavity is femoral head.In one embodiment, target organ is kidney (abdomen),
The size of detection window is set as 24 × 36 (pixels) in X-direction (Y-Z plane)；It is set as in Y direction (Z-X planes)
36×24；It is set as 24 × 24 in Z-direction (X-Y plane).In another embodiment, target organ is chest, detects window
The size of mouth is both configured to 42 × 42 in X, Y and Z-direction.In pitching one embodiment, target organ is pelvic cavity, detection window
Size is both configured to 30 × 30 in X, Y and Z-direction.
S20, using setting detection window and setting detection step-length respectively to X, Y and Z-direction fractionation two dimension
Image carries out traversal detection, obtains the testing result in X, Y and Z-direction.Detection step-length is detection window in image to be detected
On moving step length, the size of moving step length can be the distance range etc. of 2-4 pixel.In a specific embodiment,
Traversing the process detected is：When detection window is moved to upper right side from image to be detected upper left side, 2 pixels are moved down
Point, then from left direction, side direction moves 2 pixels to the right again；Using AdaBoost cascade classifiers according to the inspection of setting
Window is surveyed according to above-mentioned mobile principle to carrying out above-mentioned detection respectively in the two dimensional image of X, Y and Z-direction fractionation, is preserved respectively
Pass through the testing result of AdaBoost cascade classifiers in three directions；Judge whether the traversal of two dimensional image is completed, if
It is no, then continue above-mentioned detection process；If it is not, then traversal detection terminates.
The AdaBoost cascade classifiers used in aforementioned process are trained by AdaBoost algorithms and are generated, AdaBoost
The learning process of algorithm is mainly, and when grader is correct to the classification of certain samples, the weights of these samples are reduced；When point
When class mistake, increase the weights of these samples.It is concentrated in subsequent study and more difficult training sample is learnt, finally
Obtain a series of weak features (corresponding Weak Classifier) that most can effectively distinguish whole picture sample image, the set of weights of a large amount of Weak Classifiers
Conjunction forms strong classifier.Specially：
A) training data is built, and positive sample region and negative sample region are chosen from training sample, positive sample region is packet
Sampling window containing target organ, more specifically, positive sample region exists for the sampling window regional percentage shared by target organ
50% or more, negative sample region is the sampling window for not including target organ completely.Training sample data are：{(x1, y1), (x2,
y2) ..., (xi, yi) ..., (xl, yl), wherein xiIndicate the training sample vector of input, yiPresentation class classification, and yi∈
{ 0,1 }, yi=0 indicates that the training sample is manually determined as negative sample, y in advancei=1 indicates that the training sample is advance
It is determined as that positive sample, l are all number of training.In addition, also needing the number T that setting recycles to determine weak typing in strong classifier
The number of device, and positive sample and negative sample weight are initialized, for each positive sample weightFor each negative sample
WeightWherein m indicate positive sample number, n indicate negative sample number, total weight of all positive negative samples and be 1.Structure foot
Enough positive samples and negative sample are the trained most important condition, it is contemplated that machine learning not only wants the feature of learning objective organ to believe
Breath, it is also necessary to which the characteristic information around learning objective organ, therefore, sampling window meeting surrounding target organ central point are discharged
Sliding, to extend positive sample quantity；Negative sample sampling window then cannot generate overlapping with target organ.To ensure the complete of negative sample
Face property, negative sample detection mode of the present invention are fixed test step-length, and removal skin is with outer portion and target organ area
Step-length sampling is fixed, so that it is guaranteed that there are a certain number of negative samples at each position of body in domain in remaining image
This.
B) the Haar characteristic values of positive sample and negative sample are calculated, it is effective from Haar Feature Selections by AdaBoost algorithms
Haar features, each effectively Haar features form single Weak Classifier.The method of study image distribution rule has very much, mainly has
Learn and be based on characteristics of image learning method point by point according to pixel, wherein need to be grasped based on the point-by-point learning method of pixel each
The regularity of distribution of the pixel of organ, operand is larger for study, inefficiency；Under the recognition effect for reaching same, based on figure
As the calculation amount of feature learning is smaller, arithmetic speed is fast.In the present invention using to simple image be characterized in Haar features, such as scheme
3 show the original rectangular feature for image detection.As can be seen from Fig., there are many form, each small squares of black and white for Haar features
Shape is mutually perpendicular to or horizontally adjacent, difference of the characteristic value for the sum of white portion pixel with black portions pixel.
In the specific embodiment of the invention, the calculating of Haar characteristic values is obtained using integrogram, and the definition of integrogram is：For in image
A certain coordinate points (x, y), the sum of the pixel value of the pixel upper left corner (the left side and top) all the points, formula is：Wherein i (x ', y ') is the pixel that the coordinate of upper left corner area is (x ', y ') picture point
Value.As shown in figure 4, be an image-region integrogram schematic diagram, for four rectangles in figure, 2 points of integrogram is I+II, 3
The integrogram of point is I+III, then all pixels point of region IV and subtracts 2 points for the sum of 4 dot product components and 1 dot product component
The sum of figure and 3 dot product components.
The present invention mainly uses L-R feature, previous-next feature, left-in-right feature, upper-in-lower feature, left diagonal 45 °
Feature, right diagonal 45 ° of features and central feature this 7 kinds of Haar features.For the image window of a typical 30x30, pass through
By above-mentioned 7 kinds of Haar feature equal proportion zoom, the total Haar feature quantities that can be calculated the detection window are 394395,
The Haar features of the order of magnitude in this way are excessively huge for last test process.The present invention is in the generation for not weakening Haar features
In the case of table, on the one hand too strong, the representative feature poor, quantity is various using too small Haar feature randomnesss, ignores
Undersized Haar features, such as the sampling window size of 1x1 or 2x2；On the other hand, being needed in consideration detection process will be whole
The sliding sampling of a sampling window, there is a situation where a large amount of Haar features to be repeated calculating, ignores that position is adjacent and size is identical
Haar features.In this particular embodiment, the Haar smallest feature sizes that kidney is chosen are 3 × 3, at X-direction (Y-Z plane)
The quantity of Haar characteristic values is 6591, and positive sample quantity is 4608, and negative sample quantity is 41563；In Y direction (Z-X
Plane) Haar characteristic values quantity be 6591, positive sample quantity be 3204, negative sample quantity be 63144；In Z axis side
It it is 8493 to the quantity of (X-Y plane) Haar characteristic values, positive sample quantity is 3012, and negative sample quantity is 50211.Inspection
It is heart to survey target organ when position is chest, and the Haar smallest feature sizes of selection are 4 × 4, at X-direction (Y-Z plane)
The quantity of Haar characteristic values is 7132, and positive sample quantity is 5636, and negative sample quantity is 43780；In Y direction (Z-X
Plane) Haar characteristic values quantity be 7132, positive sample quantity be 5320, negative sample quantity be 41971；In Z axis side
It it is 7132 to the quantity of (X-Y plane) Haar characteristic values, positive sample quantity is 4761, and negative sample quantity is 56752.
In the above process, by the representative poor small size Haar features of the removal Haar feature adjacent with same size is ignored so that
The Haar feature quantities that the quantity of Haar characteristic values obtains than existing methods reduce by two orders of magnitude, are ensureing sampling window representative
Arithmetic speed is significantly improved under the premise of property.
Individually the definition procedure of Weak Classifier h is specially：It is total to have k Haar feature for piece image x, choose it
In a Haar feature t, calculate value fs of the image x about this Haar feature tt(x), and one and the relevant thresholds of this feature t are chosen
Value θt, the calculation formula of Weak Classifier is：
In formula, 1 indicates to be detected as positive sample, and 0 indicates to be detected as negative sample, ptFor polarity instruction symbol, when pt takes+1,
Characteristic value is positive sample when being more than threshold value, and when pt takes -1, characteristic value is positive sample when being less than threshold value, and x indicates detection window.Profit
With the weight of the abovementioned steps a) positive samples divided in advance and negative sample and all m+n samples.For a Haar feature
T, calculates characteristic value of all samples about t, and characteristic value is arranged according to ascending order, is denoted as f (t1), f (t2) ..., f
(tm+n), therefrom divide some threshold θtWatershed as classification.θtSelection it is related with wrong classification rate, mistake classification rate et
Calculation be：
et=min (St ++(Tt --St -), St -+(Tt +-St +))
Wherein, min expressions are minimized function, Tt +Indicate the weight and T of all positive samplest -Indicate all negative samples
Weight and St +For in θtAll characteristic values are less than θ under threshold valuetPositive sample weight and, St -For in θtAll characteristic values under threshold value
Less than θtNegative sample weight and.Under original state, all samples are all endowed equal weighted value.It should be noted that
Along with the progress of learning process, the data for being judged as negative sample are constantly dropped, and the quantity for being judged as positive sample is continuous
It changes, therefore, all positive samples and negative sample need to be standardized, weight standardization formula is：J indicates the number for being judged as positive sample so thatAll features are traversed, are chosen
The Weak Classifier of wrong resolution ratio minimum is added in strong classifier, minimal error resolution ratioThe then weight of update next round training sampleWhereinThe e if sample is correctly classifiedi=0, otherwise, then ei=1.According to upper
State AdaBoost iterative algorithms, iteration taken turns by T, selection is obtained into T Haar feature, it is of the invention in T=50.If table 1 is this
The X-direction training result effective Haar feature quantities at different levels used in inventive embodiments.
1 X-direction training result of the table effective Haar feature quantities at different levels used
Be illustrated in figure 5 the cascade classifier schematic diagram of the embodiment of the present invention, per level-one strong classifier be all according to
AdaBoost algorithm constructions, strong classifiers at different levels are from simple to complexity.When newly input piece image, if the image passes through
The test of all strong classifiers, then it is assumed that image belongs to positive sample；If divided by force not over a certain group during the test
The test of class device, then immediately by the spectral discrimination be negative sample, no longer carry out subsequent processing.In this embodiment, abdomen
Kidney is 310 in the effective Haar feature quantities of X-direction (Y-Z plane) obtained by training；In Y direction, (Z-X is flat
Face) effective Haar feature quantities be 193；It it is 146 in effective Haar feature quantities of Z-direction (X-Y plane).Chest
Heart is 146 in the effective Haar feature quantities of X-direction (Y-Z plane) obtained by training；In Y direction, (Z-X is flat
Face) effective Haar feature quantities be 133；It it is 58 in effective Haar feature quantities of Z-direction (X-Y plane).It is above-mentioned
Each effectively Haar features form a Weak Classifier, and multiple Weak Classifiers form strong classifier.As shown in figure 5, utilizing
AdaBoost cascade classifiers detect two in X, Y and Z-direction fractionation according to the detection window and detection step-length of setting respectively
Image is tieed up, is preserved respectively in three directions through the testing result of AdaBoost cascade classifiers；Judge time of two dimensional image
It goes through and whether completes, if it is not, then continuing above-mentioned detection process；If it is not, then detection terminates.It should be noted that in the present invention
The detection step-length of detection window is primary every 3 pixel point samplings, without being sampled point by point, to improve test speed.It is right
In each width two dimension test image, the cascade classifier that sharp AdaBoost is trained is tested, and is recorded each group and is passed through
Position of the image of test in former 3-D view.If a sample by all strong classifiers in cascade classifier,
The value for recording all pixels point in the window is 1, as positive (positive sample)；If detecting sample not over cascade classifier
Any one of strong classifier, then it represents that not over testing and being denoted as 0, as negative (negative sample), if Fig. 6 is the present invention
The left kidney cross section testing result obtained in one embodiment, internal shaded boxes part are the region for being detected as kidney.By three
A direction is individually handled, and the intermediate result that each direction is handled is preserved.
S30, testing result is subjected to result fusion, it is all positive pixel to be retained on X, Y and three directions of Z axis,
So that it is determined that target organ boundary.Due in aforementioned process for X, Y of 3-D view and the corresponding sagittal plane of Z axis, coronal
For face and three, cross section plane, a pixel may be marked as positive number in the graphics by traversal detection
It is 0~3 time.However, the positive test symbol in single direction and unreliable, is susceptible to the judgement of false positive, in order to reinforce most
The robustness of termination fruit, the present invention merge the test result in three directions, and reservation is all detected as sun in three directions
The pixel of property.As Fig. 7a-7c show target organ kidney front view, left view and vertical view testing result, testing result
Main region include kidney, certainly still include part kidney organ region except pixel, these pixels sagittal plane,
Coronal-plane and three, cross section direction all test positive are false positive point.
In the present invention, the pixel of the image after further merging, which is split to again on X, Y and three directions of Z axis, to be seen
It examines, by way of Gauss Distribution Fitting, the pixel number of statistics and test positive on the vertical direction in fractionation direction respectively,
The boundary of the target organ is further determined that according to the distribution of pixel, removes noise at the boundary point.It is illustrated in figure 8 left kidney
Cross section pixel Gauss curve fitting distribution, abscissa indicate that the distance along body Z axis from bottom to top, ordinate indicate vertical along body
To the pixel number of every layer of test positive, there are one apparent wave crests for the distribution of pixel in figure, but also have around wave crest
Fragmentary false positive pixel distribution.In the present embodiment by taking the distribution of left kidney as an example, select reliable wave crest up-and-down boundary can
With the Pz of the left kidney of determinationminAnd Pzmax.Pass through Gauss curve fittingX denotation coordinations in formula
Position, G (x) indicate positive pixel number, find out desired value μ and standard deviation sigma.According to P the characteristics of Gaussian Profile (μ-σ ＜ X≤
μ+σ)=68.3%, it is different to choose the sample total that different Gauss sections includes.It is the boundary of determining kidney in the present invention
Range, the fit interval of use are (μ-σ, μ+σ), μ=99.92, σ=9.20, and then can obtain kidney under Z-direction
Boundary Pzmin=μ-σ, coboundary Pzmax=μ+σ.It, can be true by the fitting result in another two direction according to above-mentioned same operation
Determine Pxmin, Pxmax, PyminAnd Pymax, the P of entire organ is just determined in this waymin(xmin, ymin, zmin) and Pmax(xmax, ymax,
zmax), and then the cuboid boundary of entire kidney organ as shown in Figure 9 is determined.
In one embodiment, target organ is selected as heart, selects chest (103 groups of sample data) and non-in experiment respectively
The data at the multiple positions of chest (head, abdomen totally 211 groups of data) are tested, manually to judge result as goldstandard, detection
The results are shown in Table 2.Wherein, it is the positive that true positives, which indicate actual sample pixel and testing result all,；True negative indicates sample
Actual pixels point and testing result are all feminine gender；False negative indicates that sample actual pixels point is the positive and testing result is negative；
False positive indicates that sample actual pixels point is feminine gender and testing result is the positive, most important for the position comprising target organ
Be identification sensitivity, and for the position not comprising target organ, it is most important that identification specificity, wherein susceptibility=true
Positive sample/(true positives sample+false negative sample), specificity=true negative sample/(true negative sample+false positive sample),
97.9% is up to for the susceptibility of chest data heart, and is 100% to the identification specificity of non-chest position organ.
2 target organ of table is the testing result of heart
In another embodiment, target organ is selected as left kidney, selects pelvic cavity (71 groups of sample data) in experiment respectively
It is tested with the data at the multiple positions of non-pelvic cavity (head, chest, abdomen totally 156 groups of data), test result such as 3 institute of table
Show, the susceptibility for pelvic cavity data kidney is 85.9%, and is 98.8% to the identification specificity of non-pelvic cavity position organ.
3 target organ of table is the testing result of left kidney
In addition, in pitching an embodiment, we select pelvis femoral head for target organ.In view of head exist part with
Femur head shapes similar structure selects 55 groups of data in the present embodiment, wherein 21 groups of header datas are as negative sample, in addition 34
Group is abdomen pelvic cavity data, the target organ window size that X, Y and Z-direction are chosen all be 30 × 30, Haar characteristic sizes most
Small is 5 × 5, and finally determining Haar feature quantities share 7716.It should be noted that left and right femoral head exists centainly
Similitude often can also identify left femur head in the test for carrying out right femoral head.Therefore, we by left femur head along body
Body central axes are turned to right side, are added in the training data of right femoral head, have not only expanded training sample amount, but also improve femur
The discrimination of head.The recognition methods of organ in medical image of the present invention, the front and rear part for calculating Haar characteristic values do not have dependence,
It can accelerate arithmetic speed with parallel computation, quickly identify target organ finally by Gauss curve fitting method, recognition accuracy is high, leads to
Often the time required to one sample of test within 1-5min.And the adaptive ability of the method for the present invention is stronger, can be adapted for
The case where Different Individual organs differences, can also obtain the sample not learnt in training process preferably to identify knot
Fruit.It should be noted that the present invention can be applied not only to the identification of target organ in CT images, magnetic resonance is applied also for
(MRI), at the imaging of equipment such as single photon emission computerized tomography (SPECT), positron emission tomography (PET)
Reason.
In above-mentioned medical image on the basis of the recognition methods of organ, the present invention also proposes organ in a kind of medical image
Dividing method, system for use in carrying includes image collection module, target organ identification module, target organ segmentation module, such as Figure 10 institutes
Show the specific steps are：
Image collection module obtains pending medical image, and medical image can be the equipment such as MRI, CT, SPECT, PET
The image of acquisition may include multiple body parts such as head, chest, abdomen, pelvic cavity in image.
Target organ identification module carries out traversal detection to medical image and obtains testing result, and obtains according to testing result
Target organ boundary, specially：Medical image is detected using effective Haar features of selection, obtains the boundary of target organ,
Medical image is split into several two dimensional images along at least two reference directions respectively, and is set and is examined according to the size of target organ
Survey window；Machine learning method based on two dimensional image chooses effective Haar features, is formed according to effective Haar features
AdaBoost cascade classifiers carry out traversal detection to two dimensional image respectively using detection window according to the detection step-length of setting,
Obtain the testing result along corresponding reference direction；Sampled result is subjected to result fusion in three-dimensional system of coordinate, is retained in along not
The pixel of all test positive on the two dimensional surface split with reference direction, so that it is determined that containing object machine in medical image
Official obtains target organ boundary, and then targetedly calls the region comprising target organ, reduces the work of follow-up organ segmentation
Use region.It should be noted that reference direction can be X-axis, Y-axis, Z axis or arbitrary combinations of directions between the two, i.e. X-Y,
Any direction between Y-Z or Z-X.In the specific embodiment of the invention, mesh reference direction is selected as X-axis, Y-axis, Z axis three and tears open
Divide direction, obtains two dimensional image in three directions.
By above-mentioned target organ identification process, the region for including target organ is can get, then using target organ point
Module is cut to be split target organ in the target organ boundary that target organ identification module obtains.Target organ is divided
Method can be：Cluster segmentation is carried out to the textural characteristics of image using Hopfield networks；Figure based on bayes method
As segmentation；The fuzzy connectedness segmentation method that plane learns aspect graph registration is conciliate based on image；Utilize the organ of 3D region growth method
Automatic segmentation etc..The method of the present invention targetedly calls partitioning algorithm to the region comprising target organ, avoids traditional images
In dividing method in the case of indefinite corresponding body part, the shortcomings that blindly calling the partitioning algorithm of all organs, section
Processing time has been saved, treatment effeciency is improved.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention
With within principle, any modification, equivalent substitution, improvement and etc. done should all be included in the protection scope of the present invention god.
Claims (8)
1. the recognition methods of organ in a kind of medical image, which is characterized in that include the following steps：
Pending medical image is obtained, the medical image is split into several two dimensional images in X, Y and Z-direction respectively,
And detection window is set according to the size of target organ；
Using AdaBoost cascade classifiers according to the detection window and the detection step-length of setting respectively to the two dimensional image
Traversal detection is carried out, the testing result in X, Y and Z-direction is obtained；
The testing result is subjected to result fusion, is retained on X, Y and three directions of Z axis all pixels of test positive,
So that it is determined that target organ boundary described in the medical image；
Further include two dimensional image being split into X, Y and Z-direction respectively to the image after result fusion, and count described two respectively
The pixel number that testing result on image is positive is tieed up, the side of the target organ is further determined that according to the distribution of pixel
Boundary；
Further determine that the boundary of the target organ is specially according to the distribution of pixel：Using the method for Gauss Distribution Fitting
Described image is determined respectively in the boundary maximum value and minimum value of X, Y and Z-direction, and the boundary maximum value and minimum value surround
Region be the target organ range.
2. the recognition methods of organ in medical image according to claim 1, which is characterized in that further include utilizing
AdaBoost algorithms generate the training process of AdaBoost cascade classifiers, specially：
A) training data is built, and positive sample region and negative sample region, the positive sample region are chosen from the training data
To include the sampling window of the target organ, the negative sample region is the sample window for not including the target organ completely
Mouthful；
B) the Haar characteristic values for calculating the positive sample and negative sample, have using AdaBoost algorithms from the Haar Feature Selections
Haar features are imitated, each effectively Haar features form single Weak Classifier；
C) several described Weak Classifiers form single strong classifier, several described strong classifier cascades constitute AdaBoost grades
Join grader.
3. the recognition methods of organ in medical image according to claim 2, which is characterized in that the Haar features pass through
Integrogram is calculated, the integrogram numerical value be on image the pixel value all put on the upper surface of coordinate points and the left side it
With.
4. the recognition methods of organ in medical image according to claim 3, which is characterized in that deletion, which is less than, to be sized
Haar features and delete position is adjacent and the identical Haar features of size.
5. the recognition methods of organ in medical image according to claim 3, which is characterized in that cascaded using AdaBoost
Grader carries out traversal detection to the two dimensional image respectively according to the detection window and the detection step-length of setting, obtain X,
The testing result of Y and Z-direction, detailed process are：
It is detected respectively in X, Y according to the detection step-length of setting using the AdaBoost cascade classifiers in the detection window
The two dimensional image split with Z-direction, and preserve respectively in three directions by the inspection of the AdaBoost cascade classifiers
Survey result；
Judge whether the traversal of two dimensional image is completed, if it is not, then continuing above-mentioned detection process；If it is, detection terminates.
6. the recognition methods of organ in medical image according to claim 5, which is characterized in that the detection step-length is phase
The distance between adjacent three pixels.
7. the dividing method of organ in a kind of medical image, which is characterized in that include the following steps：
Image collection module is provided, described image acquisition module obtains pending medical image；
Target organ identification module is provided, the target organ identification module carries out relevant treatment to the medical image, to obtain
Take the target organ boundary；
Target organ is provided and divides module, the target organ is carried out in the target organ segmentation module objectives organ boundaries
Segmentation；
The relevant treatment is：
The medical image is split into several two dimensional images, and root along at least two reference directions respectively in three-dimensional system of coordinate
Detection window is set according to the size of target organ；
Using AdaBoost cascade classifiers according to the detection window and the detection step-length of setting respectively to the two dimensional image
Traversal detection is carried out, the testing result along corresponding reference direction is obtained；
The testing result is merged in three-dimensional system of coordinate, is retained in along the two dimensional surface that different reference directions are split
All pixels of test positive, so that it is determined that the target organ boundary；
Further include two dimensional image being split into X, Y and Z-direction respectively to the image after result fusion, and count described two respectively
The pixel number that testing result on image is positive is tieed up, the side of the target organ is further determined that according to the distribution of pixel
Boundary；
Further determine that the boundary of the target organ is specially according to the distribution of pixel：Using the method for Gauss Distribution Fitting
Described image is determined respectively in the boundary maximum value and minimum value of X, Y and Z-direction, and the boundary maximum value and minimum value surround
Region be the target organ range.
8. the dividing method of organ in medical image according to claim 7, which is characterized in that the reference direction is X
Axis, Y-axis, Z axis or arbitrary combinations of directions between the two.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510729150.0A CN105389813B (en) | 2015-10-30 | The recognition methods of organ and dividing method in medical image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510729150.0A CN105389813B (en) | 2015-10-30 | The recognition methods of organ and dividing method in medical image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105389813A CN105389813A (en) | 2016-03-09 |
| CN105389813B true CN105389813B (en) | 2018-08-31 |
Family
ID=
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101526997A (en) * | 2009-04-22 | 2009-09-09 | 无锡名鹰科技发展有限公司 | Embedded infrared face image identifying method and identifying device |
| CN103295256A (en) * | 2012-01-24 | 2013-09-11 | 株式会社东芝 | Medical image processing apparatus and medical image processing program |
| CN103914697A (en) * | 2012-12-29 | 2014-07-09 | 上海联影医疗科技有限公司 | Extraction method for region of interest of breast three-dimension image |
| CN104637056A (en) * | 2015-02-02 | 2015-05-20 | 复旦大学 | Method for segmenting adrenal tumor of medical CT (computed tomography) image based on sparse representation |
| CN104751434A (en) * | 2013-12-25 | 2015-07-01 | 北京三星通信技术研究有限公司 | Method and apparatus for dividing object from image |
| CN104798107A (en) * | 2012-11-23 | 2015-07-22 | 皇家飞利浦有限公司 | Generating a key-image from a medical image |
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101526997A (en) * | 2009-04-22 | 2009-09-09 | 无锡名鹰科技发展有限公司 | Embedded infrared face image identifying method and identifying device |
| CN103295256A (en) * | 2012-01-24 | 2013-09-11 | 株式会社东芝 | Medical image processing apparatus and medical image processing program |
| CN104798107A (en) * | 2012-11-23 | 2015-07-22 | 皇家飞利浦有限公司 | Generating a key-image from a medical image |
| CN103914697A (en) * | 2012-12-29 | 2014-07-09 | 上海联影医疗科技有限公司 | Extraction method for region of interest of breast three-dimension image |
| CN104751434A (en) * | 2013-12-25 | 2015-07-01 | 北京三星通信技术研究有限公司 | Method and apparatus for dividing object from image |
| CN104637056A (en) * | 2015-02-02 | 2015-05-20 | 复旦大学 | Method for segmenting adrenal tumor of medical CT (computed tomography) image based on sparse representation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104899926B (en) | Medical image cutting method and device | |
| Zhang et al. | Review of breast cancer pathologigcal image processing | |
| CN107622492A (en) | Lung splits dividing method and system | |
| Campadelli et al. | A segmentation framework for abdominal organs from CT scans | |
| Wu et al. | AAR-RT–a system for auto-contouring organs at risk on CT images for radiation therapy planning: principles, design, and large-scale evaluation on head-and-neck and thoracic cancer cases | |
| US20230005140A1 (en) | Automated detection of tumors based on image processing | |
| Qiu et al. | Lung Nodules Detection in CT Images Using Gestalt‐Based Algorithm | |
| Ng et al. | Medical image segmentation using watershed segmentation with texture-based region merging | |
| CN113160120A (en) | Liver blood vessel segmentation method and system based on multi-mode fusion and deep learning | |
| Ramasamy et al. | Machine learning in cyber physical systems for healthcare: brain tumor classification from MRI using transfer learning framework | |
| Yang et al. | Unsupervised domain adaptation for cross-device OCT lesion detection via learning adaptive features | |
| Ramana | Alzheimer disease detection and classification on magnetic resonance imaging (MRI) brain images using improved expectation maximization (IEM) and convolutional neural network (CNN) | |
| Xiaojie et al. | Segmentation of the aortic dissection from CT images based on spatial continuity prior model | |
| Umadevi et al. | Enhanced Segmentation Method for bone structure and diaphysis extraction from x-ray images | |
| Sasikala et al. | Soft Computing based Brain Tumor Categorization with Machine Learning Techniques | |
| Rezaeitabar et al. | Automatic 3D segmentation of individual facial muscles using unlabeled prior information | |
| Li et al. | Micro-Ct synthesis and inner ear super resolution via generative adversarial networks and bayesian inference | |
| Zhou et al. | Automatic segmentation of liver from CT scans with CCP–TSPM algorithm | |
| CN105389813B (en) | The recognition methods of organ and dividing method in medical image | |
| Yan et al. | Segmentation of pulmonary parenchyma from pulmonary CT based on ResU-Net++ model | |
| Hamarneh | Towards intelligent deformable models for medical image analysis | |
| Liu et al. | Automatic Lung Parenchyma Segmentation of CT Images Based on Matrix Grey Incidence. | |
| Marone et al. | Learning the lumen border using a convolutional neural networks classifier | |
| Mughal et al. | Early lung cancer detection by classifying chest CT images: a survey | |
| TWI548401B (en) | Method for reconstruction of blood vessels 3d structure |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| CP01 | Change in the name or title of a patent holder |
Address after: 201807 No. 2258 Chengbei Road, Jiading Industrial Zone, Jiading District, Shanghai. Patentee after: Shanghai Lianying Medical Technology Co., Ltd Address before: 201807 No. 2258 Chengbei Road, Jiading Industrial Zone, Jiading District, Shanghai. Patentee before: SHANGHAI UNITED IMAGING HEALTHCARE Co.,Ltd. |