EP1709590A1

EP1709590A1 - Stochastic analysis of cardiac function

Info

Publication number: EP1709590A1
Application number: EP05702590A
Authority: EP
Inventors: Julien T. Senegas
Original assignee: Koninklijke Philips Electronics NV
Current assignee: Koninklijke Philips NV
Priority date: 2004-01-15
Filing date: 2005-01-05
Publication date: 2006-10-11
Also published as: WO2005071615A1; JP2007521862A; CN1910618A

Abstract

A diagnostic imaging apparatus (10) performs stochastic model-based segmentation of diagnostic images of a subject. A plurality of stacks of slice images (22) are generated with each stack displaced in time. Multiple solutions of the organ shapes are computed from the stacks of slice images in the form of a plurality of shape samples (26). The samples are generated (24) according to a Bayesian model describing the conditional distribution of the shape given the images and at least one functional parameter (32) is derived for each of the samples. A probability value (30) is derived for each parameter and displayed (36, 38).

Description

STOCHASTIC ANALYSIS OF CARDIAC FUNCTION DESCRIPTION The following relates to the diagnostic imaging arts. It finds particular application in the estimating of cardiac function from a sequence of magnetic resonance cardiac images.

However, it also finds application in estimating cardiac function from a sequence of heart images from other imaging means and in estimating clinically relevant parameters of other organs from diagnostic images. Estimating cardiac function by reconstructing the shape of the heart is common in the art. Among the methods currently used, a database representing temporal image slices of the heart is first generated. The database may be described as a temporal sequence of slices of the heart, or a stack of slices at temporal intervals, covering at least one whole cycle of the heart. An attempt is then made to delineate the cardiac contours in each heart phase from the stack of slices. Current methods of accomplishing this task are manual or semi-manual using visualization software. For example, a clinician marks boundary points of a heart chamber in each slice of the stack of slices at each temporal interval. The shape (volume) of the heart chamber is determined in each of the temporal intervals, at least for the temporal intervals corresponding to the cardiac phases of interest. From the estimated shapes, functional parameters such as end-diastolic and end-systolic left ventricular volumes, stroke volumes and ejection fraction are computed. With a typical sample of 200 slices, manually delineating the cardiac chambers is time-consuming and error prone. Methods of accomplishing the above-described procedure automatically are currently being tried in the research art. One automated procedure is to fit to the image slices a mathematical model of the heart. This model is based on some prior knowledge on the shape of the myocardium, and possibly a statistical description of the typical shape variations from patient to patient. The automatic segmentation technique generally consists in optimizing an energy functional made of two terms: the first one describes the fit of the model to the data, while the second one penalizes too strong deformations with respect to some reference configuration. The result of this procedure is one optimal segmentation with respect to the mathematical model used to represent the myocardium shape and the mathematical form of the energy used for the optimization. In practice, however, MRI scans are noisy, contain artifacts and only a few slices are acquired, with possible gaps between the slices. As a consequence, the image information does not permit delineating the cardiac chambers without ambiguity. Due to these problems and the like, the segmentation problem is characterized by an inherent uncertainty. This aspect is well known in clinical practice, where it is commonly accepted that the marked boundary points vary from clinician to clinician in manual techniques. One usual method in a clinical setting to deal with this uncertainty is to compare different solutions obtained by different clinicians and to check the discrepancies between the results. The important point is that the uncertainty in the reconstruction of the shape induces an uncertainty in the functional parameters to be computed. However, automated techniques ignore the possibility of multiple solutions, and produce a single solution, supposed to be optimal with respect to the model and technique used, whatever the quality of the data. The clinician is therefore lacking information about the confidence he can have in the results. The present invention contemplates an improved method and apparatus that overcomes the aforementioned limitations and others.

According to one aspect, a method of stochastic model-based segmentation of diagnostic images of a subject, from samples of the subject, is provided. The samples are generated according to a Bayesian stochastic model describing the conditional probability distribution of the organ shape given the images and functional parameters are derived for each of the respective samples. An uncertainty value is estimated for each derived parameter . According to another aspect, a diagnostic imaging apparatus is provided. The apparatus includes a means for generating the diagnostic images and a processor programmed to perform the above-described method. According to yet another aspect, an apparatus for generating stochastic model-based segmentation of diagnostic images of a subject is provided. The apparatus includes a diagnostic imaging scanner configured to perform a scan of a volume region of interest of a subject. A processing system is provided to process scanned data of the volume region of interest from the diagnostic imaging scanner and a provided reconstruction module receives the scanning data and generate a stack of slice images of the region of interest stored in the processing system. A shape module computes shape samples representing multiple solutions of the segmentation of the slice images according to a probability distribution described by a Bayesian model, and a function module derives functional parameters for each of the respective samples and estimates a probability value for each derived parameter. A display module displays the functional parameters and respective probability values on a provided display device. According to still another aspect of the present invention, a diagnostic imaging apparatus performs stochastic model-based segmentation of diagnostic images of a subject. A computing means computes a plurality of shape samples according to a probability distribution described by a Bayesian model. A function means derives at least one functional parameter for each of the respective samples and estimates a probability value for each derived parameter. One advantage resides in an improved robustness of the segmentation and the tracking of the heart chambers. Another advantage resides in the precise information provided to the clinician regarding the accuracy of the functional parameters generated by the automated process. Yet another advantage resides in the means for an improved diagnosis of cardiac function, with an improved level of confidence. Numerous additional advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments.

The invention may take form in various components and arrangements of components, and in various process operations and arrangements of process operations. The drawings are only for the purpose of illustrating preferred embodiments and are not to be construed as limiting the invention. FIGURE 1 diagrammatically shows a magnetic resonance imaging system employing automated estimating of cardiac functional parameters according to concepts of the present invention.

With reference to FIGURE 1, a diagnostic imaging scanner 10, such as a magnetic resonance scanner, includes a housing 12 defining a generally cylindrical scanner bore 14 inside of which an associated imaging subject 16 is disposed. Details of the diagnostic imaging scanner are not shown because magnetic resonance, CT, SPECT, PET, and other suitable scanners are well known in the art. It suffices to say that the diagnostic imaging scanner 10 performs a cardiac scan and communicates scan data to a reconstruction module 18 resident in a processing system 20. The processing system 20 presented herein is not inherently related to any particular computer or other apparatus. In particular, various general-purpose machines may be used with program modules in accordance with the teachings herein, or it may prove more convenient to construct more specialized apparatus to perform the required method steps. Further, the processing system 20 may be a single system or an interconnected distributed system of processors. However, one of ordinary skill in the art will recognize that there exists a variety of platforms and languages for creating modules for performing the functions outlined herein. The reconstruction module 18 generates a series of volume images at temporal intervals during the cardiac cycle. Each of the temporally offset volume images includes a stack of slice images 22. Due to possible poor contrast between the cardiac tissues and the surroundings of the myocardium, image noise, discrete pixel sizes, and the like, the boundary of the cardiac chamber is not without ambiguity. In a preferred embodiment, a shape module or means 24 computes multiple solutions of the heart shape rather than the single optimized solution of prior-art methods. According to Monte Carlo sampling techniques, the Bayesian probability distribution can be represented by a finite number of samples, for example 500. The multiple samples are stored in a myocardium shape sample memory 26. These multiple solutions can be displayed to the clinician for example in the form of an animated shape, possibly superposed on the image slices, so that a qualitative representation of the set of solutions is given to the clinician. A function module 28 then accesses the shape samples 26 and computes cardiac functional parameters and statistical parameters 30 from the multiple shapes 26, for example ejection fraction, end-diastolic volume, end-systolic volume, stroke volume, wall thickness, and the like. In the FIGURE, the exemplary functional parameter is the ejection fraction 32. Rather than computing a single ejection fraction as known in the art, the present invention computes a plurality of ejection fractions from the plurality of shapes in the shape memory 26 and statistical parameters from this plurality of the ejection fractions. In the FIGURE, for example, a histogram is computed, having a relative frequency or probability 34 for each of the computed ejection fractions 32. The functional parameters and statistical parameters 30 are then presented via a display module 36 on a display device 38 for analysis by a clinician or other user of the diagnostic imaging system 10. The display can be in graphical form as a histogram, numerical form as a median value and standard deviation, or the like. In this manner, a clinician is presented along with the solutions, a measure of the degree of confidence in the solutions. In the case of a histogram, the clinician may easily determine if the most probable solution has a high degree of certainty or whether, on the other hand, it is surrounded by a wide band of solutions of relatively high probability, indicating a lower degree of certainty in the solution. Unlike the current art, the degree of uncertainty is not easily overlooked because the relevant statistical information is provided simultaneously with the solution. Alternately, an animation module 39 displays the plurality of solutions in the aforementioned form of an animated shape, possibly superposed on the images, so that a qualitative representation of the set of solutions is given to the clinician. The animation sequence may be ordered according to a predetermined criteria such as, for example, ordered from low probability to high probability, most far from a reference shape to the most close to the reference shape, etc. The animation may also be superposed on the most probable shape. One approach that the shape module 24 uses in computing the shape samples 26 involves the use of Markov chain algorithms as described by W.R. Gilks, S. Richardson and D.J. Spiegelhalter in Markov Chain Monte Carlo in Practice, Chapman and Hall, 1966. To deal with the temporal component of the problem, sequential approaches as described by M. Isard and A. Blake in Condensation — conditional density propagation for visual tracking, International Journal of Computer Vision, 1998. The general approach is to construct a Bayesian probability distribution describing the space of solutions for the segmentation problem, and to generate samples of this probability distribution. If we denote y for the images and z for the shape of the organ, the conditional distribution π(z/y) of the shape, given the images, can be expressed as: ,_{W Λ =}≤_Jι_!__l π(y) where π(y/z) is the likelihood of the shape z, and π(z) is the prior model. The prior model describes the prior knowledge about the shape; it is a probability distribution on the parameters used to describe the shape (for example the coordinates of the nodes of a mesh representing the organ shape, or the coefficients of the decomposition of the surface based on surface harmonics). The likelihood model describes the structure of the images (for example the statistical distribution of the grey values) for a fixed shape. It can be interpreted as a measure of the goodness-of-fit of the shape to the images. The prior model and the likelihood model chosen for the computation of the segmentation form the Bayesian model. Statistics on parameters that are deduced from the shape z can be expressed as the statistical expectation under the probability distribution π(z/y) of a functional f(z): For example, if ρ(z) denotes the ejection fraction for z, the probability that it exceeds the value p₀ can be expressed as: where 1_/ denotes the indicatrice function on the interval /. Monte Carlo integration allows one to compute this integral straightforwardly by generating a sample ..., z_n) of the distribution π(z/y) and to use the approximation: Computing statistics on the functional parameters requires, therefore, generating a finite number of samples of the Bayesian probability distribution π(z/y). This can be done, for example, by generating a Markov chain which has the Bayesian probability distribution as a stationary distribution, see Markov Chain Monte Carlo in Practice for details. The samples span the range of possible solutions. If a temporal sequence of slices is to be segmented, the segmentation consists then in a temporal series of shapes. Such a sample can be sequentially generated, using sequential Monte Carlo techniques. For example, for the first temporal data set, a finite number of shapes representing volumes are generated. For the next temporal step, a motion model is used to predict the next set of shapes. For example, a simple motion model assumes that the myocardium shape contracts by a constant factor at each time step. A weight of each sample, which depends on the likelihood value of the sampled shape for the new time step, is then computed and used in the Monte Carlo integration. The sequential Monte Carlo method is well known in the arts, and used in other applications such as, for example, financial mathematics, tracking, etc. The method, also called particle filtering, is useful for following in time objects in motion. While the invention has been described with reference to magnetic resonance imaging of the heart, myocardium shapes and volumes in particular, it is to be understood that the invention applies equally well to other imaging arts, and to any shape that can be modeled. Further, while the statistical information shown in the FIGURE is a histogram, the computed statistical information can be any useful parameter such as, for instance, standard deviation, confidence interval, etc. The invention has been described with reference to a preferred embodiment. The invention has also been described with respect to several alternate embodiments. These and other variations and modifications of the invention will occur to others upon the reading and understanding of this specification. It is intended that all such variations, alterations and modifications, be included insofar as they come within the scope of the appended claims or the equivalents thereof.

Claims

CLAIMSHaving thus described the system embodiments the invention is now claimed to be:

1. A method of stochastic model-based segmentation of diagnostic images of a subject from a plurality of samples of the subject, the method comprising: generating the samples according to a Bayesian model describing a conditional distribution of the shape given the images; deriving at least one functional parameter for each of the respective samples as a result of the sampling; estimating an uncertainty value for each derived parameter.

2. The method as set forth in claim 1 , further including: propagating the samples in time using at least one of a dynamical model and a stochastic model.

3. The method as set forth in claim 1, wherein estimating the uncertainty value includes: determining a histogram of probability vs. functional parameter.

4. The method as set forth in claim 1 , wherein estimating the uncertainty value includes: determining the probability that the functional parameter is below or above a given threshold.

5. The method as set forth in claim 1, wherein estimating the uncertainty value includes: determining a confidence interval.

6. The method as set forth in claim 1 , wherein the at least one functional parameter is selected from at least one of: a cardiac functional parameter including: left ventricle volume at end-diastole; left ventricle volume at end-systole; stroke volume; ejection fraction; wall thickness; and extent of ischemic or infarcted zone; a cardiac motion parameter including: myocardial strain; and regional wall motion; and a brain property including: cortical thickness; ventricle volume; tissue volume; and volumetric evolution of a tumor.

7. The method as set forth in claim 1 , wherein the diagnostic images are selected from at least one of: X-ray images; CT images; MR images; ultrasound images; SPECT images; and PET images.

8. The method as set forth in claim 1, wherein the diagnostic images include: a plurality of sets of slice images of a region of interest, each set of slice images being displaced in time to present a time evolution of the region of interest.

9. The method as set forth in claim 8, wherein the each slice depicts an interface between an organ and a surrounding structure, a location of which interface is determinable within a range of uncertainty and the generating of the samples includes: for each set of slices, determining a plurality of shapes of the organ.

10. The method as set forth in claim 9, further including: adjusting the determined shapes in accordance with a prediction of a change in interface location from set to set.

11. A diagnostic imaging apparatus comprising: a means (10,18) for generating the diagnostic images; and a processor (20) programmed to perform the method according to claim 1.

12. An apparatus for generating stochastic model-based segmentation of diagnostic images of a subject, the apparatus comprising: a diagnostic imaging scanner (10) configured to perform a scan of a volume region of interest of a subject (16); a processing system (20) configured to process scanning data of the volume region of interest from the diagnostic imaging scanner (10); a reconstruction module (18) configured to receive the scanning data and generate a stack of slice images (22) of the region of interest stored in the processing system (20); a shape module (24) configured to compute a plurality of shape samples (26) representing multiple solutions of the segmentation of the slice images (22) according to a Bayesian model describing a conditional distribution of the shape given the images; a function module (28) configured to: derive at least one functional parameter (32) for each of the respective samples (26) as a result of the sampling; and estimate a probability value (30) for each derived parameter; and a display module (36) configured to display the at least one functional parameter (32) and respective probability value (30) on a display device (38).

13. The apparatus as set forth in claim 12, further including: an animation module (39) configured to display the multiple solutions on the display device (38) according to a predetermined sequence criteria and a predetermined supeφosition criteria.

14. A diagnostic imaging apparatus which performs stochastic model-based segmentation of diagnostic images of a subject, the apparatus comprising: a means (24) for computing a plurality of shape samples (26) according to a Bayesian model describing a conditional distribution of the shape given the images ; a function means (28) for deriving at least one functional parameter (32) for each of the respective samples (26), and estimating a probability value (30) for each derived parameter.

15. The apparatus as set forth in claim 14, wherein the computing means (24) is further configured to propagate the samples (26) in time using at least one of a dynamical model and a stochastic model.

16. The apparatus as set forth in claim 14, wherein the function means (28) is further configured to determine a histogram of probability (30) vs. functional parameter (32).

17 . The apparatus as set forth in claim 14, wherein the function means (28) is further configured to determine the probability that the functional parameter is below or above a given threshold.

18. The apparatus as set forth in claim 14, wherein the function means (28) is further configured to determine a confidence interval.

19. The apparatus as set forth in claim 14, wherein the at least one functional parameter is selected from at least one of: a cardiac functional parameter including: left ventricle volume at end-diastole; left ventricle volume at end-systole; stroke volume; ejection fraction; wall thickness; and extent of ischemic or infarcted zone; a cardiac motion parameter including: myocardial strain; and regional wall motion; and a brain property including: cortical thickness; ventricle volume; tissue volume; and volumetric evolution of a tumor.

20. The apparatus as set forth in claim 14, further including: a means (10) for scanning of a volume region of interest of a subject (16); a means (20) for processing scanned data of the volume region of interest received from the scanning means (10); a means (18) for receiving the scanning data and generating a plurality of stacks of slice images (22) of the region of interest, each stack displaced in time, the computing means (24) computing multiple solutions of the slice images (22) to form the plurality of shape samples (26); a means (36,38) for displaying the at least one functional parameter (32) and respective estimated probability value (30).

21. The apparatus as set forth in claim 14, further including: a means (39,38) for animating and displaying the multiple solutions according to a predetermined sequence criteria and a predetermined supeφosition criteria.

22. The apparatus as set forth in claim 20, wherein the scanning means (10) includes at least one of: an X-ray scanning means; a CT scanning means; an MR scanning means; an ultrasound scanning means; a PET scanning means; and a SPECT scanning means .

23. The apparatus as set forth in claim 22, wherein the each slice depicts an interface between an organ and a surrounding structure, a location of which interface is determinable within a range of uncertainty and the generating of the samples includes: for each set of slices, determining a plurality of shapes of the organ.

24. The apparatus as set forth in claim 23, wherein the computing means (24) is further configured to adjust the determined shapes in accordance with a prediction of a change in interface location from set to set.