CN101261735A - System and method for detecting an object in a high dimensional space - Google Patents
System and method for detecting an object in a high dimensional space Download PDFInfo
- Publication number
- CN101261735A CN101261735A CNA200710194441XA CN200710194441A CN101261735A CN 101261735 A CN101261735 A CN 101261735A CN A200710194441X A CNA200710194441X A CN A200710194441XA CN 200710194441 A CN200710194441 A CN 200710194441A CN 101261735 A CN101261735 A CN 101261735A
- Authority
- CN
- China
- Prior art keywords
- sorter
- candidate target
- trained
- target center
- space
- 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.)
- Pending
Links
Images
Abstract
A system and method for detecting an object in a high dimensional image space is disclosed. A three dimensional image of an object is received. A first classifier is trained in the marginal space of the object center location which generates a predetermined number of candidate object center locations. A second classifier is trained to identify potential object center locations and orientations from the predetermined number of candidate object center locations and maintaining a subset of the candidate object center locations. A third classifier is trained to identify potential locations, orientations and scale of the object center from the subset of the candidate object center locations. A single candidate object pose for the object is identified.
Description
The cross reference of related application
The application requires to submit on September 28th, 2006, sequence number is the rights and interests of 60/827,233 U.S. Provisional Application, and this application is incorporated in full with for referencial use.
Technical field
The present invention relates to be used for system and method, and more particularly, relate to and be used for using rim space study to come system and method at the higher dimensional space detected object at the higher dimensional space detected object.
Background technology
Many three-dimensionals (3D) detection and segmentation problem face at higher dimensional space to be searched for.For example, the 3D similarity transformation is characterised in that nine parameters: three location parameters, three direction parameters and three scale parameters.Searching for whole space comes detected object very expensive.Even relate to, search for all these parameters and on calculating, be under an embargo by thick strategy to essence.In addition, use and train the sorter that distinguishing ability is arranged very challenging, because the restriction of hardware once only allows the negative (negative) (about 10 of relatively little number at the positive example and the counter-example of the object that has so many parameter
6).In order to handle all possible counter-example, have to use multistage bootstrapping, thus the total system of making even slower.
But, natural in some way when the object that will detect on time, there is multiple situation.For example, the level of approximation and approximate have an identical colour of skin all of the most of faces in the picture.Similarly, the most of hearts in the CT scan are similar to and have identical size and Orientation.Need a kind of method, be used at higher dimensional space fast detecting object, wherein this search volume can be reduced greatly and still keeps result accurately.
Summary of the invention
A kind of system and method that is used at higher-dimension image space detected object is disclosed.Receive the 3-D view of object.First sorter is trained in the rim space of described object centers position, and it produces the candidate target center of predetermined number.Second sorter is trained the potential object centers position of from the candidate target center of predetermined number identification and direction and is kept the subclass of described candidate target center.The 3rd sorter is trained potential site, direction and the yardstick of the described object centers of identification from the subclass of this candidate target center.Discern the single candidate target attitude of described object.
Description of drawings
Preferred implementation of the present invention is described in more details with reference to the accompanying drawings hereinafter, the wherein same similar element of Ref. No. indication:
Fig. 1 is the block scheme that is used for being implemented in the example system of higher dimensional space fast detecting object according to of the present invention;
The projection that is used to use rim space to learn training classifier that Fig. 2 illustrates according to aspects of the present invention distributes;
Fig. 3 is the process flow diagram of describing according to the embodiment of the present invention that detects left ventricle in computed tomography images;
Fig. 4 illustrates to be performed the functional-block diagram that uses rim space study according to aspects of the present invention and detect the step of left ventricle;
Fig. 5 illustrates the example of having made (annotated) object of being explained;
Fig. 6 illustrates the functional-block diagram of LV centralized positioning method according to aspects of the present invention;
Fig. 7 be according to aspects of the present invention the LV center and the functional-block diagram of direction detection method; And
Fig. 8 illustrates the application of MSL in several other object detection problems of medical imaging.
Embodiment
The present invention relates to a kind of system and method that is used at higher-dimension image space detected object.Fig. 1 illustrates the system 100 at higher-dimension image space detected object of being used for according to an illustrative embodiment of the invention.Example in this explanation will be referred to detect anatomical structure in 3 d medical images.Yet, it will be understood by those skilled in the art that described method and system is not limited to detect anatomical structure, but can be used to detect other objects, and do not depart from scope and spirit of the present invention such as face, pedestrian, vehicle and traffic sign.As shown in fig. 1, system 100 comprises collecting device 105, personal computer (PC) 110 and operator's control desk 115, and this three connects by wired or wireless network 120.
Collecting device 105 can be computed tomography (CT) imaging device or such as any other three-dimensional (3D) high-resolution imaging equipment of magnetic resonance (MR) scanner or ultrasonic scanner.
Can be portable or the PC 110 at laptop computer, medical diagnostic imaging system or picture archive and communication system (PACS) data management station comprises CPU (central processing unit) (CPU) 125 and the storer 130 that is connected to input equipment 150 and output device 155.Described CPU 125 comprises rim space study module 145, and this rim space study module 145 comprises and will be used for detecting at 3 d medical images the method for anatomical structure at one or more that after this discuss.Although described rim space study module 145 is shown within the CPU125, described rim space study module 145 also can be positioned at outside the CPU 125.
The operation of described system 100 can be controlled by operator's control desk 115, and this operator's control desk 115 comprises controller 165 (for example keyboard) and display 160.Described operator's control desk 115 communicates with described PC 110 and described collecting device 105, and making can be by PC110 reproduction and observed on display 160 by collecting device 105 collected view data.When lacking described operator's control desk 115, by for example using input equipment 150 and output device 155 to move by some performed task of controller 165 and display 160, described PC 110 can be configured to operate and show the information that is provided by collecting device 105.
Described operator's control desk 115 can also comprise any suitable image reproducing system/tool/application, these image reproducing system/tool/application can be handled the Digital Image Data of the image data set (or its part) of being caught, to generate and display image on display 160.Or rather, image reproducing system can provide the reproduction and the visual application program of medical image, and this application program is moved on universal or special computer workstation.Described PC 110 also can comprise above-mentioned image reproducing system/tool/application.
Rim space study (MSL) utilizes most of data to have this fact of the constant attribute of some appropriateness (for example, the heart in the CT image is similar to and has identical size and Orientation).Similarly, detect in the rim space that remaining parameter is integrated away by only considering some parameter therein and get rid of most of space.Natural in some way when the object that will detect on time, also may there be multiple situation.Thisly natural distribute training classifier with as shown in Figure 2 projection to being used to, this projection distribute be present in less dimension rim space (for example, the space, left ventricle center of 3 parameters, rather than the left ventricle similarity transformation space of 9 parameters) on.Then, based on the rim space sorter of training, the search volume can be restricted to the material standed for that its projection to described rim space has the high probability value.
In MSL, sorter is trained on rim space, and in this rim space, wherein some variable is left in the basket.For example, go up the sorter of being trained at p (y) and can get rid of most of search volume fast.Another sorter is trained on its complementary space, to obtain final classification results.
Carry out therein now in the context of the example that in computed tomography (CT) image, detects left ventricle (LV) embodiments of the present invention are described.MSL is used to detect LV up to similarity transformation in 3D CT image.Fig. 3 illustrates and describes the process flow diagram that is used to detect the illustrative methods of LV according to of the present invention.Fig. 4 illustrates to be performed the functional-block diagram that uses rim space study according to aspects of the present invention and detect the step of LV.
Receive the image (402) of LV.Sorter is used to find 3D position x=(x, y, z) (step 302,404) at LV center.This sorter is trained in the rim space of LV center.For each 3D input data, 400 best positions are preserved for further assessment (step 304).The number of further assessing the position that is kept it will be understood by those skilled in the art that under the situation that does not depart from scope and spirit of the present invention, for can change.
For 400 position candidate, another sorter be used to keep the most promising 3D position of LV and direction (x, θ)=(x, y, z, θ
1, θ
2, θ
3) (step 306,406).Therefore, this sorter is trained in 6 dimension rim spaces of position and direction.Best 50 position candidate and direction are preserved for further assessment.It will be understood by those skilled in the art that under the situation that does not depart from scope and spirit of the present invention being preserved for the position candidate of further assessment and the number of direction can change.At last, trained listening group is used to detect LV until 9 dimension similarity transformations (just, position, direction and yardstick) (step 308,408)
T＝(x，θ，s)＝(x，y，z，θ
1，θ
2，θ
3，s
1，s
2，s
3) (1)
Single material standed for is the output (step 310) in this stage.
Training set comprises a plurality of image volume.The shape of each LV in the training image volume all uses the 3D grid that comprises 514 points to explain.Fig. 5 illustrates the example that utilizes the LV that 3D grid 502 and corresponding bounding box 504 explain.Described LV summit, A3C plane and main shaft are used to each shape aligned with each other.The shape of being aimed at is carried out principal component analysis (PCA) (PCA), and best 50 PCA basic points (base) are used to describe shape variable.
To the method described in Fig. 3 and Fig. 4 be described in more details now.In order to detect LV center 402, come training classifier based on 3D Ha Er (Haar) feature.At better data alignment and performance, the LV center is detected in by the image volume of double sampling (subsample) for the voxel size of 3mm.Under this resolution, volume has the size of approximate 50 * 50 * 60 voxels.Described training set comprises a plurality of image volume (for example, 100 image volume), and the size of all images volume all is adjusted to the 3mm voxel size.
3D Ha Er wave filter with the sampling location be the center, size is selected within the horizontal cube of 31 * 31 * 31 voxels.The set that about 10000 Lis Hartels are levied is selected for training.Increase the number of positive number by increasing the positive number that disturbs by 1-2 voxel, be used to stablize apart from its actual position.This allows about 6000 positive numbers to obtain from 100 training volumes.In the distance of at least 5 voxels in the real center of distance, negative is selected at random within image volume.About 1,000,000 negatives are used to training.
Probability raises, and (Probabilistic Boosting Tree PBT) is used to training and detecting to tree, and tree returns probability between 0 to 1 at each sample because this probability raises.Described PBT is to submit on March 2nd, 2006, sequence number is 11/366,722 and title be described in detail for the co-pending patented claim of " Probabilistic Bossting TreeFramework for Learning Discriminative Models (probability that is used for learning discerning model raise tree framework) ", this patented claim is incorporated in full with for referencial use.PBT utilizes five rank training, and preceding two ranks in these five ranks are implemented as cascade.
Fig. 6 graphic extension illustrates the functional-block diagram of basic LV center detection method.Input picture volume 602 is adjusted size with 3mm voxel resolution.The LV center is detected and is used the 3D Lis Hartel to levy in image volume 604.Described Lis Hartel levy by with the sample position be shown in the dotted line at center, size is selected in the frame of 31 * 31 * 31 voxels.Testing result is the stigma block element around real center 606.400 best positions are preserved for further assessment, and remaining position then is rejected.Other assessment indication of this level, all real centers are all in these 400 position candidate.This method reduces to 400 with the search volume of position from 50 * 50 * 60=150000, and this has reduced the magnitude greater than 300.
(x's second sorter works in 6 dimension spaces θ) in position and direction.(x, θ), (x, θ), position x is in 400 material standed fors that obtained by described position detector, and therefore this 6 dimension space is reduced more than 300 times at this value for this space value of being limited to.
Fig. 7 is the functional-block diagram of method of wherein having found the direction of described position.Find the position of direction to determine by the center detection method at it, this method obtains a string detected center 702.The LV direction detects uses 3D curvature feature to come the detection side to 704.It is particular voxel position in the frame of 24 * 24 * 24 voxels that each feature is calculated as size, by translation be rotated the position and the direction of sample.Within described frame, described position is organized on 9 * 9 * 9 dot matrix, therefore have wherein can calculated characteristics 729 different positions.There are 71 kinds of different combinations (summation, product, merchant, inverse trigonometric function or the like) of gradient, minimum and maximum curvature, principal direction and volume data each position in these positions.This provides 729 * 71=51759 feature.In addition, there are three yardstick: 3mm, 6mm and 12mm, calculate these features, thereby obtain about altogether 150000 features with these yardsticks.Positive number and negative are selected as position x in 400 material standed fors from the phase one.(x θ) is preserved for further assessment 706 to 50 best material standed fors.
The major issue that occurs is the significant distance that how to obtain between sample and the truth.This distance will be used to produce described positive number and negative, because those samples that approach truth really should be regarded as positive number, those samples that exceed certain threshold value then are negatives.
In order to calculate the distance apart from truth, (x θ) is increased the average dimension s that the statistical figure by described training data obtain to each sample
0Similarity transformation T=(x, θ, s that use obtains
0), obtain the reformed average shape of yardstick.Described average shape obtains when this (Procrustes) analysis is proposed in 98 training shapes execution Pu Luoke Lars.Therefore, the mean distance of the point-to-point between the corresponding reformed average shape of yardstick is two distances between the sample.
Described positive number is selected as having the distance that mostly is 6mm most, and described negative is being at least the distance of 15mm.By the output of described detecting device, (x θ) is held 50 best material standed fors, and remaining material standed for is rejected.
The dimension of search volume is increased once more, to increase the LV yardstick.Similarity transformation T=(x, θ s) are arranged in 9 dimension spaces, but preceding 6 the dimension (x, θ) be restricted to only get by on last stage the acquisition 50 values.Identical feature samples storehouse is as being used in the LV direction detection-phase, and difference now is, for each sample (x, θ, s), wherein the calculated frame of these features be now (x, θ, 4/3s).After this stage, select single material standed for.
Similar methods can be used to detect atrium sinistrum (LA).In order in 2D, to detect LA, search for 5 dimension spaces up to similarity transformation: the position (x, y), direction θ and yardstick (s, a) (yardstick and aspect ratio).Use the MSL method, three sorters are trained.First sorter is trained to detect the LA position.Because the LA in the ultrasound data shows changeability significantly in size, so also estimate very rough size with little, big three values of neutralization.1000 best material standed fors are preserved for further processing.Second sorter is trained at each material standed for to infer LA direction and a scale parameter.Then, Zui Jia 1000 material standed fors are preserved for further processing.Last sorter is trained at each the deduction aspect ratio in 1000 material standed fors, and the mean value of 20 best detections is reported as testing result.
MSL is method in common and can be used to a plurality of challenging 2D and the 3D object detection in the medical imaging and cuts apart task.Fig. 8 illustrate use MSL in radioscopic image detected catheter tip (802), in abdominal CT, detect ileocaecal sphineter (804) and liver (806), in ultrasonography, detect ventricle (808) and in MRI, detect ventricle (810).
The embodiment that is used at the system and method for higher dimensional space detected object has been described, but has it should be noted that those skilled in the art can make amendment and modification according to above-mentioned instruction.Therefore, should understand in disclosed specific implementations of the present invention and can change, these change as by claims within the defined scope and spirit of the present invention.Therefore, in detail also described the present invention particularly by the requirement of Patent Law, content required for protection and that need patent certificate to protect is illustrated in the claim of enclosing.
Claims (18)
1. method that is used at higher-dimension image space detected object, it comprises:
Receive the 3-D view of object;
Training first sorter in the rim space of object centers position, it produces the candidate target center of predetermined number;
Train second sorter to come from the candidate target center of predetermined number identification potential object centers position and direction and keep the subclass of described candidate target center;
Train the 3rd sorter to come potential position, direction and the yardstick of the described object centers of identification from the subclass of described candidate target center; And
The single candidate target attitude of identifying object.
2. the method for claim 1, wherein use in described image to be levied described first sorter is trained by the 3D Lis Hartel of double sampling.
3. method as claimed in claim 2, wherein, probability of use raises to set and trains described 3D Lis Hartel to levy.
4. the method for claim 1, wherein use the curvature feature to come described second sorter is trained.
5. method as claimed in claim 4, wherein, described second sorter is worked in the sextuple space of position and direction.
6. the method for claim 1, wherein described the 3rd sorter is trained in nine dimension spaces of position, direction and yardstick.
7. the method for claim 1, wherein described object is a left ventricle.
8. the method for claim 1, wherein described object is the atrium sinistrum.
9. method as claimed in claim 7, wherein, described 3D rendering is the 3D computed tomography images.
10. system that is used at higher-dimension image space detected object, it comprises:
Be used to catch the collecting device of the 3-D view of object;
Receive the processor of the 3-D view of being caught of object, described processor is carried out following step to every width of cloth image:
Train first sorter at the rim space of object centers position, it produces the candidate target center of predetermined number;
Train second sorter to come from the candidate target center of described predetermined number identification potential object centers position and direction and keep the subclass of described candidate target center;
Train the 3rd sorter to come potential position, direction and the yardstick of the described object of identification from the subclass of described candidate target center; And
The single candidate target attitude of identifying object;
And
Be used to show the display of detected object.
11. system as claimed in claim 10 wherein, uses in described image to be levied by the 3D Lis Hartel of double sampling described first sorter is trained.
12. system as claimed in claim 11, wherein, probability of use raises to set and trains described 3D Lis Hartel to levy.
13. system as claimed in claim 10 wherein, uses the curvature feature to come described second sorter is trained.
14. system as claimed in claim 13, wherein, described second sorter is worked in the sextuple space of position and direction.
15. system as claimed in claim 10, wherein, described the 3rd sorter is trained in nine dimension spaces of position, direction and yardstick.
16. system as claimed in claim 10, wherein, described object is a left ventricle.
17. system as claimed in claim 10, wherein, object is the atrium sinistrum.
18. system as claimed in claim 16, wherein, described 3D rendering is the 3D computed tomography images.
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US82723306P | 2006-09-28 | 2006-09-28 | |
| US60/827233 | 2006-09-28 | ||
| US11/856208 | 2007-09-17 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN101261735A true CN101261735A (en) | 2008-09-10 |
Family
ID=39962175
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CNA200710194441XA Pending CN101261735A (en) | 2006-09-28 | 2007-09-28 | System and method for detecting an object in a high dimensional space |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101261735A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106415606A (en) * | 2014-02-14 | 2017-02-15 | 河谷控股Ip有限责任公司 | Edge-based recognition, systems and methods |
| CN108225319A (en) * | 2017-11-30 | 2018-06-29 | 上海航天控制技术研究所 | The quick Relative attitude and displacement estimation system and method for monocular vision based on target signature |
-
2007
- 2007-09-28 CN CNA200710194441XA patent/CN101261735A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106415606A (en) * | 2014-02-14 | 2017-02-15 | 河谷控股Ip有限责任公司 | Edge-based recognition, systems and methods |
| CN106415606B (en) * | 2014-02-14 | 2019-11-08 | 河谷控股Ip有限责任公司 | A kind of identification based on edge, system and method |
| CN108225319A (en) * | 2017-11-30 | 2018-06-29 | 上海航天控制技术研究所 | The quick Relative attitude and displacement estimation system and method for monocular vision based on target signature |
| CN108225319B (en) * | 2017-11-30 | 2021-09-07 | 上海航天控制技术研究所 | Monocular vision rapid relative pose estimation system and method based on target characteristics |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US8009900B2 (en) | System and method for detecting an object in a high dimensional space | |
| US8135189B2 (en) | System and method for organ segmentation using surface patch classification in 2D and 3D images | |
| US10853409B2 (en) | Systems and methods for image search | |
| US7916919B2 (en) | System and method for segmenting chambers of a heart in a three dimensional image | |
| Zhang et al. | Intelligent scanning: Automated standard plane selection and biometric measurement of early gestational sac in routine ultrasound examination | |
| US8693750B2 (en) | Method and system for automatic detection of spinal bone lesions in 3D medical image data | |
| US20130223704A1 (en) | Method and System for Joint Multi-Organ Segmentation in Medical Image Data Using Local and Global Context | |
| US8363918B2 (en) | Method and system for anatomic landmark detection using constrained marginal space learning and geometric inference | |
| US20090028403A1 (en) | System and Method of Automatic Prioritization and Analysis of Medical Images | |
| US20110188715A1 (en) | Automatic Identification of Image Features | |
| US20090304251A1 (en) | Method and System for Detecting 3D Anatomical Structures Using Constrained Marginal Space Learning | |
| US9330336B2 (en) | Systems, methods, and media for on-line boosting of a classifier | |
| US7706612B2 (en) | Method for automatic shape classification | |
| US20080281203A1 (en) | System and Method for Quasi-Real-Time Ventricular Measurements From M-Mode EchoCardiogram | |
| RU2526752C1 (en) | System and method for automatic planning of two-dimensional views in three-dimensional medical images | |
| CN102132322B (en) | Apparatus for determining modification of size of object | |
| Ma et al. | Combining population and patient-specific characteristics for prostate segmentation on 3D CT images | |
| Kak et al. | Content-based image retrieval from large medical databases | |
| CN107833229A (en) | Information processing method, apparatus and system | |
| Unay et al. | Medical image search and retrieval using local binary patterns and KLT feature points | |
| CN101261735A (en) | System and method for detecting an object in a high dimensional space | |
| US7548642B2 (en) | System and method for detection of ground glass objects and nodules | |
| Huang et al. | Learning to segment key clinical anatomical structures in fetal neurosonography informed by a region-based descriptor | |
| Simonyan et al. | Immediate ROI search for 3-D medical images | |
| Davarpanah et al. | Brain mid-sagittal surface extraction based on fractal analysis |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20080910 |