CN101366059A - Cad detection system for multiple organ systems - Google Patents

Abstract

A computer aided disease detection system and method for multiple organ systems. The method performs computer aided examination of digital medical images. A patient exam type of a digital medical image is determined. Based on the patient exam type, one or more of a plurality of knowledge based anatomical segmentation blocks are invoked, each block performing image segmentation for a single organ or organ system present in the image. Based on the patient exam type, for each successfully segmented organ or organ system, one or more of a plurality of knowledge based computer aided detection blocks are invoked, each block of which is designed to search for and locate potential disease foci in a particular organ or organ system.

Description

The CAD detection system that is used for a plurality of tracts

Technical field

The present invention relates generally to the digital medical image process field, relates in particular to be used for the method and system that the medical image automated computer is checked, this method and system is in order to detect possible disease location in a plurality of human organ system.

Background technology

Technical progress in the noninvasive imaging field of human body has made the medical expert benefit aspect detection, diagnosis and the treatment disease ability, thereby brings the achievement of improvement and the reduction of M ﹠ M.Usually bring new application in the progress aspect imaging resolution, the functional and quality to the useful various imaging forms (modalities) of patient and society.Yet this progress causes the cost of manpower and financial cost, particularly imaging research and the cost that the image that produces is utilized and explains required necessary professional medical science technical skill aspect.In the past ten years, three peacekeeping four-dimensional form (for example, CT (computer tomography) art (CT), magnetic resonance imaging (MRI), single photon emission tomography (SPECT) and various other, quick sequential together with these imaging operations) quick growth has caused the raising of the quantity of view data, and has thereupon increased requirement and difficulty that the doctor lays down a definition to the image that is used for follow-up diagnosis.

The raising of technology requires a plurality of fields progressive abreast simultaneously usually.The ability that automated computer image understanding aspect reaches the recent technological advances level is being applied to the problem that medical image is understood.In a broad sense, " image understanding " is meant the ability of creating the computer based system, and the imitation human capabilitiy of these systems is to obtain (that is, non-statistical with non-numeric) information of semantic level from digitized image.Exist a large amount of examples, these examples comprise such as abilities such as facial detection and Identification and scene classifications.The enhancing that many researchs that relate to medical image robotization understanding now belong to domain-specialist knowledge (that is, anatomy and physiology) is used, and offers the notion level of the help of separating translator and other user of this image with raising.Because it is excessive that the clinical practice of radioactive ray medical science faces flourish information, so place hope on the ability of the high vision understanding technology that can help interpretation process.The TRIP proposal (changing radiological decipher process) that radiology computer utility association (SCAR) proposes, this technology is intended to encourage research, even also can improve ability, efficient and correctness from medical science Image Acquisition diagnostic message under the situation in the face of data rate, type and the complexity of increasing sharply to ancillary technique.

Some people thinks, clinical radiological practice is near " crisis " period, this time interim, quantity of increasing sharply in the medical imaging data and classification are threatening that to overwhelm available be essential professional decipher technical ability for its utilization.The research that TRIP proposal encourages to innovate seek promising technology and with people close friend's progress, to alleviate the incapability that can not handle the imaging data as the inundatory flood valuably.The potentiality that 3D rendering is handled are absorbed in this proposal especially, and expert's medical knowledge is attached in the display system lever as the efficient and the correctness that are hopeful to improve the actinology decipher.

In the task of may use medical image robotization understanding, a demand is included in the difficult task that detects disease in the big data quantity medical image.The commercialization of current existence and the computer aided detection of research grade (CAD) system, at the principal disease of tract, for example lung and colon cancer are worked with two kinds of different medical imaging forms of two and three dimensions.The purpose of this system is to improve the medical professional's (normally radiologist) of decipher imaging research correctness and time efficiency.About correctness, purpose is to improve the true detected ratios of abnormal diseases situation, and wrong discovery result is increased sharply.About time efficiency, the effect of these systems is to make reader's evaluation studies and do not lose correctness more quickly.Increase fast owing to the raising of image resolution ratio makes the data volume in the medical imaging research, back one purpose is a particular importance.

A resistance of accepting extensively that hinders CAD system relates to and stops the medical professional to distinguish to follow the law of finding the result and the responsibility of ethics,, detects the unusual pathology outside the principal focal point of research that is.The radiologist of for example, decipher virtual colonoscopy also must be noted that the anomaly in kidney and the liver.Up to now, CAD system focuses in the disease that detects in single organ or the tract (for example, colon), needs other reading time and searches for the outer finding of the colon with medical science significance.Therefore, limited the ability that this CAD system reduces the decipher time substantially widely.

With reference to following reference.

Be entitled as " System and Method for Assigning a Computer Aided DetectionApplication to a Digital Image ", the common U.S. Patent Application Publication NO.U.S.2004/0024292 that transfers the possession of people such as () Menhardt, about certain computer auxiliary detection algorithm being given the system and method for digital medical image based on the knowledge of inspect-type and imaging form.

People such as inventor Lou classify at inspect-type in that submit to, that be entitled as " Method for ClassifyingRadiographs ", the common U.S. Patent Application Publication NO.U.S.2006/0110035 that transfers the possession of on November 23rd, 2004.

The U.S. Patent Application Publication NO.U.S.2002/0164061A1 that people such as inventor Paik are disclosed on November 7th, 2002, be entitled as " Method for DetectingShapes in Medical Images ".

The U.S. Patent Application Publication NO.U.S.2002/0164060A1 that people such as inventor Paik are disclosed on November 7th, 2002, be entitled as " Method for CharacterizingShapes in Medical Images ".

Summary of the invention

The purpose of this invention is to provide the area of computer aided disease detecting system that is used for a plurality of tracts.

The present invention has disclosed a kind of area of computer aided disease detecting system that is used for a plurality of tracts, and it can reduce the reading time of the large scale anatomy imaging research in for example belly of two and three dimensions or thoracic cavity.

According to the present invention, a kind of method of area of computer aided inspection of combine digital medical image is provided, this method comprises the following steps: to determine patient's inspect-type of digital medical image; Under the supervision of checking control module,, call a plurality of one or more based in the dissection block (block) of knowledge according to patient's inspect-type; Each piece is cut apart the single organ or the tract carries out image that appear in the image; Under the supervision of described inspection control module, based on patient's inspect-type, for each organ of successfully cutting apart or tract, call a plurality of one or more based in the computer aided detection piece of knowledge, each in these pieces is that design comes to search for and locate concentrating on potential disease in certain organs or the tract.

In one embodiment, this method can comprise from the past by the result of report blocks to the certain organs considered or each all suitable report blocks as a result report in the tract, to be used for result's decipher.

The one or more step in a plurality of area of computer aided medical diagnosis on disease pieces is called in the supervision that the present invention can also be included in described inspection control module down with based on patient's inspect-type, in these pieces each designs the potential disease of estimating to concentrate in certain organs or the tract, comprises the possibility of actual disease process with the assessment organ.

Description of drawings

As shown in drawings, by the more specific description of following embodiments of the invention, above and other objects of the present invention, feature and advantage will become apparent.Element in the accompanying drawing each other must be not in proportion.

Fig. 1 illustrates explanation block scheme of the present invention.

Fig. 2 illustrates the block scheme that explanation is cut apart.

Fig. 3 illustrates the block scheme that explanation is cut apart.

Fig. 4 illustrates the problem of the linear separability with lineoid, wherein irises out the support vector with circle.

Fig. 5 illustrates the linear inseparable problem with lineoid, wherein irises out the support vector with circle.

Fig. 6 illustrates non-linear, the inseparable problem with classification surface, has wherein irised out the support vector with circle.

Embodiment

Be with reference to the accompanying drawings to the detailed description of preferred embodiment of the present invention below, in each of several accompanying drawings, identical label is represented the structure components identical.

Fig. 1 illustrates block scheme of the present invention.With reference to this accompanying drawing, digital medical image 10 appears in this method.Can be by in the known method of carrying out in the current medical domain any, for example the CT (computer tomography) art is created image.Preferably come presentation video by industrial standard DICOM form.

The phase one of handling comprises optional inspect-type classification step 20.The purpose of this step is the given patient dissection of determining in the present digital picture, in order that remaining processing is directed to specific organ or the tract that appears in the image.For example, the CT image packets of chest contains the information of relevant patient's lung, the heart and main blood vessel.The CT image packets of belly contains the information of relevant liver, stomach, big and small intestine, rectum, bladder and kidney.This step is optionally, because it can be by following any replacement: (1) is described as the inspect-type that the part of standard picture head data block (as in general DICOM form) provides significantly; Or the indication of the inspect-type of (2) user startup.

Regardless of being definite automatically or that determine by artificial input or definite by the resolution digital image head by classification step 20, inspect-type all offers checks control module 30.The purposes of checking control module is management block 40, detection/ diagnostics block 50 and 60, and the work of displaying block 70 as a result, they are suitable for certain organs or tract, and described organ or tract expection will be checked in the classification application various patients of the present invention and occur.

Then, under the supervision of checking control module 30,, call a plurality of the one or more of dissection block 40 and on digital medical image, operate based on knowledge according to the type of patient's imaging inspection.Because the knowledge of patient's inspect-type, each of these pieces is cut apart the known single organ or the tract carries out image that will appear in the digital picture.(in context, cut apart and be meant that an image pixel/voxel (voxels) marks and become to belong in certain organs or the tract or outer collection (being referred to as figure).Usually figure is expressed as binary picture, wherein non-zero pixels or voxel are represented interesting areas or structure, are referred to as prospect (foreground) traditionally.Zero pixel represents to be referred to as remaining image-region of background.In addition, cut apart and to comprise and create organ or the surface of tract or the digital geometry model of volume.) for example, will handle the CT image of chest so that respectively lung, the heart and main blood vessel are cut apart by block 41,42,43.To handle the CT image of bellies so that respectively liver, stomach, big and small intestine and kidney are cut apart by block 44,45,46 and 47.The result of block 40 is mark collection of pixel/voxel, and may be the surface or the volume-based model of each organ or tract.Owing to can not guarantee partitioning algorithm success, block also produces at potential organ or in each cut apart in the tract and shows the successfully control signal of degree.

Then, under the supervision of checking control module 30, according to the type of patient's imaging inspection with based on completing successfully that the dissection of knowledge is cut apart, with call a plurality of based on knowledge computer aided detection piece 50 and computer-aided diagnosis piece 60 in one or more.With provide digital medical image together with cut apart figure and be suitable for the optional surface of interested certain organs or tract or volume-based model as input to these pieces.For example, will provide colon (large intestine) to cut apart figure and digital geometry colon surface model as input to colon computer aided detection piece 56.The result of computer aided detection piece 50 be think the digital colon segmentation collection that may comprise superfluous natural disposition adenoma (neoplastic adenoma) or colorectal cancer (colorectal carcinoma) or surface or volume-based model the zone the locus and detect score.If diagnostics block 60 is available for certain organs of being considered or tract, check that then control module 30 can cause the operation of this module after finishing for the relevant detection module of same anatomical unit.For example, colon computer-aided diagnosis piece 66 will be handled digital medical image, colon segmentation figure and colon computer aided detection result, and each the potential candidate disease that is produced by colon computer-aided diagnosis piece 56 is produced numerical diagnostic.The sample result that produces in this situation comprises the size and the volume of estimation of candidate disease and classification and the optimum/pernicious diagnosis that is categorized into adenoma/carcinoma.In every kind of diagnosis situation, the confidence level score can be provided, this score is the real number in scope [0,1].Higher confidence level score provides bigger confidence level, and promptly the estimation diagnosis that produces by corresponding diagnostics block is real, and has medical importance.

After 60 work of computer aided detection piece 50 and computer-aided diagnosis piece, can be by report blocks 70 as a result to the result who is responsible for doctor or medical technology person's display system.These results can include but are not limited to, and are superimposed with expression position, diagnosis, and the organ of the visible mark of the confidence level of the finding of various processing block generation or the graphic presentation of tract model.Equally, can obtain result's numeral and form demonstration.Can show all these results together with the more familiar demonstration of original figure medical image, further study to allow medical personnel.

Still with reference to figure 1, about type of detection classification 20, optionally the effect of type of detection sort module 20 is medical science types of representing in automatically definite digital picture.The example of medical science type comprises the CT examination of chest, the CT examination and position (cranio-caudal) and lateral oblique position (medio-lateral oblique) the breast digital radiography projection end to end of belly.This module is optionally, because it can also provide by the artificial input of human operator or extract from the DICOM header that is associated with digital picture from the information that digital picture itself obtains.

Here quote as a reference, people such as Luo is in U.S. Patent Publication No.U.S.2006/0110021 that submit to, that be entitled as the common transfer of " Method forRecognizing Projection Views of Radiographs " on November 23rd, 2004, he it has disclosed the method for the perspective view (projectionview) that is used for automatically discerning the photography of X line.The purposes of this method is that the type of the radioactive ray imaging operation of digital picture that generation is provided is classified, and identification check in imaging at what patient anatomy.Although the method that this reference paper discloses relates generally to 2D radioactive ray imaging inspection, the technician in image understanding field can recognize, can directly the extend classification of digitized video of expression 3D imaging form of this method.Reach the purpose of inspect-type classification through the following steps: pre-service input digital image (optionally), the orientation of correcting digital image is extracted interesting areas from digital picture, and the inspect-type of discriminating digit image.The pre-service input digital image comprises original image is carried out sub sampling, makes image segmentation become prospect, background and dissection, and removes foreground area and make image intensity (intensity) normalization according to the characteristic of anatomy region from image.The orientation of correcting digital image comprises the orientation that detects digital picture and correspondingly makes the digital picture reorientation.Extract interesting areas from digital picture and comprise the intermediate shaft that detects dissection, determine center, size and the shape of interesting areas, and in radiograph, interesting areas is positioned.The perspective view of discriminating digit image is by about all possible inspect-type digital picture being classified, and the assembled classification result determines that the most probable inspect-type of representing in the digital picture finishes.

About checking control module 30, with reference to figure 1, the effect of checking control module 30 be arrange and monitor dissection block 40, be used for the computer aided detection piece 50 of organ/system based on knowledge, the finishing of report blocks as a result 70 that is used for the computer-aided diagnosis piece 60 of organ/system and is used for organ/system.Check the exam type information that control module 30 receives from inspect-type classification block 20, and use this information be arranged in each classification, relate to the certain organs that appears in the inspect-type or one or more processing of tract.If by known a plurality of organs or the tract of in digital picture, existing of exam type information, then detect control module and will be all suitable blocks 40, the computer aided detection piece 50 that is used for organ/system that belongs to organ or tract, the 70 arrangement processing of report blocks as a result that are used for the computer-aided diagnosis piece 60 of organ/system and are used for organ/system.Provide a concrete example, if the CT image of digital picture 10 expression chests, checking then that control module 30 will be called based on the dissection of large intestine knowledge cuts apart module 46, if and cut apart success, then be followed successively by large intestine computer aided detection module 56, large intestine computer-aided diagnosis module 66 and large intestine display module 76 as a result; Similarly, the similar triplets that are used for the module 47,57,67,77 of small intestine, and be used for 44,54,64,74 of liver, and or the like in the known type of detection that is present in by digital image representation, and all organs or the tract that have carried out successfully cutting apart.

Owing to dissect and to cut apart module and can fail because of a variety of causes, thus check that control module comprises steering logic, only to call the subsequent treatment module to organ or tract that success cuts apart occurring.The scope of cutting apart the reason that can fail for example can not be caught the whole anatomic region that is occupied by organ from collecting based on imaging, to the intrinsic reason of dissection of organ or tract itself, for example stops the colon evidence of folding (collapsed) cut apart fully.

Use an exemplary method to describe in detail now and cut apart 40 based on the dissection of knowledge, this method is used to use the dissection knowledge of priori to carry out cutting apart of organ or tract and surface/modeling volume.In order to effectively utilize dissection knowledge, must need to adjust cut apart with modeling make great efforts with the characteristic matching of certain organs or tract.Therefore, we have described the cutting apart and modeling algorithm of inside surface geometric model that human colon (large intestine) is cut apart and is provided in design particularly.It being understood that method described herein need revise the organ (for example, lung) to be successfully used to other inflation, and can need diverse algorithm for solid organ such as liver and main blood vessel.

Computer assisted detection and diagnostic method are used for promoting to be used for term waits the inspection colorectal cancer for " CT colon imaging (CTC) " or " virtual colonoscopy (VC) " noninvasive method just day by day.In the U.S., colorectal cancer (CRC) is second main cause of cancer mortality.For all people more than 50 years old and 50 years old of the U.S., formally simple and almost completely effective crc check is made in suggestion.The inspection method of standard comprises the endoscopy (optics Sigmoidoscope (0C)) of the colon inside surface of being done by gastroenterologist.If detected words then just just can be excised suspicious polyp and actual cancer usually when evaluation.Yet the mortality ratio of cancer is still very high, because screening test is uncomfortable, and comprises with catharsis and cleans colon excreta content.Therefore, studying the Noninvasive possible alternative of CTC, seeing whether the extensive enforcement of this method can reduce the CRC mortality ratio.Implement CTC and both can prepare method, also can clean without defaecation with the identical intestines of 0C.In the later case, remaining excreta must mark with the impenetrable material of radiation, can correctly discern this material like this in the CT image, and digitizing ground " by cutting ", in order that can not hide possible disease location.

With reference now to the block scheme of Fig. 2 and 3,, the processing 4610 of method 46 that the human colon is cut apart from the suitable representational plane of image, rectal area being cut apart.With reference to figure 3, in step 461010, select representational rectum section.After the image plane in the expection z axle zone that shows rectum, can carry out this selection by human user, or search for and automatically perform by the z axle zone of maximum coupling part (in addition, this part does not contact with the outer rim of image) to the digital code value of expression inflation area.Then, in step 461020,, upper limit threshold discerns inflation area in the section by being set with the digital value that is known as the value that surpasses general expression air.For example, in the CT image, the threshold value of 200 Hounsfield units is enough.By step before at first making to cut apart template (mask) reverse, use the known mathematical morphology computing of those skilled in the art to fill endoporus then, thereby in step 461030, the whole anatomic region of section cut apart.In step 461040, use the logical computing to make up the result of preceding two steps, cause cutting apart of independent rectal area.At last, in step 461050, the applied mathematics morphology operations is filled the inner hole of rectal area itself once more.

With reference to figure 2, according to the statistical value of the rectal area that obtains in the step 4610, in step 4620 with the expression air cutting apart of all possible voxel proceed colon segmentation.Specifically, in the rectal area through cutting apart of selected section, respectively by specifying r _MinAnd r _MaxMinimum and the maximum code value of expression, code value is observed:

r _min≤v≤r _max

The selected part of cutting apart as air section of all images voxel v.

Then, in step 4630, dividing processing identification is through the fluid mass of mark.Discern these zones with having the known digital value that drops under the value of generally representing fluid as lower threshold through marking.For example, in the CT image, the threshold value of 1500Hounsfield unit is enough.

Because so-called " partial volume effect " that causes by the finite resolving power of imaging operation, air, tissue and between the excreta of mark at the interface, be correctly the cutting apart of colon surface suitable difficulty.Volume in the imaging region of being made up of material blends can cause the centre of digital code value in those values of any individual substance.In step 4640, manage to proofread and correct at colon and the partial volume effect at the interface between the fluid mass of mark.(air and the interface between the flowing material of mark are not very important for the detection task).For this reason, cutting apart the morphology dilation operation of carrying out on the figure for colon and material through marking.In each case, only on single direction, occur expanding-for through the material of mark with respect to gravity downward (thereby entering tissue), and make progress with respect to gravity for organizing.The typical sizes of the structural element that is used to expand is one or several voxels at most.Then, calculate two common factors between the figure cut apart through expanding.This intersection area is represented some voxels, and wherein part and parcel volume effect and interruption disease detection process may occur together.In step 4650, these zones are added get back to colon and dissect and cut apart.

Then, in step 4650, the whole anatomic region of imaging volume is cut apart, used the mathematical morphology computing to fill internal holes then, comprise the inner chamber of colon by making to cut apart oppositely from the air section of step 4620.In step 4660, by to cut apart actuating logic as a result and the computing with the logical OR computing of cutting apart from the relevant portion volume area of step 4640 with from total air section of step 4620 from the whole dissection of step 4650, create cutting apart of whole colonic lumen.At last, in step 4670, use the standard operation of mathematical morphology fill colon regions cut apart in any remaining hole.

In step 4680, to (marking the colonic lumen zone now with binary one by what complete scale-of-two 3D cut apart that figure represents, and marking all other zones with Binary Zero) the whole zone of the imaging volume contour surface (iso-surface) that carries out standard handles, to create the geometric model of colon inside surface.It will be apparent to one skilled in the art that and to use the contour surface of cutting apart figure obtain value 1/2 of standard " mobile cube (marching cubes) " algorithm from considering as scalar field.Geometric model comprises the polygon surface mesh that contains summit (vertices), and this mesh comprises some summits and becomes polygonal their set.Except surface model, full volumetric scale-of-two colonic lumen is cut apart the output result that figure also is this step.

In step 4690, in each place, summit gauging surface gradient of colon surface model.Use known means from the raw digital image data compute gradient.

The computer aided detection method 50 that is used for organ or tract is described now in more detail.Especially, be used for detecting the method for the potential disease position of medical image as shown in Figure 1.It must be understood that when being applied to detect lung or colon cancer disease location (piece 51 and 56) situation, this method also has effectiveness widely for other potential use.Here this method is described as the single example of the computer aided detection block type that the present invention can comprise.Be noted that the computer aided detection method that can comprise many other types is as a part of the present invention.

Detection method is based on such notion, promptly, in the situation of lung, or in the situation of colonic lumen (empty inner space), interested disease target, colon or lung cancer generally comprise spherical or irregular spherical piece of tissue, and these piece of tissue show higher x ray attenuation than normal soft tissue (lung tissue).Therefore, in these disease situations, consider that digital picture is the mathematics scalar field, calculate the mathematics gradient vector of x ray attenuation field, can see that the gradient vector adjacent with spherical disease body sphere or irregular will trend towards as one man being orientated, thereby point to the barycenter of the volume of this body together.

In people's such as Paik the U.S. Patent Application Publication NO.U.S.2002/0164061A1 that is entitled as " Method for Detecting Shapes in Medical Images ", the method that is used for detecting by the incident of locating accumulation surface normal vector in the locus of colon polyp of colon has been described.Although this method might be sought the disease target on the colon inside surface, its difficulty is detected surface normal lap position is mapped to the definite position of colon surface, and this makes that this overlapped surfaces normal vector is risky.Therefore, the candidate's of description searching now polyp of colon improves one's methods.

Now continue to be described in this detection method in the particular case of colon, can be modeled as long hollow tubular or cylindrical to human colon's anatomical structure.Though in fact, in human body, be that distortion is twined and spiral that still, this organ inside surface is with the appropriate mathematical model of inner polyp still topological circle cylindricality just can see the substantially cylindrical shape of colon.Colorectal cancer and its tendency adenoma neoplasm (precursoradenomatous neoplasm) occur on the inside surface of colon organ.Therefore, in this case, check that the state of the digitized video in the colon inner surface area is just enough.As the result who handles colon segmentation module 46, colon surface cut apart figure and the model both can occur.This detection method structure is (for example, in this case, for general important clinical damage, size is about 10mm), the vector vertical with locating surface of regular length with respect to the disease candidate's who is searched for size.Spherical or irregular spherical disease composition place are occurring, the regular length vector should trend towards in groups consumingly at the volume center place near structure.Near normal anatomical structures, such as the colon inside surface of part plan, or the local half-cylindrical haustrum on the similar face is folding, and vector can be so closely not in groups.In disclosed technology, also has other detection method based on this principle, for example, see D.S.Paik, C.F.Beaulieu, G.D.Rubin, B.Acar, R.B.Jeffrey, Jr., J.Yee, J.Dey, and S.Napel " Surface Normal Overlap:AComputer-Aided Detection Algorithm with Application to Colonic Polyps andLung Nodules in Helical CT ", IEEE Trans.Med.Imag., Vol.23, pp.661-675, June 2004.

By the center that identification normal vector group is centered on, at least one group of this detection method identification normal vector.This method is based on notion so, that is, each center that is called barycenter is associated with the subclass of surface vertices, and these surface vertices produce around barycenter normal vector in groups.Can use the modification of Linde, Buzo and Gray (LBG) vector quantization (VQ) algorithm for design to calculate 2 iteratively ⁿ(n=0,1,2 ..., nsplit) individual barycenter, these are the points (representative value of nsplit is 8) in the 3d space that centers on of surface normal vector group.The xyz position of barycenter k has the N in its group _kThe minimum mean square distance of vector

(may be interpreted as geometrical line cuts apart).In first iteration, all vectors are all distributed to single group.In each follow-up iteration, make each already present barycenter split into two by the single position of disturbance barycenter a little; By Euclidean distance vector is reallocated to nearest barycenter again; And recomputate centroid position.

For each barycenter k (k=1,2,3 ..., 2 ^Nsplit+1-1), can come computing computer auxiliary detection (CAD) score S with following 1-D Gaussian function _k:

$S_k=\frac{N_k}{\mathrm{\σ}\sqrt{2\mathrm{\π}}}\·e(-\frac{D_k^2}{2{\mathrm{\σ}}^2})$

The size (representative value is 1 or 2) of can operation parameter σ selecting distinct group.Maximum CAD score identification is as the most interested barycenter of potential disease point.

Describe in detail now and be used for the method for the computer aided diagnosing method 60 of organ or tract about diagnostic medicine image potential disease position.This method will notice that this method has effectiveness widely for other potential application in the situation of the disease location that can be applicable to diagnose lung or colon cancer (piece 61 and 66).This method is described as comprising the single example of computer-aided diagnosis block type in the present invention here.Be noted that the computer aided diagnosing method that can comprise its many other types is as a part of the present invention.

Diagnostic method at first obtains candidate shape from the detection step of the barycenter of determining normal vector in groups, and the surface is associated with candidate shape, and this candidate shape comprises unique tabulation of the surface vertices that is associated with in groups all barycenter in candidate shape.Characterizing method (characterization method) calculates the one or more parameters for the surface of candidate shape then.According to these parameter values, this method determines that whether candidate shape is corresponding to interested shape.The detailed description of the preferred embodiment of characterizing method will be described now.To expand the previous explanation that in the 3-D volumetric image, detects the detection method of shape now, so that characterizing method to be shown at the anatomical surface place.

First step is that the CAD score of barycenter is shone upon back on the surface vertices.This can be that the CAD score of maximum realizes by distributing to each summit for nsplit+1 barycenter (surface normal with described barycenter in groups).This produces the subclass of the barycenter that is associated with surface vertices.

Next procedure is to seek CAD collision (hit).This can realize by making the integrated subgroup of barycenter that be associated with surface vertices and more than threshold value CAD score, like this so that (for example, distance 10mm) connects the barycenter of giving grouping less than fixed threshold.

The surface that can calculate each CAD collision is as unique tabulation of the surface vertices that is associated with in groups all barycenter in this collision.Can calculate this surperficial parameter with the generation characteristic of division, thereby this characteristic of division is determined the feature of each candidate shape and is helped to determine that whether it is corresponding to interested shape except the CAD score.These parameters can comprise, but be not limited to, from the candidate parameter of following tabulation (this tabulation comprise appearance with colon cancer be closely related most the surface parameter of connection): the 3D shape index, concentration gradient (gradient concentration), the CT value, bendability (curvedness), gradient amplitude (gradientmagnitude), the Hough distortion, edge dislocation (edge displacement), streamline convergence (streamline convergence), the best-fit radius of circle, best-fit circle remaining (best fit circleresidue), best-fit secondary remnants (best fit quadratic residue), parallel lines angle (parallel line angle), best-fit parallel lines remnants, the 3rd rank square invariant (3 ^RdOrder momentinvariant), sphericity, Gaussian curvature, areal concentration, summit counting (vertex count), wall thickness, volume, average neck (neck) curvature.

Then can according to the standard machine learning method (such as at Duda, R.O., Hart, P, and Stork, D.G., Pattern Classification, 2 ^NdEd., John Wiley and Sons, New York is described in 2001) carry out actual disease candidate diagnosis.For integrality, general sorting machine is described below; Support vector sorter (SVM), this is a preferable sorting algorithm of the present invention.

Describe the result now in more detail and report 70.

After handling dissection block 40, be used for the computer aided detection piece 50 of organ/tract and be used for the suitable set of computer-aided diagnosis piece 60 of organ/tract based on knowledge, check will to cause control module as a result the suitable collection of report blocks starts that each is used for the organ the disease candidate checked by former piece or each of tract.These as a result each of report blocks will provide target organ or tract by the mode of the particular anatomical that is suitable for its target cut apart, detection and diagnostic result.

Describe in detail now and be used for reporting the method for medical image potential disease position.This method can be applicable to detect and diagnose the situation of colon cancer disease location (piece 77).This method is described for comprising the single example of the type of report blocks as a result in the present invention here.Should be noted that the reporting step as a result of many other types can be included in the part of the present invention.

For the result of CT colon imaging inspection is offered medical science decipher professional, need provide multiple means, so that primitive medicine view data and computer aided detection and diagnostic result are visual.Current, in medical circle to mainly being that 2D or 3D rendering show and decipher is suitable for CTC most and exists arguement.For this reason, reporting modules 77 can provide the 2D image of colon inside surface to show and the demonstration of 3D graphical model simultaneously as a result.Two kinds of demonstrations all comprise the mark and the note of the disease target detection of system, and measured value and other instrument, separate the translator with raising and distinguish true and the existing ability of wig.At least one commercial company provides exemplary show tools, for example, and the Viatronix of New York Stony Brook.

Support vector machine device sorter is described now.The present invention uses the many sorters according to the principle of support vector machine device (SVM).SVM is the conventional data pattern classifier, and these sorters learn to carry out their classification task through the training example of mark in a large number according to what show to them.Machine is used to expand the interior kernel method of linear classification algorithm.Generally, at the vector space (its dimension coding mapping target signature) of some high dimension, SVM carries out linear lineoid sorter.In the present invention, the pixel value that the scope of target " feature " can be from image window is to more complicated calculating (for example, the vector of Gabor filter response or wavelet conversion coefficient).Can train SVM as input vector with these combination of features.The simple development of the training that the present described SVM of being is used to classify and the ultimate principle of application.

Linear SVM is known.In some cases, data can be divided, and really not so in other situation.

About situation about can divide, the simple scenario of svm classifier device is that the data of two classes that expression is separated linearly by a plane (or the lineoid in higher dimension) are trained and classified.Consider training data { x _i, y _i, I=1...1, y _i∈ 1,1}, x _i∈ P ^d, y wherein _iIt is the classification mark.Separately the lineoid of data satisfies:

w·x+b＝0 S1

The purpose of training SVM is to determine free parameter w and b.Fig. 4 illustrates R ²In the situation of dividing.

Always can go up the ratio (scale) that ratio puts on w and constant b so that all data all to observe inequality right:

w·x _i+b≥+1，y _i＝+1

S2

w·x _i+b≤-1，y _i＝-1

This one-tenth capable of being combined:

$\begin{array}{cc}{y}_i(w\·x_i+b)-1\≥0,& \∀i\end{array}$ S3

By proper proportion is set, might need at least one trained vector to satisfy each of equation (S2).Then, nargin (margin) (being defined as the vertical range between the lineoid that satisfies (S2) in the situation of equating) is 2/||w|| definitely.The purpose of SVM training is to make nargin big as much as possible, and this is by reducing to suffer constraint (S3) || and w|| realizes.For convenience's sake, let us is considered to minimize alternative, equivalently, objective function || w|| ², suffer constraint (S3) once more.(it is to protrude (convex) that this selection makes objective function).Optimization problem falls into classical protrusion optimization field (being also referred to as quadratic programming) then.Use the method for lagrange's method of multipliers, obtain having positive multiplier α _iLagrangian function.

$L=\frac{1}{2}{\left|\right|w\left|\right|}^2-\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{a}_i{y}_i(x_i\·w_i+b)+\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i$ S4

Under the constrained optimization theory, we must make objective function minimize, and requiring simultaneously becomes zero about the derivative of the Lagrangian function of all multiplier α, and same α 〉=0.

Optimization problem (S4) is protruding (convex) programming problem: its objective function || w|| ²Be protruding, the territory of α also is protruding, is limited in the positive quadrant.For these problems, there is the other form that is referred to as Wolfe antithesis (dual), wherein substitute by the equality constraint on the Lagrange's multiplier in the restriction of the inequality on the primary variables (S3).(S4) Wolfe antithesis needs us to make L maximization, and suffering derivative with respect to the L of α and major parameter w and b is zero restriction.Carry out differential and produce two constraints:

$w=\underset{i}{\mathrm{\Σ}}{a}_i{y}_i{x}_i$

S5

$\underset{i}{\mathrm{\Σ}}{a}_i{y}_i=0$

To retrain (S5) substitution (S4) and provide the following two Lagrangian formulations that are formulated again, and have and training data x only in dot product, occurs _iInteresting characteristic.This allows the SVM method to expand to the situation that data non-linearly can be divided.

$L_D=\underset{i}{\mathrm{\Σ}}{a}_i-\frac{1}{2}\underset{i,j}{\mathrm{\Σ}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_j{x}_i\·x_j$ S6

Because the advantage of its structure, the Hessian matrix of this optimization problem is a positive definite.For original protruding optimization problem (S4), can set up one group of necessary and sufficient condition that is called the KKT condition.The KKT condition comprises that the ormality of original inequality restriction, Lagrange's multiplier and Lagrangian function (S6) partial derivative must be zero requirement about all primary variabless.Enumerate these conditions below.

$w=\underset{i}{\mathrm{\Σ}}{a}_i{y}_i{x}_i$

$\underset{i}{\mathrm{\Σ}}{a}_i{y}_i=0$

S7

$\begin{array}{cc}{y}_i(w\·x_i+b)-1\≥0,& \∀i\end{array}$

α≥0

${\mathrm{\α}}_i\begin{array}{cc}(y_i(w\·x_i+b)-1)\≥0,& \∀i\end{array}$

Whether the KKT conditions permit is checked any individualized training situation, consistent with separating of optimization to observe its multiplier value.Final condition is called supplementary condition.It states that inactive constraint (not being in those constraints that their feasible zone boundaries are located) must have zero multiplier.Otherwise, can upset constraint and may move to objective function on the wrong direction.

The particular algorithm of separating quadratic programming problem (S6) is known, and occurs as the standard capability of software library and instrument (such as MATLAB).Yet according to the big or small l of training set, training is calculated and can be surpassed apace or even the ability of optimality criterion software program.For example, if l=50000, this is rational value for the face detection situation, and then the Hessian of quadratic problem (second derivative) matrix needs l ²=2.5 * 10 ⁹Individual (entry).For this reason, the increase that SVM uses has caused being used for the development of " (divide and conquer) divides and rule " algorithm of optimization step.

During training process, parameter alpha is interesting especially.With zero α that finishes _iThose trainings, for judging that by (S7, first party formula) lineoid w has no contribution.These situations are incoherent, and can abandon from training set and have no effect.α _iFor the situation of non-zero is called the support vector, this situation influences sorter really.Being contemplated to be of training time: really to suitable little part in the contributive just training in final decision surface.In fact, the result who obtains from Statistical Learning Theory shows that more little as the part of the training of supporting vector, the universal performance of sorter is good more.If many or most of trainings are to support vector, then sorter may just be remembered its training data, and the unitized hope of success is very little.(this situation being called " crossing training ").

When solving optimization problem, can write out in the expression formula of the w that provides in (S7, first party formula) and insert in the equation (S1) that lineoid is classified according to support vector with nonzero coefficient, be used for to provide the SVM decision function:

$f\left(x\right)=w\·x+b=\underset{i=1}{\overset{l_s}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i{x}_i\·x+b$ S8

L wherein _sBe the quantity of supporting vector.

New vector x one of is assigned in two classifications the sign (sign) that is based on decision function.

About non-situation of dividing, SVM expand to when data be not that the situation of linear separability need be introduced non-negative slack variable ζ _i, and the parameters C that is used for the classification error on punishment (penalize) training data.Notice that slack variable is not a Lagrange's multiplier.Substitute inequality (S2) by following formula:

w·x _i+b≥+1-ζ _i，y _i＝+1

S9

w·x _i+b≤+1+ζ _i，y _i＝-1

ζ≥0

Slack variable (it be can't see in the training process outside at SVM) can remain on zero giocoso, unless classification error appears on training set, when they must get value greater than one (unity).Then, for punishment being awarded these mistakes, substitute objective function by following formula || w|| ²:

$\frac{{\left|\right|w\left|\right|}^2}{2}+C\underset{i}{\mathrm{\Σ}}{\mathrm{\ζ}}_i$ S10

The particular value of employed C is provided with the relative importance of classification error on the training data.Upper limit of value only by giving multiplier α, the introducing of slack variable has influenced the Wolfe dual representation of training problem:

0≤α≤C S11

Whenever trained vector has been carried out wrong classification, its corresponding multiplier α will the value of being limited to C.The SVM decision function is not subjected to the influence of slack variable.

Slack variable has bigger influence to the KKT condition.Because they are not Lagrange's multipliers, the viewpoint of optimizing from restriction be it seems so, and they are the primary variabless with inequality restriction.Therefore, the essential new collection of introducing non-negative Lagrange multiplier μ causes following some difficult KKT condition set:

$w=\underset{i}{\mathrm{\Σ}}{a}_i{y}_i{x}_i$

$\underset{i}{\mathrm{\Σ}}{a}_i{y}_i=0$

$\begin{array}{cc}C-{\mathrm{\α}}_i-{\mathrm{\μ}}_i=0,& \∀i\end{array}$

$\begin{array}{cc}{y}_i(w\·x_i+b)-1+{\mathrm{\ζ}}_i\≥0,& \∀i\end{array}$

S12

α≥0

ζ≥0

μ≥0

${\mathrm{\α}}_i\begin{array}{cc}(y_i(w\·x_i+b)-1)=0,& \∀i\end{array}$

$\begin{array}{cc}{\mathrm{\μ}}_i{\mathrm{\ζ}}_i=0,& \∀i\end{array}$

Once more, two final conditions are complementarity conditions of inertia constraint.Notice that in case trained sorter, the user of SVM both cannot see lagrange's variable and also cannot see slack variable.But these variablees are interim artefacts of training process.

Fig. 5 illustrates linearity, inseparable situation.The decision function of SVM is as keeping in (S8).

Non-linear SVM is described now.

Recognize that the task of face detection can not cause the linear separability problem, no matter selected feature set.Fortunately be that by introducing the method for kernel function, the Wolfe dual representation of training problem (S6) causes the almost trivial extension to nonlinear situation.Consider mapping function Φ: the natural vector space P of input data from its (low) dimension d ^dBe mapped to (possibility) more Λ → H of the Space H of higher-dimension.Make the people have some surprised, H even can have infinite dimension though this situation is delivered to the mathematics travelling in the unfamiliar territorial waters, and ignores it here.Suppose that we at first are mapped to Space H to data via function phi before applying above-mentioned SVM method.So, because original training data occurred in dot product in the past only, the training data of mapping also is so now, and these data only appear in the dot product among the H now, and this is form Φ (x _i) Φ (x _j) function.

We introduce kernel function now, and there is such characteristic in assumed function K:

K(x _i，x _j)＝Φ(x _i)·Φ(x _j) S13

That is, when being applied to lower dimensional space, function K provides the identical scalar result of dot product that is mapped to the argument amount of higher dimensional space with it.Then, data will appear in the training process, and decision function is the argument amount of function K, there is no need to use mapping Φ at training period.Whenever in the time of in the equation before the dot product of two vectors appears at, with K (x _i, x _j) substitute it.Simple case from [12] has Φ: P ²→ P ³, and have:

$\mathrm{\Φ}\left(x\right)=\left(\begin{array}{c}{x}_1^2\\ \sqrt{2}{x}_1{x}_2\\ x_2^2\end{array}\right)$ S14

And K (x then _i, x _j)=(x _iX _j) ²

Clarified the almost effortless expansion SVM of they permissions to not being on the classification problem that to divide linearly about mapping and the discussion of kernel function.After introducing kernel function, the sorter that is produced still produces lineoid in H.Yet, in the Λ of natural data space, under the reverse situation of mapping function Φ, judge that now the surface is " the previous image " of lineoid.This judge surface can be in Λ hardly imaginably complexity, non-linear bunch (manifold).

The selection of kernel K does not automatically hint the mapping Φ and the higher dimensional space H of unique correspondence: zero or a plurality of mappings can be arranged, and are kernel function for its given K.The result who has the functional selection provide some conditions, under these conditions, by (S13) specific kernel K corresponding to unique mapping Φ and Space H; These conditions are called as the Mercer condition, by Scholkopf, and B, Burges, C., and Smola, A., Advances In Kernel Methods, MIT Press, Cambridge, given in 1999.These conditions guarantee to have element Kij ≡ K, and (xi, kernel matrix K xj) must be the matrix of positive definite.This character of interior nuclear matrix forces the training problem (being non-linear now) of SVM to remain the quadratic programming task.Some typical kernels comprise the polynomial function of the dot product of vector, radially basic function (exponential form) and S type (sigmoidal) function.The specific selection of kernel function allows SVM to simulate the behavior of general pattern sorter (such as feedforward nervous system network and radial primary function network).Yet, possible SVM behavior all much abundant than action span any in other method.

Under the situation of introducing kernel mappings, the Wolfe dual objective function becomes:

$L_D=\underset{i}{\mathrm{\Σ}}{a}_i-\frac{1}{2}\underset{i,j}{\mathrm{\Σ}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_{j}K(x_i\·x_j)$ S15

The Hessian matrix of Mercer condition assurance objective function remains positive definite, and therefore, optimization problem is the quadratic equation with unique global optimization (global optimum).Provide SVM lineoid among the H now by following formula:

$w=\underset{i}{\mathrm{\Σ}}{\mathrm{\α}}_i{y}_i\mathrm{\Φ}\left(x_i\right)$ S16

And correspondingly adjust the SVM decision function:

$f\left(x\right)=w\·\mathrm{\Φ}\left(x\right)+b=\underset{i=1}{\overset{l_s}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_{i}K(x_i\·x)+b$ S17

Yet, have faint difference at (S8) with (S17).Because kernel function K non-linear, in (S17) and no longer can be as (S8) in and dot product exchange: kernel function is in the way.Therefore, all l of essential execution _sKernel is estimated to estimate (S17).Fig. 6 illustrates R ²In non-linear, inseparable situation.

In description, preferred embodiment of the present invention is described as a software program.Those skilled in the art will appreciate that the equivalent that also can in hardware, constitute this software.Because image manipulation algorithms and system are known, thus this description especially at a part that forms according to the inventive method, or more directly with the algorithm and the system of the inventive method cooperation.Can from these systems well known in the prior art, algorithm, assembly and element, select others that comprised, that do not illustrate especially or describe, that be used to produce and be used in addition handle these algorithms and system and the hardware and/or the software of picture signal here.

Computer program can comprise one or more storage mediums, for example, and the magnetic storage medium such as disk (such as floppy disk) or tape; Optical storage media such as CD, light belt or machine readable go out bar code; Solid-state electronic memory device such as random-access memory (ram) or ROM (read-only memory) (ROM); Or be used to store and have any other physical device or the medium of one or more computing machine of control with the computer program of the instruction of carrying out the method according to this invention.

Claims (20)

1. the method checked of the area of computer aided of a combine digital medical image comprises:

Determine patient's inspect-type of digital medical image;

According to described patient's inspect-type, call a plurality of one or more based in the dissection block of knowledge, each piece is that the single organ or the tract carries out image that appear in the image are cut apart; And

According to described patient's inspect-type, for each organ of successfully cutting apart or tract, call a plurality of one or more based in the computer aided detection piece of knowledge, wherein each piece is designed and searches for and locate concentrating on potential disease in certain organs or the tract.

2. the method for claim 1, wherein to be suitable for watching certain organs or tract each the report of report blocks as a result from before the result of piece, so that described result is carried out decipher.

3. the method for claim 1 also comprises the following steps:

According to described patient's inspect-type, call one or more in a plurality of area of computer aided medical diagnosis on disease pieces, each piece in these pieces is designed to estimate to concentrate on the potential disease in certain organs or the tract, to assess the possibility that described organ comprises the actual disease process.

4. the method for claim 1 is wherein carried out determining of described patient's inspect-type by automatic classification method.

5. the method for claim 1, wherein said patient's inspect-type determine the result that the user imports.

6. the method for claim 1, wherein said patient's inspect-type determine to be included in result of information in the digital picture head.

7. the method for claim 1 was wherein carried out pre-service to described image before by the inspect-type classification.

8. the method for claim 1 is wherein cut apart and is comprised the described organ of establishment or the surface of tract or the digital geometry model of volume.

9. the method for claim 1, wherein said detection piece to think the image-region that may comprise lysis the locus and/or detect score and discern.

10. method as claimed in claim 9, wherein said detection piece is discerned the volume center that has regular length vector group to center on.

11. method as claimed in claim 10, wherein said diagnostics block provide the size of detected unusual estimation.

12. method as claimed in claim 11, wherein each diagnostics block is designed to estimate and show the type and the seriousness of the lysis that detects existence.

13. method as claimed in claim 12 wherein provides the confidence level score to each diagnosis situation.

14. method as claimed in claim 13, wherein said report blocks is as a result reported the result by two peacekeeping 3-D display simultaneously.

15. method as claimed in claim 14, wherein said result shows with the original figure medical image.

16. the system that the area of computer aided of a combine digital medical image is checked.Comprise:

The device that is used for patient's inspect-type of definite digital medical image;

Check control module;

A plurality of dissection blocks based on knowledge, each piece is cut apart known single organ or the tract carries out image that is present in the digital picture; And

A plurality of computer aided detection pieces based on knowledge, each piece have the specific dissection knowledge for target organ/tract customization.

17. system as claimed in claim 16 also comprises:

A plurality of reports and displaying block provide demonstration to allow the decipher to the result to the result from former piece, and each piece all is customized to the visualization technique of representing best about the information of each specific anatomical target.

18. system as claimed in claim 17 also comprises:

A plurality of area of computer aided medical diagnosis on disease pieces, each piece all have the specific anatomical knowledge of organ/tract customization according to target.

19. system as claimed in claim 18 also comprises the apparatus for automatically sorting that is used for determining patient's inspect-type.

20. system as claimed in claim 19 also is included in by inspect-type and image is carried out pretreated device before classifying.

Images