CN1983332A - Improved segmentation of nodules for computer assisted diagnosis - Google Patents

Abstract

By testing for nodule segmentation errors based on the scan data, juxtapleural cases are identified. Once identified, the scan data or subsequent estimation may be altered to account for adjacent rib, tissue, vessel or other structure effecting segmentation. One alteration is to shape a filter as a function of the scan data. For example, an originally estimated ellipsoid for the nodule segmentation defines the filter. The filter is used to identify the undesired information, and masking removes the undesired information for subsequent estimation of the nodule segmentation. Another possible alteration biases the subsequent estimation away from the incorrect information, such as the rib, tissue or vessel information influencing the original estimation. For example, a negative prior or probability is assigned to data corresponding to the originally estimated segmentation for the subsequent estimation.

Description

The improved nodule segmentation that is used for computer-aided diagnosis

Related application

Patent document require in the sequence number that on April 18th, 2005 submitted to be 60/672,277 interim U.S. Patent application, according to the rights and interests of applying date of 35U.S.C. § 119 (e), this application is incorporated herein by reference.

Technical field

Present embodiment relates to be cut apart.Especially from the scan-data such as computed tomography data, discern tubercle or other structure.

Background technology

The lung nodule segmentation is a target of the computer-aided diagnosis (CAD) that is used to discern lung neoplasm.For example, CAD system is discerned the lung tubercle from chest computed tomography (CT) data.The automanual sane solution of cutting apart can realize a reliable cubing part, tubercle as lung cancer screening and management.

In CAD system, cutting apart solution, tubercle cut apart based on brightness such as local density's maximal value algorithm.Although this solution can be carried out satisfactorily for the isolatism tubercle, these solutions can not be isolated tubercle owing to similar brightness from contiguous surrounding structure, for example wall and vascular.Advised that more perfect method is next in conjunction with how much specific restrictions of tubercle.Yet, near pleura (juxtapleural), or being attached on the wall, tubercle still is counted as challenge, because this tubercle may non-compliant geometry supposition.Another root of problem is the rib that occurs with high luminance values in the CT data.This high luminance area that approaches possible tubercle may make the deviation that estimates at of nodule center.

Two kinds of methods provide near steadily and surely cutting apart of the situation of pleura.In first method, whole lung of execution or rib are cut apart before nodule segmentation.This holistic approach may be effectively, but very complicated and depend on the precision that whole lung is cut apart on calculating.In the second approach, before nodule segmentation, carry out local non-object removal or avoidance.This partial approach may be more effective than holistic approach, but because the available information of the limited quantity of non-object construction and more be difficult to realize high-performance.

Summary of the invention

As introduction, following preferred embodiment comprises method, system or the computer-readable medium that is used for carrying out in computer-aided diagnosis improved nodule segmentation.By checking the nodule segmentation error according to scan-data, near the situation the pleura is identified.In case be identified, scan-data or estimation subsequently can be modified rib, tissue, vascular or other structure of the vicinity of cutting apart with the explanation influence.A kind of modification is to form wave filter according to scan-data.For example, the ellipsoid of cutting apart that comes from original estimation defines described wave filter.Described wave filter is used for discerning undesired information, and for the estimation of subsequently nodule segmentation, shelters the undesired information of removing.Another kind of possible modification makes estimation subsequently deviate from incorrect information, for example rib, tissue or the vessel information that influences original estimation.For example, the priori of Fou Dinging (negative prior) is assigned to corresponding to the original data of estimating of cutting apart to be used for estimation subsequently.The combination that any or they in the modification were revised, were departed from check, filtering can be used.

In first aspect, provide a kind of method that is used for carrying out improved nodule segmentation in computer-aided diagnosis.Processor is according to first cutting apart that the scan-data test handler is determined.In first cutting apart under the situation that makes the check failure that processor is determined, processor is revised scan-data, parameter or scan-data and parameter.According to second cutting apart that amended scan-data, parameter or amended scan-data and parameter determine that processor determines.

In second aspect, provide a kind of system that is used for carrying out improved nodule segmentation in computer-aided diagnosis.Processor can be operated and be used for cutting apart according to scan-data check first, can operate to be used for cutting apart under the situation that makes check failure and revise scan-data, parameter or scan-data and parameter, and can operate and be used for determining that according to amended scan-data, parameter or scan-data and parameter second cuts apart first.Display can be operated and be used to export second indication of cutting apart.

In the third aspect, computer-readable recording medium has the data that are stored in wherein, and this data representation can be by the instruction of the processor execution that is programmed, and this instruction is used for carrying out improved nodule segmentation in computer-aided diagnosis.Storage medium has instruction, be used for: check first tubercle to estimate according to scan-data, if first tubercle estimates to make the check failure, then revise scan-data, tubercle estimation or scan-data and tubercle and estimate both, and estimate that according to amended scan-data, the estimation of amended tubercle or amended scan-data and amended tubercle both determine the estimation of second tubercle.

In fourth aspect, computer-readable recording medium has the data that are stored in wherein, and this data representation can be by the instruction of the processor execution that is programmed, and this instruction is used for carrying out improved nodule segmentation in computer-aided diagnosis.Storage medium has instruction, is used for: determine filter shape according to scan-data, according to filter shape scan-data is carried out filtering, and cut apart first tubercle according to filtered scan-data.

In aspect the 5th, computer-readable recording medium has the data that are stored in wherein, and this data representation can be by the instruction of the processor execution that is programmed, and this instruction is used for carrying out improved nodule segmentation in computer-aided diagnosis.Storage medium has instruction, is used for: determine that first tubercle estimates, with first tubercle estimate to be identified as with possible pleura near tubercle relevant, and make that second tubercle estimates determine that deviating from first tubercle estimates.

The present invention is limited by following claim, and any content in this part should not be considered to be the restriction to those claims.Other aspect of the present invention and advantage are discussed below in conjunction with the preferred embodiments.

Description of drawings

Parts and figure there is no need proportionally to draw, but emphasis should be placed on and illustrates on the principle of the present invention.In addition, in the drawings, similar Reference numeral is represented corresponding part in different views.

Fig. 1 is the process flow diagram chart of an embodiment that is used for carrying out in computer-aided diagnosis the method for improved nodule segmentation;

Fig. 2 is the graphic example that improves an embodiment of cutting apart according to scan-data;

Fig. 3 is the graphic example of an improved embodiment of cutting apart; With

Fig. 4 is the block diagram of an embodiment that is used for carrying out in computer-aided diagnosis the system of improved nodule segmentation.

Embodiment

Embodiment can be in computed tomography (CT) by improving near the lung nodule segmentation that provides sane of cutting apart of the situation pleura.By the check segmentation result, can discern near the situation of pleura.Because can avoid incorrect the cutting apart that surrounding structure caused by revising.At near the situation pleura, provide modification by non-object removal in part and/or avoidance method.In one approach, detect and remove the interior lung wall district of input sub-volumes.For example, use the open operation of three-dimensional scale-of-two form (binary morphologicalopening operation).By using scan-data to discern structural element such as the ellipsoid three-dimensional structure element of data-driven, morphological operations more may cause the removal of undesired information.In another approach, the mean shift framework of expansion comprises repulsion (negating) priori, and its trend is away from the convergence of one or more particular data point.No matter this prior-constrained mean shift is used to correctly detect nodule center and the existence of rib.Segmentation precision can be enhanced under the data conditions of not removing expression wall and rib.

In one embodiment, proposed two kinds of solutions are implemented as the expansion of sane anisotropic Gauss curve fitting solution, but other partitioning algorithm also can be used.Ellipsoid structural element and/or repel sub-priori and from Gauss curve fitting, obtain.

This embodiment can be used to other imaging pattern, for example magnetic resonance imaging, ultrasound wave, x ray, positron emission or other imaging pattern now known or that newly advance to research and develop.Can use the data of bidimensional or 3-D scanning.Embodiment can alternatively be used near the situation the pleura that isolate or non-.

Fig. 1 illustrates an embodiment who is used for carrying out in computer-aided diagnosis the method for improved nodule segmentation.System or the different system such as PC, network affair device or imaging workstation by Fig. 4 realize described method.Described method is carried out action with shown order or different orders.In addition, action different or still less may be set.For example, have or the situation of action without any other under, action 26 rather than 28 is performed, or vice versa.As another example, action 22,24 and 30 is performed under 26 and/or 28 the situation not moving.

In action 22, determine the tubercle estimation.Tubercle estimates it is cutting apart of possible tubercle.Automatically or semi-automatically carry out and to determine by processor, thereby cause definite the cutting apart of processor.

This is cut apart is local, for example estimates tubercle in the subregion of the scan-data of representing plane or volume.For example, use the sub-volumes of 33 * 33 * 33 three-dimensional pixels.This zone is greater than possible tubercle, for example on diameter greater than 30 millimeters, 50 millimeters or other value.Can use have cube, the zone of other size of spherical or other shape, for example spheric region.Scan-data is the data of CT data or another type.

Import by processor or in response to the user and automatically to select regional area.According to the radiologist by watch evaluation (eye-appraisal) read or automatically the result of nodule detection system set up mark.For example, use a kind of algorithm globally so that determine may the zone relevant with tubercle.As another example, point, area or volume that user's input may be relevant with tubercle.Regional area is centered on user's input position or determined area or the volume.In the embodiment that substitutes, carry out global segmentation.

After the identification regional area, regional area is carried out nodule segmentation.Regional area may comprise isolated tubercle or near the tubercle pleura.Scan-data to regional area is carried out lung or other nodule segmentation.Can use any partitioning algorithm, for example filtering, housebroken neural network now known or that newly advance research and development, Bayes is cut apart or other cutting apart based on sorter.

In one embodiment, utilizing the Gauss curve fitting function to carry out cuts apart.Gaussian function, for example the two dimension or three-dimensional Gaussian function match scan-data.For example, publication number be 2005/0036710,2005/0096525,2005/0135663,2005/0201606 or _ _ _ _ _ _ _ _ _ _ _ _ (application serial no is 11/184,590) the one or more dividing processing described in the U.S. Patent application are performed, and the disclosure of these patented claims is incorporated herein by reference.A kind of sane anisotropic Gauss curve fitting identification is by the represented tubercle of computed tomography data.Automanual (for example, clicking) three-dimensional nodule segmentation is provided.Click the mark x of cutting apart the approximate location of using the indicating target tubercle _pThe three-dimensional Gaussian function match scan-data relevant with this mark.Determine that rule of thumb the three-dimensional boundaries of tubercle is similar to the ellipsoid of the Gauss's of institute's match 35%, 50% or other degree of confidence.Can use other structure except ellipsoid, for example spheroid, cube or irregular structure.

For the efficient of calculating, this algorithm is applied to being centered in mark x _pAnd from CT volume data I (x): R ³→ R ₊, the sub-volumes V (x) of 12 CT scan extracting data for example.This algorithm provides Gaussian function, its best or the local luminance of match target nodule fully distribute.This match is represented as: I (x) ≈ α x Ф (x; U, ∑) | _{X ε S}, Ф (x wherein; U, ∑)=| 2 π ∑s | ^-1/2Exp (1/2 (x-u) ^t∑ ^-1(x-u)) be anisotropic three-dimensional Gaussian function, α is positive quantity coefficient, S is the local neighborhood that forms the attraction basin (basin) of target, u is the Gaussian mean of institute's match of the estimated nodule center of indication, and ∑ is Gauss's covariance matrix of institute's match of the anisotropy expansion of indication tubercle.U is to use the convergence of the local maximum (for example gradient) of fitting function.

Can not calculate α and S.Alternatively, this algorithm is carried out multiscale analysis by Gauss's metric space of considering the input sub-volumes.Gauss's metric space L (x; H) be to have initialization L (x; 0)=diffusion equation of I (x) ${\∂}_{H}L=\frac{1}{2}{\▿}^{2}L$ Separate.This metric space is by I (x) and the gaussian kernel K with bandwidth matrices H _H(x) convolution defines: L (x; H)=I (x) * K _H(x; H=h ₁).Described algorithm is considered the discrete analysis yardstick { h at one group of intensive sampling _k| k=1 ... Gauss's metric space of the last structure of K}.Analyze on the yardstick at each, in each metric space image, carry out fixed size and steadily and surely analyze to be used for the anisotropic Gaussian function of match.Determine described average, be u and covariance, be ∑ at each level and smooth yardstick or rank (that is bandwidth).

Fixed size analysis and utilization metric space mean shift is carried out sane Gauss curve fitting, and the mean shift of the convergent weighting that defines in Gauss's metric space is represented as:

$m(x;H_k)=\frac{\∫{x^{\text{'}}K}_H(x-x^{\text{'}};H_k)I\left(x^{\text{'}}\right)d{x}^{\text{'}}}{\∫K_H(x-x^{\text{'}};H_k)I\left(x^{\text{'}}\right)d{x}^{\text{'}}}-x=h_k\frac{\▿L(x;h_k)}{L(x;h_k)}---\left(1\right)$

Gaussian mean u as nodule center passes through at mark x _pThe convergence of Cai Yang most of initial seeds is on every side estimated.Around estimated average u, one group of new seed is sampled.Carry out the mean shift process from each seed then.Gauss's covariance by have unknown ∑, estimate with affined least square solution only along the linear system of the mean shift vectorial structure of convergence track.Described linear system can be constructed with the metric space Hai Sai (Hessian) of response criteriaization.

Given one group of estimated Gauss, the most stable final result of estimation decision on yardstick.In one embodiment, to provide by Gauss estimated cut apart and scan-data between error calculate.In other embodiments, such as being under the big or unsettled situation in error, the estimated Gauss of this group is checked so that according to the most stable estimation of bandwidth identification that changes.Given one group to analyze yardstick { (u _k, ∑ _k) estimated Gauss, search the most stable estimation based on the stability test of dispersing and realize multiscale analysis by especially utilizing, but can use other check.A kind of exemplary algorithm adopts with each Jensen Shannon that analyzes three contiguous Gausses of yardstick calculating disperses (JSD).Suppose the distribution of canonical form, JSD represents with following simple form:

$\mathrm{JSD}\left(k\right)=\frac{1}{2}\log \frac{\left|\frac{1}{3}{\mathrm{\Σ}}_{i=k-1}^{k+1}{\mathrm{\Σ}}_i\right|}{3\sqrt{{\mathrm{\Π}}_{i=k-1}^{k=1}\left|{\mathrm{\Σ}}_i\right|}}+\frac{1}{2}\underset{i=k-1}{\overset{k+1}{\mathrm{\Σ}}}{(u_i-u)}^t{\left(\underset{i=k-1}{\overset{k+1}{\mathrm{\Σ}}}\right)}^{-1}(u_i-u)---\left(2\right)$

Wherein $u=\frac{1}{2}{\mathrm{\Σ}}_{k-1}^{k+1}\mathrm{ui}.$ At yardstick h _kMinimizing that last JSD distributes causes (most-stable-over-scales) the most stable on yardstick to estimate (u ^*, ∑ ^*).

Robustness is two aspects of this algorithm due to.The first, fixed size Gauss curve fitting solution is used the convergence of metric space mean shift to analyze and carried out has the sane model fitting that exceptional value is removed.At each h, determine u according to equation 1, wherein exceptional value is not as the estimative information of the part of tubercle.M with exceptional value removal more may reduce the influence of other structure to the convergence of peak value.This help to relax proximity structure arranged side by side, such as the problem of rib, the wall of the chest or vascular.The second, the use of selecting based on the yardstick of stability makes even not follow the process of fitting treatment of Luminance Distribution of Gauss's supposition preferably sane.This is convenient to choose at random and is used to cut apart the important clinically still frosted glass type of technical complexity or the solution of other type tubercle.

Although sane, near some pleura or other situation may cause coarse cutting apart.Sane Gauss curve fitting solution or other the estimation of cutting apart can be expanded further, so that not only handle isolated situation and handle near the situation of pleura.Sane Gauss curve fitting or other are cut apart estimation and are performed, and the result is verified in action 22, such as using testing goodness of fit.Processor automatically or semi-automatically checks tubercle to estimate according to scan-data.Described check with tubercle estimate to be identified as with possible pleura near tubercle relevant, discern the possible tubercle that is attached on the wall such as this check.

Computed tomography or other scan-data are compared with lung or other nodule segmentation.For example, the error between cutting apart of determining of processor and the scan-data is analyzed.In one embodiment, the card side's error between data and the model is calculated, but other Error Calculation also can be used.In another embodiment, it is estimated that linear DC departs from (bias).Can use other didactic check, such as using housebroken neural network or other sorter.Combination that can service test departs from check such as error analysis of card side and linear DC and combines.Can be provided with to cut apart not and determine, such as based on training or have the error threshold or the didactic result of the data set of known true value by check or any threshold value by check or other.

When initially fitting result is failed by check in action 22, carry out further to handle and to cut apart more accurately or the tubercle estimation to provide.Can use one or more different further processing, comprise any processing now known or that research and develop recently.Generally cutting apart for sane Gauss, may be because near the situation pleura by departing from the thickest failure of cutting apart that card side's error analysis that testing goodness of fit combines detected with linear DC.Initial fitted Gaussian at this failure may tend to be similar to wall and framing system.Adopt initial fitted Gaussian as handling cutting apart solution and can utilizing these observations of input.

In one embodiment, the further processing execution of processor utilization is revised in action 24.The parameter that scan-data, tubercle estimate to handle, be used to estimate to handle, their combination or other modification are performed.In one embodiment, revise scan-data according to morphic function based on initial segmentation.In another embodiment, the estimation of cutting apart subsequently is weighted away from initial segmentation.Because the check failure, initial segmentation is considered to out of true.This failure can be used to influence later cutting apart with identification tubercle rather than other structure.Other is or is not that the embodiment of the function of initial segmentation can be used.

Be used in the embodiment of action 24 modification scan-datas, selected with the scan-data of undesired structure-irrelevant in action 26, perhaps relevant with undesired structure scan-data is removed.For example, remove by the open wall of carrying out of three-dimensional configuration.Selecting or removing is the function of morphic function.Morphic function is in response to scan-data, all functions that may represent the initial segmentation of undesired structure in this way.Any morphic function can be used, such as according to scan-data or according to determining filter shape in response to the ellipsoid of initial match.

The input sub-volumes of failure scenarios may comprise lung wall district near the pleura.This wall district is usually expressed as the big connected region that has than the CT value that the lung soft tissue is high on every side.Near the pleura tubercle may show as the nodular structure that is partially embedded in the wall.Use morphological operations from sub-volumes, to remove the wall district.Then the scan-data of removing wall is carried out sane Gauss curve fitting or other is cut apart, thereby cause the improved of target nodule to be cut apart.

Fig. 2 is with a kind of possible algorithm that is used for selecting or removing according to morphic function scan-data of diagrammatic representation.Input comprises sub-volumes V (x), mark x _pWith the fitted Gaussian (u that makes the testing goodness of fit failure ^*, ∑ ^*).Images with 34 expressions illustrate sub-volumes with two-dimensional cross sectional, wherein represent to be embedded in tubercle on the wall than bright area.With undesired structure, from V (x), removed, thereby cause V represented on image with 37 marks such as the relevant scan-data in wall district _r(x).

As at 35 the expression images shown in, scan-data is by binarization or be converted into binary representation.Threshold value, be applied to scan-data such as the value 500 of the scan-data that is used to have 0 to 4,095 dynamic range.This threshold value is a luminance threshold.Can use other threshold value.

Morphic function calculates based on initial tubercle estimation, spheroid, ellipsoid, their combination or other shape at least in part.For example, if mean diameter greater than threshold value, then estimates to come initialization three-dimensional structure element according to initial tubercle, and if mean diameter less than threshold value, then come initialization three-dimensional structure element according to predetermined structure.To by initial segmentation covariance ∑ ^*Defined ellipsoidal mean diameter d _AveCalculate.If mean diameter is greater than threshold value, then morphic function is by the initial segmentation ∑ ^*The three-dimensional structure E of definition, the E=∑ ^*Any threshold value can be used, such as 16.6 three-dimensional pixels or specific size.Otherwise three-dimensional structure E is set to have radii fixus, such as the three-dimensional sphere of 14 three-dimensional pixels or specific size.Can use other threshold value, radius, shape or size.Can use two-dimensional process.At each sub-volumes or the undesired structure of possibility, estimate the ellipsoid structural element that these data are relevant.Structural element has same or different size.

Structure E represents filter shape.Utilize the filtering of the filter shape that data derive to allow list or hyperchannel is level and smooth and/or sharpening to select or to remove scan-data.According to filter shape scan-data is carried out filtering.It is open that this filtering is carried out form according to morphic function.For example, it is open to carry out three-dimensional scale-of-two form according to three-dimensional structure element E, thereby causes only keeping the level and smooth volume B in big wall district _s(x)=[B _o(x)  E]  E.Image with 36 expressions shows the result who the scale-of-two scan-data is carried out filtering according to structural element.

According to the scan-data with the initial sub-volumes of 34 expressions being sheltered with the open output of the three-dimensional scale-of-two forms of 36 expressions.This is sheltered and selects interested data or remove undesired data.For example, by using B _s(x) the non-V of sheltering (x) carries out wall and removes: V _r(x)=V (x) xNOT[B _s(x)].In action 30, amended scan-data carried out cut apart, such as to having x _pV _r(x) carry out sane Gauss curve fitting algorithm.This is cut apart can provide improved nodule segmentation (u _Wr, ∑ _Wr).

Among the embodiment that substitute or additional in the action 28 of Fig. 1, make to cut apart and estimate to deviate from cutting apart or the tubercle estimation of failure.For near the situation pleura, depart from can as U.S. Patent No. (sequence number _ _ _ _ _ _, submitted (attorney reference 2005P05271US) on March 9th, 2006) in disclosedly be performed like that, the disclosure of this patent is incorporated herein by reference.The convergence of cutting apart subsequently is influenced or pushed away wrong result.For example, the ellipsoid by initial segmentation output influences later cutting apart.For near the tubercle the pleura, wall that make that tubercle estimates determine to depart from represented by computed tomography data or framing system and adjacent with near the possible pleura tubercle.Initial tubercle is estimated to be assumed that relevant with wall or framing system.Do not removing clearly under the data conditions of expression wall and/or rib, subsequently nodule center is being detected.

Based in the cutting apart of Gauss curve fitting, Gauss repels priori (repelling prior) constraint mean shift.Prior-constrained mean shift is attached to spatial prior information in the mean shift analysis of data-driven.The priori of negating is assigned at least one with initial tubercle estimation or cut apart the scan-data of relevant position.Sub-volumes V (x) is carried out previous sane Gauss curve fitting, thereby cause nodule center and spread estimation (u ^*, ∑ ^*).It is the normal probability paper distribution Q (x) of the likelihood at estimated center that this fitted Gaussian is interpreted as indicating x, and it is represented as:

Q(x)＝N(x；u ^*，∑ ^*)＝|2π∑ ^*exp(-(x-u ^*) ^t∑ ^*-1(x-u ^*)) (3)

Work as x=u ^*The time Q (x) have mxm..The estimated position u of testing goodness of fit failure expression ^*Not in the center of target nodule and estimated expansion ∑ ^*Generally the scope of expression (rib/wall) structure has attracted to described construction error the nodule center of mean shift convergence away from reality.Nodule center can utilize affined mean shift to reappraise, and by the understanding to Q (x) convergence of this mean shift is departed from, so that described convergence is pushed away the estimation u of failure ^*

For in conjunction with this repulsion (negative) priori, equation 3 be converted and subsequently cut apart or tubercle is used as parameter modification in estimating.Scan-data I (x) is taken a sample or is associated with weight to represent that some data points are than other data point idea more suitably.The positive weight that priori derives is represented as by non-definition of Q (x):

w _Q(x)＝1-|2π∑ ^*| ^1/2Q(x) (4)

The metric space that causes the following data space of the usefulness discretize of sampling again to be represented in conjunction with the priori of negating

$\tilde{L}(x;h)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{w}_Q\left(x_i\right)I\left(x_i\right)K_b(x-x_i)---\left(5\right)$

Converge on

In the mean shift m of pattern _r(x; H Q) is the prior-constrained metric space mean shift being negated.The prior-constrained metric space mean shift being negated is defined by following formula:

$m_r(x;h,Q)=\frac{{\mathrm{\Σ}}_i{x}_i{K}_h(x-x_i)I\left(x_i\right)w_Q\left(x_i\right)}{{\mathrm{\Σ}}_i{K}_h(x-x_i)I\left(x_i\right)w_Q\left(x_i\right)}-x---\left(6\right)$

Convergence property is held, because w _Q(x _i) 〉=0  x _i

Utilize the priori of negating, carry out once more and cut apart.For example, according to estimating that with initial tubercle relevant deviation is applied to computed tomography data with Gaussian function.In action 30, construct new Gauss curve fitting scheme by in sane fitting algorithm, replacing initial gauges spatial mean value translation (replacing equation 1) with equation 6 with prior-constrained mean shift.The given initial Gauss (u that makes the testing goodness of fit failure ^*, ∑ ^*), raw data V (x) is carried out this have m _r(x; H, new solution Q), thus cause tool (u _Ms, ∑ _Ms) improved cutting apart.

Action 26 and 28 can be used independently or together.Come from both results of action 26 and 28 can be merged, such as by the determined result of match on average or best.Action 26 processing can be used to guide or improve the processing of action 28 or vice versa.Alternatively, under the situation of not carrying out action 28, action 26 is performed, or action 28 is performed under the situation of not carrying out action 26.

In action 30, determine definite the cutting apart or the tubercle estimation of another processor according to amended scan-data, amended parameter, the estimation of amended tubercle or their combination.Identical or different cutting apart is performed, such as determining according to three-dimensional Gauss curve fitting function.Determine cutting apart subsequently according to the priori of negating or scan-data with information removal or that select.At the data of removing or select, cut apart tubercle according to filtered scan-data, this filtered scan-data is masked more may the data relevant with tubercle with identification.At data devious, cut apart the scan-data of the mean shift that influences and cut apart tubercle according to having the priori of being subjected to.Modification can allow to have near the nodule segmentation of lung preferably of the tubercle pleura of computed tomography data.

Fig. 3 illustrates some illustrative examples.Leftmost image comprise center 41 with the undesired structural ellipse of big tubercle adjacency cut apart 42.Center image illustrate remove with the corresponding scan-data of described structure after subsequently cut apart 43.Rightmost image illustrate after in conjunction with the priori of negating subsequently cut apart 44.This processing is performed once, but can be such as being carried out repeatedly under the situation of a plurality of upsets (distractor) at tubercle.Can in different situations, carry out preferably the different further processing of revising.At each mark, possible tubercle or sub-volumes (that is, regional area), this processing is used independently.

Open and be subjected to solution prior-constrained mean shift, that extend out from sane Gauss curve fitting method can cut apart near the situation pleura effectively based on form.When tubercle is attached to non-wall construction or influenced by non-wall construction or very large tubercle when being attached to the thin part of lung wall, the use of negative priori can be worked better than form is open.The form opening can be in other situation, for example carry out near the better off ground the little pleura.The method of Fig. 1 can be tested and selects suitable further processing concrete situation.Alternatively, no matter make what the reason of this check failure is, carry out identical further processing.

Cutting apart can assisted diagnosis.For example, correctly cut apart the more accurate cubing that tubercle can be provided.Tubercle volume, shape or variation can help diagnosis clinically.For the Gauss curve fitting method of above-mentioned discussion, use the ellipsoid border approximate.Yet, utilize the nonparametric of Gauss's priori to cut apart by combination, perhaps, the further improvement of cutting apart quality is possible, wherein this Gauss's priori is by using the method for being advised in publication number is 2005/0201606 U.S. Patent application to derive, and the disclosure of described patented claim is incorporated herein by reference.Tubercle is similar to by Gaussian function.In order to obtain more cutting apart of fine-grained, utilize the Gauss curve fitting that is used as priori to carry out different cutting apart.Can use other modification or difference.

In one embodiment, the method for Fig. 1 comprises other check or action.For example, multilayer or stage division are employed.Can use the Gauss of different size.For example, two different sub-volumes and cutting apart of being associated are performed (for example, 33 * 33 * 33 and 66 * 66 * 66).Big sub-volumes can be such as by level and smooth and down-sampling is reduced, with the application parameter identical with less sub-volumes.Less sub-volumes at first is used.If cut apart the smaller volume failure, then bigger sub-volumes is used to bigger tubercle is tested.For example, if ∑ ^*Greater than threshold value, less cutting apart can be defined as failure, and use bigger cutting apart.

Fig. 4 illustrates an embodiment who is used for carrying out in computer-aided diagnosis the system 50 of improved nodule segmentation.System 50 is workstation, personal computer, network, server, computer-aided diagnosis system, imaging system, computed tomograph scanner system, medical diagnostic imaging system or other disposal system known now or that research and develop recently.For example, the Local or Remote workstation receives the image that is used for computer-aided diagnosis.The method or the diverse ways of system's 50 execution graphs 1.

System 50 comprises processor 52, storer 54, display 56 and user input apparatus (user input) 58.Can be provided with other, different or less components.For example, system 50 does not comprise user input apparatus 58 and/or display 56.As another example, system 50 comprises sensor, such as the computed tomography images former.Parts are shown adjacent one another are, such as in identical room, and on identical cart, or in identical house.In other embodiments, one or more parts are long-range, are that remote data base or display 56 are on networking or wireless device such as storer 54.

User input apparatus 58 is keyboard, button, slider, mouse, touch pads, touch-screen, trace ball, dial (of a telephone) or other input media now known or that research and develop recently.User input apparatus 58 is the parts by processor user interface 52 generations or control.User and computer-aided diagnosis system 50 are alternately to discern tubercle or number of computations based on cutting apart.For example, user input apparatus 58 receives the user and imports the tubercle mark position.

Processor 52 is one or more general processors, digital signal processor, application-specific IC, field programmable gate array, server, network, digital circuit, mimic channel, their combination or other device that is used to handle medical image now known or that research and develop recently.Processor 52 software program for execution, such as manual generation or the programming code or such as housebroken classification or model system.Software identification tubercle border.Alternatively, hardware or firmware are carried out identification.Processor 52 obtains scan-data, operational order and/or out of Memory from storer 54.

Processor 52 can be operated and be used for according to scan-data check nodule segmentation.Scan-data is a computed tomography data, but can use the data of other type.Scan-data represents tubercle, such as knurl or other structure.Example is near the tubercle the pleura in the lung computed tomography.In order to check, nodule segmentation is compared with computed tomography data.Can use any check, such as error of fitting or didactic check.

Processor 52 can be operated and be used for cutting apart under the situation that makes described check failure and revise scan-data/parameter or scan-data and parameter formerly.For example, processor 52 is selected scan-data according to morphic function.As another example, processor 52 make subsequently cut apart determine deviate from cutting apart that previous processor determines, such as deviating from wall or rib is cut apart.Modification can be the function that is used to the scan-data of tubercle estimation, for example previous estimated tubercle is used for revising.

Processor 52 can be operated and be used for determining previous and/or cutting apart subsequently.Three-dimensional Gauss curve fitting is performed, but can use other partitioning algorithm.In an example, processor 52 is determined cutting apart subsequently according to selected scan-data under the removed situation of other scan-data.In another example, processor 52 is determined cutting apart subsequently according to the parameter of being revised or dividing processing, for example with use the priori of negating relevant.

Storer 54 is computer-readable recording mediums.Computer-readable recording medium comprises various types of volatibility and non-volatile memory medium, including, but not limited to random access memory, ROM (read-only memory), programmable read only memory, EPROM, electricallyerasable ROM (EEROM), flash memory, tape or disk, optical medium or the like.Storer 54 storages are used for by scan-data processor 52 processing or during being handled by processor 52.Scan-data is imported in processor 52 or the storer 54.In one embodiment, scan-data is a view data.In other embodiments, scan-data be before being converted to picture format data, such as sensing data or detected data.

In one embodiment, storer 54 is computer-readable recording mediums, and this computer-readable recording medium has the instruction that can be carried out by the processor that is programmed 52 that is stored in wherein.Processor 52 is realized automatic or automanual operation discussed herein with instruction at least in part.Instruction makes processor 52 realize described herein any, all or some function or action.Function, action or task are independent of instruction set, storage medium, processor or the processing policy of specific type and can be carried out in the mode that works alone or in combination by software, hardware, integrated circuit, membrane equipment (film-ware), microcode or the like.Similarly, processing policy can comprise multiple processing, multitask, parallel processing or the like.

In one embodiment, instruction is stored in and is used for by medical diagnostic imaging system, computer-aided diagnosis system or the removable media drive that reads with the workstation of imaging system networking.Teletype command on imaging system or the workstation.In another embodiment, instruction is stored on the remote location, is used for being transferred to imaging system or workstation by computer network or via telephone communication.In other embodiments, instruction is stored in hard disk, random access memory, cache memory, impact damper, removable medium or other system of installing.

Display 56 is monitor, CRT, LCD, plasma, flat screen, touch-screen, projector, printer or other display device known now or that research and develop recently.The indication that display 56 outputs are cut apart.For example, display 56 is exported the image that generates based on cutting apart according to the scan-data with superposition boundary.As another example, display 56 outputs are based on the value of cutting apart, such as volume.Other output can be provided.

Though described the present invention with reference to various embodiment in the above, should be understood that: many changes and modification can made without departing from the scope of the invention.Therefore mean that above-mentioned detailed description should be considered to illustrative rather than restrictive, and it should be understood that the following claim of all equivalents that comprises is intended to limit the spirit and scope of the present invention.

Claims (30)

1. method that is used for carrying out improved nodule segmentation in computer-aided diagnosis, described method comprises:

Utilize processor according to first cutting apart that computed tomography data comes that test handler determines;

In first cutting apart under the situation that makes the check failure that processor is determined, utilize this processor to revise scan-data, parameter or scan-data and parameter; And

According to second cutting apart that amended scan-data, parameter or amended scan-data and parameter determine that processor determines.

2. the method for claim 1 further comprises:

First cut apart and be defined as the lung nodule segmentation what processor was determined according to scan-data, described scan-data comprises computed tomography data;

Wherein check comprises according to computed tomography data and checks the lung nodule segmentation.

3. method as claimed in claim 2, that determines wherein that processor determines first cuts apart and comprises according to tubercle mark position and three-dimensional Gaussian function and carry out match.

4. the method for claim 1 is wherein determined to comprise according to three-dimensional Gaussian function and is carried out match.

5. the method for claim 1, wherein check comprises near the tubercle the pleura is tested.

6. the method for claim 1, wherein check comprises:

Analysis processor determine first cut apart and scan-data between error;

The execution linear DC departs from;

Check heuristicly; Or

Their combination.

7. the method for claim 1 is wherein revised and is comprised according to morphic function and remove scan-data, and determine wherein that processor determines second cut apart to be included under the situation that does not have the scan-data removed and determine according to scan-data.

8. method as claimed in claim 7, wherein removal comprises:

Determine morphic function based on processor first cutting apart of determining, spheroid or their combination at least in part;

It is open to carry out form according to this morphic function; With

According to the output that form is open scan-data is sheltered.

9. the method for claim 1, wherein revise comprise make that processor determines second cut apart determine to deviate from that processor determines first cut apart.

10. method as claimed in claim 9 wherein departs from and comprises the priori of negating distributed to processor is determined and first cut apart relevant scan-data, and determine wherein that processor determines second cut apart the priori that comprises according to negating and determine.

11. method as claimed in claim 9, wherein said scan-data comprise near the tubercle the pleura, and wherein depart from and comprise and deviate from wall or rib.

12. a system that is used for carrying out in computer-aided diagnosis improved nodule segmentation, described system comprises:

Processor, can operate and be used for cutting apart according to scan-data check first, can operate to be used for cutting apart under the situation that makes check failure and revise scan-data, parameter or scan-data and parameter, and can operate and be used for determining that according to amended scan-data, parameter or scan-data and parameter second cuts apart first; And

Display can be operated and is used to export second indication of cutting apart.

13. system as claimed in claim 12, wherein said processor can be operated to be used for cutting apart first according to scan-data and be defined as nodule segmentation, described scan-data comprises computed tomography data, and wherein can operate and be used for check and comprise operating and be used for checking nodule segmentation according to computed tomography data.

14. system as claimed in claim 13 further comprises:

User input apparatus can be operated and is used to receive the tubercle mark position;

Wherein said processor can be operated and be used for determining that according to three-dimensional Gauss curve fitting first and second cut apart.

15. system as claimed in claim 12, wherein processor can be operated to be used for revising and comprises operating and be used for selecting scan-data according to morphic function, and wherein processor can be operated and is used for determining that second cuts apart and comprise operating and be used for determining according to selected scan-data.

16. system as claimed in claim 12, wherein scan-data comprise near the pleura tubercle and wherein processor can operate be used to revise comprise can operate be used to make second cut apart determine to deviate from that processor determines first cut apart, this departs from away from wall or rib.

17. in computer-readable recording medium, this computer-readable recording medium has the data of the instruction that the expression that is stored in wherein can be carried out by the processor that is programmed, described instruction is used for carrying out improved nodule segmentation in computer-aided diagnosis, and described storage medium comprises instruction, is used for:

Check first tubercle to estimate according to scan-data;

If first tubercle estimates to make the check failure, then revise scan-data, tubercle estimation or scan-data and tubercle and estimate both; And

Estimate or amended scan-data and definite second tubercle estimation of amended tubercle estimation according to amended scan-data, amended tubercle.

18. instruction as claimed in claim 17 further comprises:

First tubercle is estimated to be defined as near the lung nodule segmentation of the tubercle the pleura, and the lung nodule segmentation determines that according to tubercle mark position, Gauss curve fitting function and scan-data described scan-data comprises computed tomography data;

Wherein check comprises according to computed tomography data check lung nodule segmentation; With

Wherein definite second tubercle is estimated to comprise according to the Gauss curve fitting function and is determined.

19. instruction as claimed in claim 17 is wherein revised the morphic function comprise according in response to scan-data and is selected scan-data, and determines wherein that second tubercle is estimated to comprise according to selected scan-data and determine.

20. instruction as claimed in claim 17, wherein modification comprises first tubercle estimation of determining to deviate near the tubercle of pleura that second tubercle is estimated.

21. in computer-readable recording medium, this computer-readable recording medium has the data of the instruction that the expression that is stored in wherein can be carried out by the processor that is programmed, this instruction is used for carrying out improved nodule segmentation in computer-aided diagnosis, and described storage medium comprises instruction, is used for:

Determine filter shape according to scan-data;

According to filter shape scan-data is carried out filtering; With

Cut apart first tubercle according to filtered scan-data.

22. instruction as claimed in claim 21, wherein definite filter shape comprises makes Gaussian function match scan-data, and determines filter shape according to the ellipsoid in response to match.

23. instruction as claimed in claim 21 further comprises:

Determine second nodule segmentation according to the Gaussian function of match scan-data;

Check second nodule segmentation according to scan-data; With

If second nodule segmentation makes the check failure, then execution is determined filter shape, filtering and is cut apart, described definite filter shape and filtering comprise removes some scan-datas, and described cut apart to be included under the situation that does not have the scan-data removed determine first nodule segmentation according to another Gaussian function of match scan-data.

24. instruction as claimed in claim 21, wherein first tubercle comprises near the lung nodule segmentation of the tubercle that pleura is, and scan-data comprises computed tomography data.

25. instruction as claimed in claim 21, wherein definite filter shape comprises according to first threshold makes the scan-data binarization, determine the mean diameter that initial tubercle is estimated, if mean diameter is greater than second threshold value, then estimate to come initialization three-dimensional structure element according to initial tubercle, if and mean diameter then comes initialization three-dimensional structure element according to predetermined structure less than second threshold value, filter shape comprises the three-dimensional structure element;

Wherein filtering comprise according to the three-dimensional structure element carry out three-dimensional scale-of-two form open and according to three-dimensional scale-of-two form open output scan-data is sheltered; With

Wherein cutting apart first tubercle comprises according to the scan-data of being sheltered and estimates first tubercle.

26. in computer-readable recording medium, this computer-readable recording medium has the data of the instruction that the expression that is stored in wherein can be carried out by the processor that is programmed, this instruction is used for carrying out improved nodule segmentation in computer-aided diagnosis, and described storage medium comprises instruction, is used for:

Determine the estimation of first tubercle;

With first tubercle estimate to be identified as with possible pleura near tubercle relevant;

What make the estimation of second tubercle determines that deviating from first tubercle estimates.

27. instruction as claimed in claim 26, wherein definite first tubercle is estimated to comprise makes Gaussian function The Fitting Calculation machine tomoscan data, wherein depart from comprise make that second tubercle estimates determine to deviate from the wall represented by computed tomography data or framing system and in abutting connection with near the tubercle the possible pleura, the first tubercle estimation is relevant with wall or framing system

Wherein depart from and comprise according to estimating that with first tubercle relevant departing from makes Gaussian function The Fitting Calculation machine tomoscan data.

28. instruction as claimed in claim 26 wherein departs from and comprises the scan-data of the priori of negating being distributed at least one position relevant with the estimation of first tubercle.

29. instruction as claimed in claim 26 wherein departs from and comprises that utilizing Gauss to repel priori retrains mean shift.

30. instruction as claimed in claim 26 determines that wherein the estimation of first tubercle comprises the function of the Gaussian function of determining the match scan-data;

Wherein identification comprises according to scan-data and checks first tubercle to estimate, check identification is possible is attached to and estimates relevant tubercle with first tubercle on the wall.

Images