CN1703711A - System and method for using delayed enhancement magnetic resonance imaging and artificial intelligence to identify non-viable myocardial tissue - Google Patents

Abstract

A system and method for imaging and identifying non-viable myocardial tissue in a patient's myocardium is disclosed. Images of a section of the myocardium are obtained. An endocardial border and epicardial border of the section of the myocardium is segmented. The section of the myocardium is divided into sectors. One or more selected features of the sectors of the myocardial wall are measured and applied to a decision surface. A determination is made as to whether each sector contains viable or non-viable myocardial tissue. An image that shows each sector of the myocardial wall and an indication of its viability is displayed.

Description

Utilize the system and method for delayed enhancement magnetic resonance imaging and artificial intelligence to identify non-viable myocardial tissue

The cross reference of relevant application

The application requires the U.S. Provisional Application No. of the sequence number 60/415,840 of submission on October 3rd, 2002, and this provisional application all is included in herein as a reference.

Invention field

The present invention is directed to and utilize the system and method for medical image (for example delayed enhancement magnetic resonance, film magnetic resonance) cardiac muscular tissue's segmentation, in more detail, at using support vector machine to come the system and method for identify non-viable myocardial tissue intelligently by utilizing the distinctive one or more features of described types of organization.

Background of invention

After heart attack, need discern and identify non-activity (necrosis) tissue, so that research and develop the intervention strategy and the therapeutic scheme of the heart disease of some type effectively.Healthy maybe can pass through coronary artery bypass, and the tissue of inserting recovery such as support should be distinguished with the tissue of non-activity or irreversible damage and come.Like this, just can predict which patient can be benefited from revascularization, thereby improve their cardiac function and survival rate.

The physician relies on several Noninvasive indicator (indicator) to determine the activity of cardiac muscular tissue.The form of cardiac muscle, specifically, myocardium attenuation is the evidence of slough.In addition, abnormal motion, for example passive movement in a zone, or not also expression cardiac muscle damage of motion under extreme case.But though the change of form and function can show the unusual of tissue, they are for distinguishing unusual and non-activity (necrosis) tissue sensitivity not enough.

The imaging technique that also can utilize contrast to strengthen helps discern non-viable regions.Positron emission tomography (PET) can be by different signal intensity indication activated information with single photon emission tomography (SPECT).But the purposes of these physiotherapy apparatus is limited, because their resolution is quite low, and unavailable usually under the PET situation.

The imaging technique that recently a kind of novel contrast strengthens, delayed enhancement magnetic resonance (DEMR) has shown the non-viable myocardial of can directly showing one's color.The DEMR imaging is a kind of technology that non-viable myocardial tissue occurs with the signal intensity that increases.DEMR normally (for example: Gd-DTPA) utilize after 20-30 minute standard conversion to recover that the MRI acquisition sequence carries out is taking paramagnetic contrast agent.And DEMR has enough spatial resolutions, can accurately will have the cardiac muscle of activity (normal or ischaemic) and the cardiac muscle difference of non-activity to come in the left locular wall.Radiologists obtain these images in conjunction with other functional physiotherapy apparatus (for example MR film) usually, and utilize the magnetic domain knowledge and experience to find non-viable tissue.

Even utilize above-mentioned information, determine that the activity of organizing is still challenging.At first, DEMR is prone to false negative.That is, non-viable regions may not have the signal intensity of enhancing.The second, determine that based on morphology and discriminating morphology activated state (thickening or ventricular wall motion) requires experience and intuition to a certain degree.And the indicator of indication locular wall abnormal motion may be merged.In other words, is the specific region of heart to spur at autokinetic movement or by adjacent domain? in brief, this can to make the expert be non-activity with an area marking.

Although have any problem in the identification, people still extremely pay close attention to the location and quantize the problem of non-viable tissue because shown the degree of infraction reproduce with coronary artery after the long-term improvement of cardiac function closely related.But DEMR is the state-of-the-art technology of still ratifying without FDA, can as the brainstrust of testing and develop DEMR personally experience not arranged to the unfamiliar clinician of DEMR.It is desirable to, the clinician is no matter new hand whether should obtain expert's feedback opinion.

DEMR is done a large amount of work, but seldom had scheme that its automatic segmentation problem is described.In one known technique,, obtain its mean flow rate, brightness is designated as non-activity greater than the pixel of two standard deviations by the zone of manually finding active myocardium in the DEMR image.MRI technology in the past or utilize cardiac shape separately and function is determined the activity of tissue is perhaps checked the quantity of non-viable tissue and the relation between the cardiac function as a result.

As mentioned above, DEMR is prone to false negative.And the traditional fragmentation technique for example threshold value boundary method (thresholding) of region growing or DEMR can produce incompetent result.

Summary of the invention

The present invention is directed to the employing machine learning techniques, for example support vector machine (SVM) reaches a kind of partition strategy by merging a plurality of indexs, so that identify non-viable myocardial tissue.The present invention utilizes DEMR, and morphology and discriminating morphologic information are automatically with non-viable myocardial zone segmentation.Use artificial intelligence technology " study " how the expert to carry out segmentation.Like this, the clinician in described field just can obtain second suggestion of following form: promptly, " our expert have percent so so confidence will classify non-activity as with lower area ".Be that activity or non-activity are clinician's responsibilities at last with tissue typing.But this class feedback is of great value input and can assists the clinician to make suitable medical diagnosis.Under the poorest situation, accelerate the clinical speed of passing through by solution being provided good editing (editable) conjecture.

Brief Description Of Drawings

Below will consult accompanying drawing the preferred embodiments of the present invention are described in more detail, same label is represented same element in the accompanying drawing:

Fig. 1 is the block scheme according to the system architecture of demonstration magnetic resonance imaging of the present invention (MRI) system;

Fig. 2 is the image that utilizes the patient's heart left ventricle of DEMR generation;

Fig. 3 is that graphic extension utilizes support vector machine that cardiac muscular tissue's segmentation is determined that also which tissue is the method flow diagram of non-activity (if any);

Fig. 4 is the curve map of graphic extension based on the demarcation of the decision surface of the measurement result of three kinds of myocardium characteristics; And

Fig. 5 is the graphic extension of the user interface of demonstration, and described user interface illustrates the data of DEMR screening.

Describe in detail

The present invention is directed to and use support vector machine to come the system and method for identify non-viable myocardial tissue intelligently by utilizing the peculiar myocardium characteristic of one or more non-viable tissue.Fig. 1 is the block scheme according to demonstration magnetic resonance imaging of the present invention (MRI) system.An example of MRI system is the MAGNETROM class MRI system that Siemens Company (Siemens AG) makes.

Fig. 1 graphic extension is arranged in the synoptic diagram of each parts of the MRI system of scanning room 100.Magnet 108 produces first magnetic field that is used for imaging process.Gradient coil 110 is arranged in magnet 108, is used at X, and Y produces magnetic field gradient on the Z direction.Radio frequency (RF) coil 112 is arranged in gradient coil 110.Coil 112 produces and makes spin half-twist or 180 ° of second required magnetic fields.Coil 112 also detects the signal of ex vivo spin.By computer-controlled patient table 106 patient 102 is placed in the magnet 108.Patient table 106 has the bearing accuracy of 1mm.Scanning room 100 is surrounded by RF radome 104.Radome 104 prevents that high power RF impulse radiation is to the hospital outside.It prevents that also the various RF signals in TV or radio station from being detected by the MRI system.Some scanning room is also surrounded by magnetic shielding cover, and magnetic shielding cover containment magnetic field makes its unlikely diffusion cross far and enter in the hospital.In some novel magnets, described magnetic shielding cover is the integral part of magnet.

The central module of MRI system is a computing machine 126.Computing machine 126 is being controlled all parts in the MRI system.RF parts under computing machine 126 controls are radio frequency source 138 and pulse programmer 134.Radio frequency source 138 produces the sine wave of required frequency.Pulse programmer 134 is change mark sine pulse with the RF pulse shaping.RF amplifier 136 is brought up to pulse power kilowatt from milliwatt.Computing machine 126 is also controlled gradient pulse programmable device 122, and programmable device 122 is set the shape and the amplitude in each magnetic field in three gradient magnetics.Gradient amplifier 120 is brought up to the power of gradient pulse the level that is enough to drive gradient coil 110.

The array processor (not shown) that comprises in some MRI systems is the device that can carry out two-dimensional Fourier transform in less than one second time.Computing machine 126 is unloaded to this with Fourier transform and installs faster.The operator of MRI system provides input signal by control desk 128 to computing machine 126.Select and customize at 128 pairs of imaging sequences of control desk.The operator can see image on the video display that is positioned at control desk 128, perhaps the hard duplicate of construction drawing picture on film printer.

Shown delayed enhancement magnetic resonance imaging (DEMR) non-viable myocardial of can directly showing one's color.The DEMR imaging is a kind of technology that non-viable myocardial tissue occurs with the signal intensity that increases.Usually utilize after 20-30 minute standard conversion to recover the MRI acquisition sequence to carry out DEMR taking paramagnetic contrast agent (Gd-DTPA).And DEMR has enough spatial resolutions, can accurately will have the cardiac muscle of activity (normal or ischaemic) and the cardiac muscle difference of non-activity to come in the left locular wall.Radiologists obtain these images in conjunction with other functional physiotherapy apparatus (for example MR film) usually, and utilize the magnetic domain knowledge and experience to find non-viable tissue.

The professional and technical personnel understands, can use other contrast preparation in the DEMR process.For example manganese or iron contrast preparation also can use.Research for various contrast preparation is described in following article: " Tissue-specific MR Contrast Agents ", the author is WeinmannHJ., Ebert W., Misselwitz B. and Schmitt-Willich H, be published in the European Journal of Radiology in April, 2003, Vol.46, Issue 1, the 33-44 page or leaf, described article is included in this paper as a reference.

Fig. 2 illustrates the image of the left ventricle 202 that utilizes the DEMR generation.As seen from the figure, by taking paramagnetic contrast agent, the zones of different of left ventricular wall demonstrates different brightness.As seen from the figure, darker ventricular wall tissue 206 expression tissues have activity.It is non-activity that the brighter described tissue of ventricular wall tissue 204 expressions has than high likelihood.

By utilizing artificial intelligence technology " study " how the expert to carry out segmentation, further strengthens the present invention.Specifically, the present invention uses support vector machine (SVM) to help the clinician to determine whether heart tissue is non-viable tissue.SVM is a kind of artificial intelligence technology that is monitored, training computer understanding provides a kind of phenomenon of a series of examples.The feature that training is selected by the user based on a cover.This study is considered to " monitored ", because in our situation, these examples all are added with the label that belongs to or do not belong to a certain particular category by the expert.

So, be not an instruction list as traditional computer program is done, stipulating how to discern something, but computing machine comes by a series of examples " and study ".SVM had been used in medical imaging classification and detection task in the past.Be used for distinguishing the identification feature of polyp and health tissues in their brain classification in CT colon photography and in the PET image.In addition, the Microcalcification in the breast x radiograph also once utilized SVM to detect.But as far as we know, this is that the SVM first Application is on cardiac segmentation.

The present invention adopts SVM and applies it to high-dimensional feature space, so that based on predict the classification of expert to heart tissue in the feature described in the preamble: activity or non-activity are arranged.Specifically, use the binary decision function:

Described formula will be imported (raw information) map to output (classification),

$\stackrel{\‾}{x}=(x_1,\·\·\·,x_n)---\left(2\right)$

F in the formula (x) 〉=0 represents positive class.Learning algorithm is selected a kind of decision function from candidate's space of decision functions.

Decision function is got following form:

$f\left(\stackrel{\‾}{x}\right)=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_{i}k(\mathrm{\φ}\left({\stackrel{\‾}{x}}_i\right),\mathrm{\φ}\left(\stackrel{\‾}{x}\right))+b---\left(3\right)$

φ in the formula (x) is the input function that is called feature, y _i{ 1,1} is the classification that is provided by the expert to ∈, and l is an instance number, α _iWith b be the weight that will learn.We use Gaussian radial basis function to kernel function k, and its form is:

$k(\mathrm{\φ}\left(\stackrel{\→}{x}\right),\mathrm{\φ}\left({\stackrel{\→}{x}}^{\′}\right))=e^{-{|\mathrm{\φ}\left(x\right)-\mathrm{\φ}\left(x^{\′}\right)|}^2/2{\mathrm{\σ}}^2}---\left(4\right)$

Obviously, the nuclear of this form (condition that meets Mercer) has corresponding optimization problem, and they are protruding, do not have local minimum.

So for example neural network is different with other pattern recognition system, guarantee that SVM can concentrate on the step of limited quantity.We adopt the Matlab embodiment of SVM.

Below be that cardiac muscle is divided into the used logic of a plurality of segmentations.Two layers of radial at cardiac muscle fan-shaped (interior and outer) goes up the classification that active and non-viable regions occur.Cardiac muscle is described as heart membrane boundary endo (θ)=(x _Endo(θ), y _Endo(θ)) and epicardial border epi (θ)=(x _Epi(θ), y _Epi(θ)), center line centerline (θ)=(x _Centerline(θ), y _Centerline(θ)) in the two centre.

Cardiac muscle is divided into even number n fan-shaped S.Fan-shaped S in n/2 is arranged _InnerWith n/2 outer fan-shaped S _Outer,

S＝S _inner+S _outer (5)

Interior fan-shaped $S_{\mathrm{inner}}=\underset{i=0}{\overset{\frac{n}{2}-1}{\mathrm{\Σ}}}{s}_{\mathrm{inne}{r}_i}$ With π i/n≤θ＜π (i+1)/n and

Endo (θ)≤r (θ)＜centerline (θ) is the boundary, and fan-shaped outward $S_{\mathrm{ouler}}=\underset{i=0}{\overset{\frac{n}{2}-1}{\mathrm{\Σ}}}{s}_{\mathrm{oule}{r}_i}$ With π i/n≤θ＜π (i+1)/n and centerline (θ)≤r (θ)＜epi (θ) is the boundary.

Adopt the difficulty of SVM to be to select feature.Feature is meant and can makes the measurement that appropriate medical diagnosis is carried out to activity or non-viable myocardial tissue.Some features that can comprise are brightness of image, chamber wall thickness, the evidence that heart chamber wall thickens, the homogeneity of image-region, the degree (saturating wall degree) of clear zone (scar) extend through ventricle wall and the position of scar.In addition, also can use contrast take in speed, from movable information that obtains such as physiotherapy apparatus such as Tagged-MR and the brightness that from nuclear research, obtains etc.The professional and technical personnel understands that described characteristic only is demonstration, and other characteristic also can be included, and these do not deviate from scope and spirit of the present invention.Can also understand, can adopt weight scheme to distinguish one or more features importance of other one or more features relatively.

Select correct feature will make the described decision function can separate instance; Not enough or too much feature can cause subregion not good.Definition described feature φ (x) on above definition fan-shaped.Internal sector S _Inner, the thickness of definition is expressed as:

$T_{S_{\mathrm{inner}}}=\underset{\frac{m}{n}}{\overset{\frac{\mathrm{\π}(i+1)}{n}}{\∫}}(\mathrm{centerline}\left(\mathrm{\θ}\right)-\mathrm{endo}\left(\mathrm{\θ}\right))\mathrm{d\θ}---\left(6\right)$

In like manner, the thickness of the outer sector of definition.Should be pointed out that a pair of interior sector and outer sector will have equal one-tenth-value thickness 1/10 because for given radially boundary, center line is equidistant between myocardial boundary.Thicken the change that is defined as from diastasis to end-systole sector thickness.

The mean flow rate M of sector _SBe defined as

$M_s=\frac{\underset{p\∈s}{\mathrm{\Σ}}{I}_p}{\underset{p\∈s}{\mathrm{\Σ}}1}---\left(7\right)$

I in the formula _pBe illustrated in the brightness of pixel p among the s of sector.Homogeneity H _SBe expressed as

$H_s=\frac{\underset{p\∈s}{\mathrm{\Σ}}{\mathrm{\σ}}_1^2\left(p\right)}{\underset{p\∈s}{\mathrm{\Σ}}1}---\left(8\right)$

σ in the formula ₁ ²(p) represent the p variation of 3 * 3 adjacent areas on every side.The saturating wall degree U of sector _sIn being also included within

$U_s=\underset{\frac{\mathrm{\πi}}{n}}{\overset{\frac{\mathrm{\π}(i+1)}{n}}{\∫}}g\left(\mathrm{\θ}\right))\mathrm{d\θ}---\left(9\right)$

G in the formula (θ) representative the longest continuous band from endocardial border (if interior sector) to the brightness enhancing display pixel of center line on the θ direction.

The present invention adopts minimum (SMO) technology of optimizing of sequence to determine weight from equation (3) _i, b.SMO comes work by in each iteration optimization only being limited to the subclass of two weights.This method has advantage aspect the use (kernel matrix does not need to be stored in the storer) of speed (because optimization problem only comprises 2 points that can analyze solution) and storer.

Be definite kernel σ and C in the heart, adopted margin maximization and training organize wrong number minimize between the strategy of compromise, promptly " stay a strategy ".

Consult Fig. 3-5, the existing demonstration screening of SVM and data set and the many recognition features that obtains just utilized illustrates the present invention.According to example with consult Fig. 3, to 14 patients three short-axis slice positions (promptly perpendicular to the ventricle major axis) get DEMR image and cinema MR (Flash or TrueFisp) image (step 302).Utilize the Argus software package of Siemens, the expert is to all images myocardial boundary of drawing.Software is to the internal membrane of heart and the external membrane of heart segmentation (step 304) of left ventricle.The Argus software package provides the full-automatic segmentation of high resolving power TrueFisp image.Cardiac muscle is divided into 36 radially covering of the fans, and each covering of the fan along the periphery segmentation, obtains 72 covering of the fans (step 306) altogether again.Each covering of the fan is defined as activity or non-activity by the expert.

Obtain 38 DEMR image slices (3 sections of each object are because unusual cause has been got rid of 4 sections) altogether and have the true foundation in basis that the expert provides.The basis truth is exactly expert's medical diagnosis.Each experiment is made up of 38 branch experiments, and one of them section is used as " test ", other all sections are as training set (" staying a strategy ").The average accuracy of all 38 test sets is used for determining required parameter σ and the C of SVM algorithm.According to these experiments, we set σ=0.1 and C=10 and obtain average accuracy 87%.

Fig. 4 illustrates the projection of result's 6 dimension decision surfaces to three-dimensional 404, and this is obtained by 38 determined measurement results of DEMR image slices.According to the present invention, measured features be chamber wall thickness (x axle 406), cardiac muscular tissue thicken (y axle 402) and DEMR brightness (z axle 408).Use the feature of being surveyed to set up decision surface then.Decision surface 404 has negative edge (negative margin), illustrates that the feature covering of the fan according to us can not be separated into active and non-activity class fully.Each x represents the non-activity covering of the fan, and each represents active covering of the fan.By measuring the myocardium feature that the patient has discerned, just can utilize decision surface to help determine whether non-viable tissue and their position.

Fig. 5 illustrates the user interface of the demonstration that is used to show myocardial data.After the patient made image and measured its feature (step 308), collected data are applied to decision surface (step 310).According to the result, just can whether contain activity or non-viable tissue make a prediction (step 312) to each covering of the fan.Gained data interface display (step 314).

The data that also comprise relevant MRI scanning on the display.In the upper left corner of display is DEMR image and the particularly patient's of heart left ventricle (LV) 504.As seen from the figure, artery 504 walls seem inhomogeneous in brightness.In other words, some zone of LV wall is brighter than the other parts of wall.

Curve map 526 is drawn brightness, the thickness on each covering of the fan of DEMR and is thickened.Dot-and-dash line is represented basic truth (descending) and the prediction that utilizes SVM.In predicted picture 518, the lower right corner at the interface, white sectors 524 expression non-viable tissue, gray sectors 522 is active uncertain sectors, black sectors 520 is represented active mass.As previously mentioned, these predictions all utilize SVM to carry out.

The medical diagnosis that ground truth image 506, the 512 expression doctors in the lower left corner do according to DEMR.In ground truth image 506, high-brightness region 508,510 expression non-viable tissue.In ground truth sector image 512, white sectors 514 expression non-viable tissue, black sectors 516 expression active masses.

Below to being used for cardiac muscular tissue's segmentation and utilizing artificial intelligence technology to determine describedly to organize whether to be the embodiment of the method for non-viable tissue be described, should be pointed out that the professional and technical personnel can make amendment and changes according to above content.Therefore, obviously, can change disclosed specific embodiment of the present invention, these are changed still within the scope and spirit by appended claims definition of the present invention.More than the invention has been described according to the desired details of Patent Law and feature, desired right and need in appended claims, proposing by the patent certificate protection.

Claims (27)

1. method that is used for the patient's heart non-viable myocardial tissue is carried out imaging and identification said method comprising the steps of:

Obtain the image of the part of described cardiac muscle;

Endocardial border and epicardial edge segmentation with described cardiac muscle part;

The described segmentation of described cardiac muscle part is divided into covering of the fan;

Measure the one or more selected feature of the described covering of the fan of described cardiac muscle;

With one or more feature application of recording in decision surface;

Determine that each covering of the fan contains the active myocardium tissue or contains non-viable myocardial tissue; And

Each covering of the fan that demonstration can illustrate described cardiac muscle with and the image of active mark.

2. the method for claim 1, the image of wherein said acquisition utilizes delayed enhancement magnetic resonance (DEMR) imaging.

3. the method for claim 1, wherein said cardiac muscle part is a left ventricle.

4. the method for claim 1, wherein selected feature is a brightness of image.

5. the method for claim 1, wherein selected feature is the thickness of described cardiac muscle.

6. the method for claim 1, wherein selected feature is the discriminating thickness of described cardiac muscle.

7. the method for claim 1 is wherein selected the uniformity coefficient in the DEMR image that feature is described cardiac muscle.

8. the method for claim 1, wherein selected feature is the saturating wall degree of scar (non-activity) tissue.

9. the method for claim 1, wherein selected feature is described position of organizing covering of the fan.

10. the method for claim 1, wherein selected feature is the absorption speed of contrast preparation.

11. the method for claim 1, wherein selected feature is the material strain message form that is labeled as MR.

12. the method for claim 1, wherein said decision surface utilize artificial intelligence technology to set up.

13. method as claimed in claim 12, wherein said artificial intelligence technology is a support vector machine.

14. the method for claim 1 wherein shows active mass's covering of the fan with first kind of color, and shows the non-viable myocardial tissue covering of the fan with second kind of color.

15. one kind is used for patient's the myocardium part imaging and the system of identify non-viable myocardial tissue, described system comprises:

MR imaging apparatus is used to obtain the DEMR image of described cardiac muscle;

Processor, be used for myocardium part segmentation with the patient, be divided into endocardial border and epicardial edge, described processor also is divided into covering of the fan with the myocardial wall of described cardiac muscle part, described processor measure the one or more selected feature of described myocardial wall covering of the fan and with one or more selected feature application in decision surface, so that definite each covering of the fan contains the active myocardium tissue or contains non-viable myocardial tissue; And

Display is used to show each covering of the fan that described myocardial wall can be shown and its activity mark's image thereof.

16. system as claimed in claim 15, wherein said cardiac muscle part is a left ventricle.

17. system as claimed in claim 15, wherein selected feature is a brightness of image.

18. system as claimed in claim 15, wherein selected feature is the thickness of described myocardial wall.

19. system as claimed in claim 15, wherein selected feature is thickening of described myocardial wall.

20. system as claimed in claim 15, wherein selected feature is the uniformity coefficient of described myocardial wall.

21. method as claimed in claim 15, wherein selected feature is the saturating wall degree of scar (non-activity) tissue.

22. method as claimed in claim 15, wherein selected feature is described position of organizing covering of the fan.

23. the method for claim 1, wherein selected feature is the absorption speed of contrast preparation.

24. the method for claim 1, wherein selected feature is the material strain message form that is labeled as MR.

25. system as claimed in claim 15, wherein said decision surface utilize artificial intelligence technology to set up.

26. system as claimed in claim 26, wherein said artificial intelligence technology is a support vector machine.

27. system as claimed in claim 15 wherein shows active mass's covering of the fan with first kind of color, and shows the non-viable myocardial tissue covering of the fan with second kind of color.

Images