CN104809465A - Classifier training method, target detection, segmentation or classification method and target detection, segmentation or classification device - Google Patents

Abstract

The invention discloses a classifier training method, a target detection, segmentation or classification method and a target detection, segmentation or classification device. The classifier training method comprises the steps of acquiring a plurality of training samples, wherein each training sample corresponds to a classifier; and carrying out training on the classifiers through minimizing a predetermined objective function so as to acquire each classifier.

Description

Sorter training method, target detection, segmentation or sorting technique and device

Technical field

The present invention relates to the technology of automatic synchronization target detection, segmentation and classification, more particularly, relate to a kind of the sorter for target detection is trained method, the target detection applying described sorter, segmentation or sorting technique and target detection, segmentation or sorter.

Background technology

Ultrasonic technique due to its Noninvasive, can carry, the feature such as low cost, be the topmost instrument of human body internal object.Because manual detection is very expensive, and consuming time, and therefore developing full automatic object detection systems is current study hotspot.But such research is still in the face of a lot of challenge, this is because ultrasonoscopy itself is containing noise, and the object of inside of human body itself has diversity.

At present, the full-automatic object detection systems of great majority comprises the module of following three serials: the position detecting module of the position of (1) detected object; (2) border of object is obtained, thus the segmentation module that object is split from background; (3) sort module of Properties of Objects (tumour such as, in human body is pernicious or optimum) is judged.

Below the prior art of object detection, segmentation and classification is described in detail.

In medical image, the object (such as, tumour, organ, tissue etc.) of human body inside is the problem of a very challenging property.Early stage system is all generally based on some didactic rules.Such as, Drukke(Drukker, K., Giger, ML., Horsch, K., Kupinski, MA., Vyborny, CJ., Menelson, EB.:Computerized lesion detection on breast ultrasound.Med.Phys., 29 (7): 1438-46 (2002)) a strict hypothesis etc. is proposed: " tumor region is more black than background ".Therefore, the method cannot detect the tumor region of high echo response, this is because this tumour is brighter than background.In recent years, machine learning algorithm is introduced into detection field.Based on sliding window technique, (Zhang, J.D., the Zhou S.K. such as Zhang, Brunke, S., Lowery, C., Comaniciu, D.:Database-guided breast tumor detection and segmentation in2Dultrasound images.SPIE, 7624 (3), 1-7 (2010)) use Harr characteristic sum boosting algorithm to judge whether each window comprises tumour.DPM(Deformable Part-based model, Felzenszwalb, P.F, Girshick, R.B, McAllester, D., Ramanan, D.:Object Detection with discriminatively trained part-based models.PAMI, 32 (9), 1627-1645 (2009)) be general object detecting method in most popular image in recent years.DPM extracts HOG feature (Histogram of Oriented Gradient, Dalal, N., Triggs, B.:Histograms of Oriented Gradients for Human Detection.CVPR, 886-893 (2005)) and support vector machine classifier judge whether tumour exists.DPM algorithm requires to be categorized into 2 or 3 templates according to length breadth ratio in advance to target classification, because which limit possible tumour template number.In other words, this algorithm only considers 2 or 3 possible length breadth ratio templates, and tumour can be any shape and length breadth ratio.ESVM algorithm (Tomasz Malisiewicz, Abhinav Gupta, and Alexei A.Efros.Ensemble of exemplar-svms for object detectionand beyond.In ICCV, pages86 – 91,2011) DPM algorithm can be regarded as when the maximized distortion of template number.ESVM is considered as a possible template each training sample, trains corresponding sorter.But this algorithm does not consider the local message of object, the performance therefore in object detection is not as DPM algorithm.

In Medical Image Processing, the dividing method of view-based access control model technology conventional at present can be divided into roughly two classes: curve drives and region drives.Last class methods, such as level set (Horsch, K., Giger, M.L., Venta, L.A., Vyborny, C.J.:Automatic segmentation of breastlesions on ultrasound.Med.Phys., 1652 (28), 1652-1659 (2001); Chan, T.F., Vese, L.A.:Active contours without edges.IEEE Trans.on ImageProcessing, 10 (2): 266-276 (2001)) method develops from an initial edge curve, until reach the optimal value of energy function.And the main difference of various Level Set Method is just the definition of energy function.In general, energy function needs to consider factor usually: length of curve, area under the curve, line smoothing degree, inside or outside of curve gray scale difference etc.(the Zhihua Liu such as Liu, Lidan Zhang, Haibing Ren, and Ji-Yeun Kim.A robust region-based active contourmodel with point classification for ultrasound breast tumorsegmentation.In SPIE Medical Imaging, 2013) edge pixel is classified, thus improves the accuracy of segmentation.Based on region evolves as markov random file (Varun Gulshan, Carsten Rother, Antonio Criminisi, Andrew Blake, and Andrew Zisserman.Geodesic star convexity for interactive image segmentation.In CVPR, pages3129 – 3136,2010) first some super-pixel are divided into image, and constructing a figure, node of graph is each super-pixel.According to this figure, an energy function can be defined, thus to mark each super-pixel be foreground/background.

Current most of object classification is by extracting contour feature, and design corresponding sorter (such as, Chang, Ruey-Feng, Wen-Jie Wu, Woo Kyung Moon, Yi-Hong Chou, and Dar-RenChen.Support vector machines for diagnosis of breast tumors on USimages.Academic radiology10, no.2 (2003): 189-197) classify.

Such tandem system considers each subproblem independently, and uses distinctive technology to independently solve each subproblem.The key issue of this serial type system is error propagation.Because each module relies on the result of previous step module completely, and each module has its oneself error rate, and therefore the error rate of whole system will be the accumulation of each module error rate.

Summary of the invention

An aspect of of the present present invention is to provide one to detect simultaneously, split and class object, thus the target detection of the sorter avoiding the error propagation sorter training method that produces due to the use of submodule and application to obtain according to described sorter training method, segmentation or sorting technique and target detection, segmentation or sorter.

According to an aspect of the present invention, provide a kind of sorter training method, comprising: obtain multiple training sample, wherein, each training sample corresponds to a sorter; By minimizing predetermined objective function, sorter is trained, thus obtain each sorter; Use achievement fusion method to correct each sorter, thus make each sorter phase-splitting is compatible.

Preferably, described predetermined objective function is:

L(β _E)=||β _E|| ²+h(score(E))+∑ _x∈N(E)h(-score(x))，

Wherein, β _erepresent the weight vectors of sample, h (x)=max (0,1-x) represents hinge loss function, and score (E) represents the score of positive sample E, N (E) represents the set of negative sample, and score (x) represents the score of negative sample x.

Preferably, score (E) calculates according to following equation:

$\mathrm{score}\left(E\right)=\mathrm{score}(e_0,e_1,...,e_n)={\mathrm{\Σ}}_{i=0}^n{\mathrm{\β}}_i\mathrm{\φ}\left(e_i\right)-{\mathrm{\Σ}}_{i=1}^n{d}_i{\mathrm{\φ}}_d\left(e_i\right)+b_E,$

Wherein, E=(e ₀, e ₁..., e _n) represent sample, e ₀represent the overall template of sample, e _irepresent the local template of sample, φ (e _i) representation feature function, φ _d(e _i) represent the proper vector of current local template relative to the upper left corner displacement d of overall template, β _i, d _iand b _efor parameter, wherein, (dx _i, dy _i)=(x _i, y _i)-(2 (x ₀, y ₀)+v _i), (x ₀, y ₀) represent the top left co-ordinate of overall template, (x _i, y _i) represent the top left co-ordinate of current local template, v _irepresent (x _i, y _i) relative to (x ₀, y ₀) displacement.

Preferably, optimum β is obtained by Latent support vector machine _i, d _iand b _e.

Preferably, each sorter is corrected by matching sigmoid distribution.

Preferably, by using following equation to correct each sorter:

$s\left(x\right|A_E,B_E)=\frac{1}{1+\exp (A_E\·(\mathrm{score}\left(x\right)+B_E)},$

Wherein, s (x|A _e, B _e) represent that the score of each sorter after correcting, score (x) represent the score of each sorter before correcting, A _eand B _eit is the parameter obtained by stand-alone training.

Preferably, described fundamental function is Harr fundamental function or HOG fundamental function.

According to a further aspect in the invention, a kind of target detection, segmentation or sorting technique are provided, comprise: by for each classifier calculated input amendment in multiple sorter score and the score of each input amendment is corrected, obtain each other compatible sorter; By affined transformation, the profile of corresponding sorter is mapped to the profile that input amendment obtains destination object; Based on the classification of the classification determination destination object of corresponding sorter.

Preferably, the score of input amendment is calculated by following equation:

${\mathrm{\φ}}_d\left(e_i\right)=({\mathrm{dx}}_i,{\mathrm{dy}}_i,{\mathrm{dx}}_i^2,{\mathrm{dy}}_i^2),({\mathrm{dx}}_i,{\mathrm{dy}}_i)=(x_i,y_i)-(2(x_0,y_0)+v_i),$

Wherein, score (E) represents the score of positive sample E, E=(e ₀, e ₁..., e _n) represent sample, e ₀represent the overall template of sample, e _irepresent the local template of sample, φ (e _i) representation feature function, φ _d(e _i) represent the proper vector of current local template relative to the upper left corner displacement d of overall template, β _i, d _iand b _efor parameter, wherein, (dx _i, dy _i)=(x _i, y _i)-(2 (x ₀, y ₀)+v _i), (x ₀, y ₀) represent the top left co-ordinate of overall template, (x _i, y _i) represent the top left co-ordinate of current local template, v _irepresent (x _i, y _i) relative to (x ₀, y ₀) displacement.

Preferably, the score by using following equation to correct each input amendment:

$s\left(x\right|A_E,B_E)=\frac{1}{1+\exp (A_E\·(\mathrm{score}\left(x\right)+B_E)},$

Wherein, s (x|A _e, B _e) represent that the score of each input amendment after correcting, score (x) represent the score of each input amendment before correcting, A _eand B _eit is the parameter obtained by stand-alone training.

Preferably, the sorter corresponding to top score is defined as corresponding sorter.

Preferably, described fundamental function is Harr fundamental function or HOG fundamental function.

Preferably, affined transformation is carried out by following equation:

P'=UP+V，

Wherein, P={p ₁, p ₂..., p _nrepresenting that the profile of corresponding sorter, U and V represent change of scale and coordinate position skew respectively, P' represents the profile of destination object.

According to a further aspect in the invention, a kind of target detection, segmentation or sorter are provided, comprise: sorter determining unit, be configured to by for each classifier calculated input amendment in multiple sorter score and the score of each input amendment is corrected, determine each other compatible sorter; Cutting unit, is configured to, by affined transformation, the profile of corresponding sorter is mapped to the profile that input amendment obtains destination object; Taxon, is configured to the classification of the classification determination destination object based on corresponding sorter.

Preferably, sorter determining unit is configured to the score being calculated input amendment by following equation:

$\mathrm{score}\left(E\right)=\mathrm{score}(e_0,e_1,...,e_n)={\mathrm{\Σ}}_{i=0}^n{\mathrm{\β}}_i\mathrm{\φ}\left(e_i\right)-{\mathrm{\Σ}}_{i=1}^n{d}_i{\mathrm{\φ}}_d\left(e_i\right)+b_E,$

Wherein, score (E) represents the score of positive sample E, E=(e ₀, e ₁..., e _n) represent sample, e ₀represent the overall template of sample, e _irepresent the local template of sample, φ (e _i) representation feature function, φ _d(e _i) represent the proper vector of current local template relative to the upper left corner displacement d of overall template, β _i, d _iand b _efor parameter, wherein, (dx _i, dy _i)=(x _i, y _i)-(2 (x ₀, y ₀)+v _i), (x ₀, y ₀) represent the top left co-ordinate of overall template, (x _i, y _i) represent the top left co-ordinate of current local template, v _irepresent (x _i, y _i) relative to (x ₀, y ₀) displacement.

Preferably, sorter determining unit is configured to the score by using following equation to correct each input amendment:

$s\left(x\right|A_E,B_E)=\frac{1}{1+\exp (A_E\·(\mathrm{score}\left(x\right)+B_E)},$

Wherein, s (x|A _e, B _e) represent that the score of each input amendment after correcting, score (x) represent the score of each input amendment before correcting, A _eand B _eit is the parameter obtained by stand-alone training.

Preferably, sorter determining unit is configured to the sorter corresponding to top score to be defined as corresponding sorter.

Preferably, described fundamental function is Harr fundamental function or HOG fundamental function.

Preferably, cutting unit is configured to carry out affined transformation by following equation:

P'=UP+V，

Wherein, P={p ₁, p ₂..., p _nrepresenting that the profile of corresponding sorter, U and V represent change of scale and coordinate position skew respectively, P' represents the profile of destination object.

Increase significantly according in the standard error of the sorter training method of exemplary embodiment of the present invention, target detection, segmentation or sorting technique and target detection, segmentation or sorter area, area under Receiver operating curve's (ROC curve), susceptibility, specificity, accuracy and F-measure, can be applicable to the various target detection of medical field, and can be applicable to other real-time systems and detect general object.

Accompanying drawing explanation

By the description carried out embodiment below in conjunction with accompanying drawing, these and/or other aspect of the present invention and advantage will become clear and be easier to understand, in the accompanying drawings:

Fig. 1 illustrates to be the process flow diagram of the sorter training method illustrated according to exemplary embodiment of the present invention according to exemplary embodiment of the present invention;

Fig. 2 is the process flow diagram that target detection according to exemplary embodiment of the present invention, segmentation or sorting technique are shown;

Fig. 3 is the block diagram that target detection according to exemplary embodiment of the present invention, segmentation or sorter are shown;

Fig. 4 illustrates and the profile of corresponding sorter is mapped to input amendment to obtain the example of the profile of destination object;

Fig. 5 is the diagram illustrating that the ROC curve of various detection algorithm compares;

What Fig. 6 and Fig. 7 illustrated the present invention and level set and markov random file dividing method compares diagram;

Fig. 8 is the diagram that quantitative segmentation result is shown.

Embodiment

More fully the present invention is described hereinafter with reference to accompanying drawing, exemplary embodiment of the present invention shown in the drawings.But the present invention can implement in many different forms, and should not be interpreted as being confined to proposed embodiment here.On the contrary, provide these embodiments to make the disclosure will be thoroughly with completely, and scope of the present invention is conveyed to those skilled in the art fully.

Although it should be understood that and term first, second, third, etc. can be used here to describe different elements, assembly, region, layer and/or part, these elements, assembly, region, layer and/or part should by the restrictions of these terms.These terms are only used to an element, assembly, region, layer or part and another element, assembly, region, layer or part to make a distinction.Therefore, when not departing from instruction of the present invention, the first element discussed below, assembly, region, layer or part can be referred to as the second element, assembly, region, layer or part.As used herein, term "and/or" comprises one or more combination in any and all combinations of lising of being correlated with.

Term used herein only in order to describe the object of specific embodiment, and is not intended to limit the present invention.As used herein, unless the context clearly indicates otherwise, otherwise singulative be also intended to comprise plural form.It is also to be understood that, " comprise " when using term in this manual and/or " comprising " time, there is described feature, entirety, step, operation, element and/or assembly in explanation, but does not get rid of existence or additional one or more further feature, entirety, step, operation, element, assembly and/or their group.

Unless otherwise defined, otherwise all terms used herein (comprising technical term and scientific terminology) have the meaning equivalent in meaning usually understood with those skilled in the art.It will also be understood that, unless clearly defined here, otherwise term (term such as defined in general dictionary) should be interpreted as having the meaning that in the environment with association area, their meaning is consistent, and will not explained them with desirable or too formal implication.

Hereinafter, the present invention is explained in detail with reference to the accompanying drawings.

Fig. 1 is the process flow diagram of the sorter training method illustrated according to exemplary embodiment of the present invention.

With reference to Fig. 1, in step S101, first obtain multiple training sample.Here, by using ultrasonic probe to obtain multiple training sample.Such as, in an experiment, from 395 ages by 24 to 81 user collect 480 mammary gland and have children outside the state plan images.All images are for be obtained from 2006 to 2010 years by Philip ATL iU22 ultrasound machine, and the ultrasonic probe of this machine is 5 to 12MHz, 6cm size.The size of each B-mode image is 1024x768 pixel, and spatial resolution is 0.23mm/ pixel.All images can be divided into training sample (exemplar) and test sample book at random.In described experiment, 320 images can be used as training sample.Each image can have one or more destination object not of uniform size (such as, tumour, tissue etc.).Each destination object carefully marks through experienced medical practitioner, and is verified by biopsy.Here, the detecting device based on exemplar can train an independently sorter for each training sample.But above example is only illustrative.More image can be adopted as training sample.For the destination object in each image, existing most of detection algorithm is all based on sliding window technique.By a linear function, such sliding window technique judges whether window (that is, ultrasonic probe window) comprises target object:

score(w)=β·φ(w)+b （1），

Wherein, φ (w) is the fundamental function (such as, Harr fundamental function or HOG fundamental function) extracted from window w.β and b represents weight vectors and straggling parameter respectively.

Next, in step s 102, by minimizing predetermined objective function, sorter is trained, thus obtain each sorter.Specifically, for each sorter, trained by the objective function minimized below:

L(β _E)=||β _E|| ²+h(score(E))+∑ _x∈N(E)h(-score(x)) （2）

Wherein, β _erepresent the weight vectors of sample, h (x)=max (0,1-x) represents hinge loss function, and score (E) represents the score of positive sample E, N (E) represents the set of negative sample, and score (x) represents the score of negative sample x.

More specifically, the error-detecting sample that obtains in normal picture (such as, tumor free image) for utilizing existing model in training process of negative sample N (E).In order to extract HOG feature, the window of each training sample under the condition keeping length breadth ratio, will be normalized to the cell of general 100 8x8 sizes in advance.But, from the window of the overall situation, only extract feature be not enough to various object (such as, tumour) to distinguish from the negative sample of large amount of complex.Therefore, need from high-definition picture, catch local detail feature.Such as, in star-structured model, sample E=(e ₀, e ₁..., e _n) overall template (e can be expressed as ₀) and n local (e _i) template.Such as, but not limited to this, overall template corresponds to whole window, and local template corresponds to the cell of 8x8 size one by one.Therefore, the score of each sample can calculate according to following equation (3):

$\mathrm{score}\left(E\right)=\mathrm{score}(e_0,e_1,...,e_n)={\mathrm{\Σ}}_{i=0}^n{\mathrm{\β}}_i\mathrm{\φ}\left(e_i\right)-{\mathrm{\Σ}}_{i=1}^n{d}_i{\mathrm{\φ}}_d\left(e_i\right)+b_E---\left(3\right),$

Wherein, E=(e ₀, e ₁..., e _n) represent sample, e ₀represent the overall template of sample, e _irepresent the local template of sample, φ (e _i) representation feature function, φ _d(e _i) represent the proper vector of current local template relative to the upper left corner displacement d of overall template, β _i, d _iand b _efor parameter, (dx _i, dy _i)=(x _i, y _i)-(2 (x ₀, y ₀)+v _i), (x ₀, y ₀) represent the top left co-ordinate of overall template, (x _i, y _i) represent the top left co-ordinate of current local template, v _irepresent (x _i, y _i) relative to (x ₀, y ₀) displacement.According to exemplary embodiment of the present invention, obtain optimum β by Latent support vector machine _i, d _iand b _e.

In training method as above, a series of sorter can be obtained, but the achievement of these sorters is incompatible, needs to regulate each sorter achievement.Therefore, in step s 103, use achievement fusion method to correct each sorter, thus make each sorter phase-splitting is compatible.Specifically, correct sorter by matching sigmoid distribution, redistribute each sorter achievement.After calibration, the exemplar mated most will obtain the highest sorter achievement.Specifically, by using equation (4) to correct each sorter:

$s\left(x\right|A_E,B_E)=\frac{1}{1+\exp (A_E\·(\mathrm{score}\left(x\right)+B_E)}---\left(4\right),$

Wherein, s (x|A _e, B _e) represent that the score of each sample after correcting, score (x) represent the score of each sample before correcting, A _eand B _eit is the parameter obtained by stand-alone training.According to exemplary embodiment of the present invention, in the training process, positive sample and doctor mark Duplication can be more than or equal to 0.7, and negative sample and mark Duplication are then less than or equal to 0.2.

Selectively, other achievement fusion methods also can be used for correcting each sorter.

By sorter training method as above, high-precision sorter can be obtained.Below, the target detection of the sorter applied according to sorter training method acquisition as above, segmentation or sorting technique and target, segmentation or classification and Detection device is specifically described with reference to Fig. 2 and Fig. 3.

Fig. 2 is the process flow diagram that target detection according to exemplary embodiment of the present invention, segmentation or sorting technique are shown.

With reference to Fig. 2, in step s 201, by the score for each classifier calculated input amendment in multiple sorter, and the score of each input amendment is corrected, obtain sorter compatible each other.Here, the image of user is obtained as input amendment by ultrasonic probe.The score of input amendment is obtained by above-mentioned equation (3):

$\mathrm{score}\left(E\right)=\mathrm{score}(e_0,e_1,...,e_n)={\mathrm{\Σ}}_{i=0}^n{\mathrm{\β}}_i\mathrm{\φ}\left(e_i\right)-{\mathrm{\Σ}}_{i=1}^n{d}_i{\mathrm{\φ}}_d\left(e_i\right)+b_E---\left(3\right).$

Selectively, the score by using equation (4) to correct each sample:

$s\left(x\right|A_E,B_E)=\frac{1}{1+\exp (A_E\·(\mathrm{score}\left(x\right)+B_E)}---\left(4\right).$

Here, the sorter corresponding to top score can be defined as corresponding sorter.

Next, in step S202, by affined transformation, the profile of corresponding sorter is mapped to the profile that input amendment obtains destination object.As mentioned above, for each sorter, mark its profile through experienced medical practitioner.According to exemplary embodiment of the present invention, carry out affined transformation by equation (5):

P'=UP+V （5），

Wherein, P={p ₁, p ₂..., p _nrepresenting that the profile of corresponding sorter, U and V represent change of scale and coordinate position skew respectively, P' represents the profile of destination object.Fig. 4 illustrates and the profile of corresponding sorter is mapped to input amendment to obtain the example of the profile of destination object, wherein, the left side of Fig. 4 illustrates detected image (input amendment) and detection window, the segmentation result that the right side of Fig. 4 illustrates the below sorter edge projection of top obtained to input amendment and the classification results obtained to input amendment by tumour category mappings corresponding for the sorter of top.

Finally, in step S203, can based on the classification of the classification determination destination object of corresponding sorter.That is, the classification of input amendment is consistent with the classification of the sorter determined in step S201.

Like this, the sorter of training in advance can be utilized simultaneously to complete detection, segmentation and classification, thus significantly improve the precision of target detection.

Fig. 3 is the block diagram that target detection according to exemplary embodiment of the present invention, segmentation or sorter are shown.

With reference to Fig. 3, described target detection, segmentation or sorter comprise sorter determining unit 301, cutting unit 302 and taxon 303.Sorter determining unit 301 is configured to by the score for each classifier calculated input amendment in multiple sorter, and corrects the score of each input amendment, obtains sorter compatible each other.As mentioned above, sorter determining unit 301 obtains the score of input amendment by above-mentioned equation (3) and passes through the score that equation (4) corrects each sample, then the sorter corresponding to top score is defined as corresponding sorter.Cutting unit 302 is configured to, by affined transformation, the profile of corresponding sorter is mapped to the profile that input amendment obtains destination object.Such as, cutting unit 302 carries out affined transformation by equation (5), the profile of corresponding sorter is mapped to input amendment to obtain the profile of destination object.Finally, taxon 303 is configured to the classification of the classification determination destination object based on corresponding sorter, that is, the classification of sorter sorter determining unit 301 determined is defined as the classification of destination object.

Below, Binding experiment data verify effect of the present invention.

In order to carry out Performance comparision with prior art, achieve (Zhang, the J.D. such as Zhang here, ZhouS.K., Brunke, S., Lowery, C., Comaniciu, D.:Database-guided breasttumor detection and segmentation in2D ultrasound images.SPIE, 7624 (3), 1-7 (2010)) detection algorithm, and six evaluation indexes provided by following table 1 carry out quantitative comparison detection perform.Result shows, and by introducing minutia, sorter can remove a large amount of error-detecting.In addition, not calibrated sorter is still better than boost and ESVM method.And the AUC value after correcting brings up to 85.77% from 82.25%.Fig. 5 is the diagram illustrating that the ROC curve of various detection algorithm compares.

Table 1

Here, prior art 1 represents Dalal, N., Triggs, B.:Histograms ofOriented Gradients for Human Detection.CVPR, 886-893 (2005), prior art 2 represents Drukker, K., Giger, ML., Horsch, K., Kupinski, MA., Vyborny, CJ., Menelson, EB.:Computerized lesion detection on breast ultrasound.Med.Phys., 29 (7): 1438-46 (2002), prior art 3 represents Dixon, A.M.:Breastultrasound:how, why and when.Elsevier Health Sciences (2007).In Table 1, AUC represents area under Receiver operating curve's (ROC curve), SE represents the standard error of this area, Sensitivity represents susceptibility, Specificity represents specificity, Accuracy accuracy, they and F-Measuer are the important technology indexs that those skilled in the art weigh classifier performance, repeat no more here.

For segmentation performance, here the present invention and level set and MRF dividing method are compared.What Fig. 6 and Fig. 7 illustrated the present invention and level set and MRF dividing method compares diagram.As shown in Figure 6 and Figure 7, the present invention difficulty (such as obscure boundary Chu or tumour itself due to calcification point cause uneven) in, significantly better than level set and MRF method.In figure 6 and figure 7, (a) illustrates detected image (input amendment), and wherein, irregular obstacle body is the profile of the manual mark of doctor, and (b) illustrates MRF segmentation result, and (c) illustrates level-set segmentation result, and (d) illustrates segmentation result of the present invention.

Fig. 8 is the diagram that quantitative segmentation result is shown.Here, have selected three evaluation indexes: dice likeness coefficient (DSC), false positive rate (FPR) and average absolute distance (MAD).Here, optimum sample set and pernicious sample set and be divided into two subsets at random.Fig. 8 illustrates that the present invention is better than additive method, particularly for the set comprising a large amount of difficult tumour.In fig. 8, transfer instruction is according to the result of the affined transformation of exemplary embodiment of the present invention, and levelset indicates the result of Level Set Method, and MRF indicates the result of markov random file.

For classification performance, the classify accuracy detecting correct sample is 87.58%, and the classify accuracy for all detections is 75.00%.It is worth mentioning that, the present invention only relies on training knowledge, and does not use any sorter.Can predict, by extracting the contour feature be partitioned into, design complex classifier, will improve classification performance.

Following table 2 illustrates that overall performance compares.

Table 2

As shown in above Fig. 5 to Fig. 8 and table 1 and table 2, according to the sorter training method of exemplary embodiment of the present invention, target detection, segmentation or sorting technique and target detection, segmentation or sorter overall performance outstanding, especially increase significantly in AUC, SE, Sensitivity, Specificity, Accuracy and F-measure.And have wide range of applications according to the sorter training method of exemplary embodiment of the present invention, target detection, segmentation or sorting technique and target detection, segmentation or sorter, can be applicable to the various target detection of medical field.Such as, under different mode (as MRI, CT etc.), in Different Organs (as liver, brain, lymph etc.), target (as tumour, knot, block etc.) detects.In addition, also can be applicable to other real-time systems according to the sorter training method of exemplary embodiment of the present invention, target detection, segmentation or sorting technique and target detection, segmentation or sorter and detect general object (as car, people etc.).

The above-mentioned sorter training method according to exemplary embodiment of the present invention and target detection, segmentation or sorting technique can be implemented as software or computer code or their combination.In addition, software or computer code also can be stored in non-transitory recording medium (ROM (read-only memory) (ROM), random-access memory (ram), compact disk (CD)-ROM, tape, floppy disk, optical data storage device and carrier wave (such as being transmitted by the data of internet)) in or by the computer code of web download, wherein, described computer code is initially stored in remote logging medium, computer readable recording medium storing program for performing, or non-transitory machine readable media also will be stored on local recording medium, thus method described herein can use multi-purpose computer, digital machine or application specific processor are to store such software on the recording medium, computer code, software module, software object, instruction, application program, applet, app etc. implement, or be implemented in programmable hardware or specialized hardware (such as ASIC or FPGA).As understood in the art: computing machine, processor, microprocessor controller or programmable hardware comprise volatibility and/or nonvolatile memory and memory assembly (such as RAM, ROM, flash memory etc.), wherein, described storer and memory component can store or receive software or computer code, wherein, described software or computer code will be will be implemented disposal route described herein by computing machine, processor or hardware access when performing.In addition, will recognize: when the code for being implemented on the process shown in this accessed by multi-purpose computer, multi-purpose computer is changed into the special purpose computer for being executed in the process shown in this by the execution of described code.In addition, program can pass through any medium (such as, by wire/wireless connect send signal of communication and equivalent) electronically transmitted.Described program and computer readable recording medium storing program for performing also can be distributed in the computer system of networking, thus store and computer readable code executed with the form of distribution.

Although shown and described some embodiments, it should be appreciated by those skilled in the art that without departing from the principles and spirit of the present invention, can modify to these embodiments, scope of the present invention is by claim and equivalents thereof.

Claims (19)

1. a sorter training method, is characterized in that, comprising:

Obtain multiple training sample, wherein, each training sample corresponds to a sorter;

By minimizing predetermined objective function, sorter is trained, thus obtain each sorter;

Use achievement fusion method to correct each sorter, thus make each sorter phase-splitting is compatible.

2. sorter training method according to claim 1, is characterized in that, described predetermined objective function is:

L(β _E)=||β _E|| ²+h(score(E))+∑ _x∈N(E)h(-score(x))，

Wherein, β _erepresent the weight vectors of sample, h (x)=max (0,1-x) represents hinge loss function, and score (E) represents the score of positive sample E, N (E) represents the set of negative sample, and score (x) represents the score of negative sample x.

3. sorter training method according to claim 2, is characterized in that, score (E) calculates according to following equation:

$\mathrm{score}\left(E\right)=\mathrm{score}(e_0,e_1,...,e_n)={\mathrm{\Σ}}_{i=0}^n{\mathrm{\β}}_i\mathrm{\φ}\left(e_i\right)-{\mathrm{\Σ}}_{i=1}^n{d}_i{\mathrm{\φ}}_d\left(e_i\right)+b_E,$

Wherein, E=(e ₀, e ₁..., e _n) represent sample, e ₀represent the overall template of sample, e _irepresent the local template of sample, φ (e _i) representation feature function, φ _d(e _i) represent the proper vector of current local template relative to the upper left corner displacement d of overall template, β _i, d _iand b _efor parameter,

Wherein, (dx _i, dy _i)=(x _i, y _i)-(2 (x ₀, y ₀)+v _i), (x ₀, y ₀) represent the top left co-ordinate of overall template, (x _i, y _i) represent the top left co-ordinate of current local template, v _irepresent (x _i, y _i) relative to (x ₀, y ₀) displacement.

4. sorter training method according to claim 3, is characterized in that, obtains optimum β by Latent support vector machine _i, d _iand b _e.

5. sorter training method according to claim 1, is characterized in that, corrects each sorter by matching sigmoid distribution.

6. sorter training method according to claim 5, is characterized in that, corrects each sorter by using following equation:

$s\left(x\right|A_E,B_E)=\frac{1}{1+\exp (A_E\·(\mathrm{score}\left(x\right)+B_E)},$

Wherein, s (x|A _e, B _e) represent that the score of each sorter after correcting, score (x) represent the score of each sorter before correcting, A _eand B _eit is the parameter obtained by stand-alone training.

7. sorter training method according to claim 3, is characterized in that, described fundamental function is Harr fundamental function or HOG fundamental function.

8. target detection, segmentation or a sorting technique, is characterized in that, comprising:

By for each classifier calculated input amendment in multiple sorter score and the score of each input amendment is corrected, obtain each other compatible sorter;

By affined transformation, the profile of corresponding sorter is mapped to the profile that input amendment obtains destination object;

Based on the classification of the classification determination destination object of corresponding sorter.

9. target detection according to claim 8, segmentation or sorting technique, is characterized in that, calculated the score of input amendment by following equation:

$\mathrm{score}\left(E\right)=\mathrm{score}(e_0,e_1,...,e_n)={\mathrm{\Σ}}_{i=0}^n{\mathrm{\β}}_i\mathrm{\φ}\left(e_i\right)-{\mathrm{\Σ}}_{i=1}^n{d}_i{\mathrm{\φ}}_d\left(e_i\right)+b_E,$

Wherein, score (E) represents the score of positive sample E, E=(e ₀, e ₁..., e _n) represent sample, e ₀represent the overall template of sample, e _irepresent the local template of sample, φ (e _i) representation feature function, φ _d(e _i) represent the proper vector of current local template relative to the upper left corner displacement d of overall template, β _i, d _iand b _efor parameter,

Wherein, (dx _i, dy _i)=(x _i, y _i)-(2 (x ₀, y ₀)+v _i), (x ₀, y ₀) represent the top left co-ordinate of overall template, (x _i, y _i) represent the top left co-ordinate of current local template, v _irepresent (x _i, y _i) relative to (x ₀, y ₀) displacement.

10. target detection according to claim 8, segmentation or sorting technique, is characterized in that, the score by using following equation to correct each input amendment:

$s\left(x\right|A_E,B_E)=\frac{1}{1+\exp (A_E\·(\mathrm{score}\left(x\right)+B_E)},$

Wherein, s (x|A _e, B _e) represent that the score of each input amendment after correcting, score (x) represent the score of each input amendment before correcting, A _eand B _eit is the parameter obtained by stand-alone training.

11. target detection according to claim 9 or 10, segmentation or sorting technique, is characterized in that, the sorter corresponding to top score is defined as corresponding sorter.

12. target detection according to claim 9 or 10, segmentation or sorting technique, it is characterized in that, described fundamental function is Harr fundamental function or HOG fundamental function.

13. target detection according to claim 8, segmentation or sorting technique, is characterized in that, carries out affined transformation by following equation:

P'=UP+V，

Wherein, P={p ₁, p ₂..., p _nrepresenting that the profile of corresponding sorter, U and V represent change of scale and coordinate position skew respectively, P' represents the profile of destination object.

14. 1 kinds of target detection, segmentation or sorters, is characterized in that, comprising:

Sorter determining unit, be configured to by for each classifier calculated input amendment in multiple sorter score and the score of each input amendment is corrected, determine each other compatible sorter;

Cutting unit, is configured to, by affined transformation, the profile of corresponding sorter is mapped to the profile that input amendment obtains destination object;

Taxon, is configured to the classification of the classification determination destination object based on corresponding sorter.

15. target detection according to claim 14, segmentation or sorter, is characterized in that, sorter determining unit is configured to the score being calculated input amendment by following equation:

$\mathrm{score}\left(E\right)=\mathrm{score}(e_0,e_1,...,e_n)={\mathrm{\Σ}}_{i=0}^n{\mathrm{\β}}_i\mathrm{\φ}\left(e_i\right)-{\mathrm{\Σ}}_{i=1}^n{d}_i{\mathrm{\φ}}_d\left(e_i\right)+b_E,$

Wherein, score (E) represents the score of positive sample E, E=(e ₀, e ₁..., e _n) represent sample, e ₀represent the overall template of sample, e _irepresent the local template of sample, φ (e _i) representation feature function, φ _d(e _i) represent the proper vector of current local template relative to the upper left corner displacement d of overall template, β _i, d _iand b _efor parameter,

Wherein, (dx _i, dy _i)=(x _i, y _i)-(2 (x ₀, y ₀)+v _i), (x ₀, y ₀) represent the top left co-ordinate of overall template, (x _i, y _i) represent the top left co-ordinate of current local template, v _irepresent (x _i, y _i) relative to (x ₀, y ₀) displacement.

16. target detection according to claim 14, segmentation or sorter, is characterized in that, sorter determining unit is configured to the score by using following equation to correct each input amendment:

$s\left(x\right|A_E,B_E)=\frac{1}{1+\exp (A_E\·(\mathrm{score}\left(x\right)+B_E)},$

Wherein, s (x|A _e, B _e) represent that the score of each input amendment after correcting, score (x) represent the score of each input amendment before correcting, A _eand B _eit is the parameter obtained by stand-alone training.

17. target detection according to claim 15 or 16, segmentation or sorter, it is characterized in that, sorter determining unit is configured to the sorter corresponding to top score to be defined as corresponding sorter.

18. target detection according to claim 15 or 16, segmentation or sorter, it is characterized in that, described fundamental function is Harr fundamental function or HOG fundamental function.

19. target detection according to claim 14, segmentation or sorter, is characterized in that, cutting unit is configured to carry out affined transformation by following equation:

P'=UP+V，

Wherein, P={p ₁, p ₂..., p _nrepresenting that the profile of corresponding sorter, U and V represent change of scale and coordinate position skew respectively, P' represents the profile of destination object.