JP2632501B2 - Method for extracting internal organs contours in vivo using multiple image frames of internal organs - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention generally relates to a method for extracting a boundary or an outline of an internal organ based on image data. More specifically, the present invention is directed to processing image data by a plurality of image frames. , For extracting contours of internal organs.

[0002]

2. Description of the Related Art Contrast ventriculography is a routine procedure for performing cardiac catheterization in a clinical setting. A catheter is inserted percutaneously into the heart to perform, for example, cardiac volume measurement and / or flow rate measurement. A ventricular angiogram is an X-ray image representing the inside (or endocardium) surface of the ventricle as a graphic. These images typically estimate the boundaries within the heart at end-diastole, ie, when the heart is full of blood, and at end-systole, ie, when the heart is at the end of contraction in the cardiac cycle. By tracing the contours or boundaries of the inner surface of the heart at the beginning and end of the cardiac cycle, a physician can measure the size and function of the left ventricle and diagnose certain abnormalities or defects in the heart. A radiopaque fluid is injected into a patient's left ventricle to produce a ventricular contrast diagram.
The X-ray source is located on the same line as the heart, and a projection is created that outlines the endocardial surface of the myocardium (myocardium). The left ventricular silhouette image is visualized by contrast between the radiopaque fluid and other surrounding physiological structures. Usually, the contour is extracted by manual intracardiac delineation. However, this method requires time and a considerable amount of training and experience to perform correctly. Alternatively, healthcare professionals can visually evaluate the ventricular angiogram and estimate the intracardiac contour, but such evaluations are often little different from empirical guesses, and This is especially true if the contrast diagram was generated at or near end systole. Automated boundary detection techniques that yield much more accurate results in much less time than manual evaluations will clearly dominate. A number of automatic boundary detection algorithms have been developed to address the above problems. In U.S. Pat. No. 5,268,967, a variety of prior image processing techniques were considered to improve the extraction techniques by which images could be analyzed to identify specific parts of the body. Thus, density-gradation conversion based on histograms is a simple and effective way to adjust the contrast of the image, but distinguishes the desired foreground part of the image from the background clutter and identifies the object in question. It has been found that other techniques are required to distinguish between foreground and background. After examining what else has been done to achieve this goal and the problems of these techniques, this patent application discloses a method that can be applied to any digital X-ray image. The methods disclosed in this patent application include methods of edge detection, region division, region classification, region reclassification, and bitmap creation. More specifically, after first detecting the edges of the object, the image is divided into non-overlapping adjacent pixel regions, and classified into foreground, background, and object for each region. In the region classification process, the state of each region in each of the 10 states is estimated based on a series of clinically and empirically determined determination rules. By the evaluation of the fine structure in each area, the area classification is re-classified, and a binary image that functions as a template for image processing performed thereafter is created. Another automatic border detection technique is based on identifying the density gradients that make up the image. In this prior art, a density gradient threshold is applied to the processing of the density image data of the ventricular angiogram to identify the left ventricle boundary. Other processing is applied to increase the accuracy of identifying the boundaries. Another possibility is that the combination of the index, the recognizable shape, and the density value can always be tracked, thereby determining the direction and speed of the movement represented as a flow vector. By analyzing the flow vector pattern, the movement of the internal organs can be evaluated. However, this flow vector does not directly indicate the contour of the internal organs. Still other techniques sometimes extract the contours of internal organs by digital subtraction. The mask image is recorded before the radiopaque contrast agent is injected into the internal organs. The mask image may include radiopaque internal organs that tend to impair the recognition of internal organ contours such as ribs and spine. A radiopaque contrast agent is injected into the internal organs, a second image is created, and the mask image is digitally subtracted from the second image. Thereby, the clutter of the second image, that is, the internal organs that are not interested is subtracted. However, it is difficult to match the mask image and the second image recognizable by the internal organs with the radiopaque contrast agent so that they do not shift.
It is difficult to actually implement this technique. An application of this technique is time interval delay subtraction. In this technique, an image created immediately before is subtracted from an image to be analyzed, and an image showing a difference between two images includes only an internal organ part that has moved within a time interval between the two images. But,
Internal organs that have not moved within the time to create the two images cannot be described. The morphological opening / closing process can also be used for image data processing for extracting the outline of the object. Such techniques are often more generally applicable, such as, for example, with regard to artificial vision systems, and are not restricted to physiological aspects. "M
edical Image Analysis using Model-Based Optimizati
on ", James S. Duncan, Lawrence H. Staib, Thomas Bir
kholzer, Randall Owen, P. Anandan, IsilBosma (IEEE,
1990) proposes the use of mathematical models based on empirically estimated data for the analysis of medical diagnostic images. In the second example discussed in this paper, a parametric shape model with a measure of the length of the boundary line determined from the image is used. The use of empirical data in the model improves its accuracy, but the results still have some instability.

[0003]

Most previous techniques for contour extraction have focused on the analysis of a single image. Even the best current automatic algorithms for contouring internal organs such as the heart have very low success rates and usually require human correction to prevent significant errors. The biggest problem with existing automation techniques is that they do not recognize the expected shape or movement of individual rooms or internal organs that a physician should analyze when performing an image evaluation. At least in order to obtain the accuracy of the image analysis performed by experts, the information obtained from the images must be used as much as possible to describe the surface of the internal organs. In addition, automated systems must perform tasks more efficiently and faster than humans. Finally, it has been clarified that the accuracy with which the automated technique extracts internal organ contours is enhanced by more information than can be obtained from a single image. Analysis of multiple images can provide more information needed to establish higher accuracy, and can provide physicians with more information about the heart and other internal organs than current technology.

[0004]

SUMMARY OF THE INVENTION The method of the present invention, made to solve the above-described problems, comprises an in-vivo method for in-vivo image data of a heart including a series of image frames generated over one cardiac cycle of the heart. Wherein the consecutive image frames include a region of interest in which the ventricles are located, are successively created at approximately equal intervals in time within the cardiac cycle, and each image frame has a density value. Consisting of a plurality of pixels, (a) using learning data obtained by manually determining the outline of a target ventricle of a plurality of other hearts, in order to determine an initial estimation of a region delimited by the outline, As a function of the probability of relying on the density values for the pixels in successive image frames of the training data, the pixels in the image frame of the in-vivo heart are more likely to be inside or outside the contour of the ventricle. Can be in By indicating high classifies pixels in each image frame of the heart in vivo. (b) For successive image frames, the initial estimation of the area delimited by the contour of the in-vivo ventricle is performed for a plurality of points spaced over the contour of the initial estimation, and It is improved by applying dynamic constraints that determine the limits of motion for each point above. In other words, the method of extracting a contour of a ventricle in a living body based on heart image data uses a continuous image frame of one cardiac cycle or more. A series of image frames, including the region of interest where the ventricles are located, are recorded at approximately regular intervals in time. Each image frame is composed of a plurality of pixels, and each pixel has a density value. To perform the initial contour estimation, the pixels of each image frame of the heart in the living body are classified as a probability as a function of the probability of being likely to be inside or outside the ventricular contour. In this process, learning data obtained by manually analyzing the contour of a target ventricle in a plurality of hearts of another individual is used. The probability of the location of each pixel associated with the contour depends on the pixel density values of successive image frames associated with the classification determined by the learning data. The initial estimation of the contour is improved by applying dynamic constraints on the contour in successive image frames. These dynamic constraints limit the motion of each point in the initial estimation of the contour between successive image frames.

[0005] Further, the contour of the ventricle extracted in this manner is determined in order to determine the coordinates of the contour part with low reliability.
It is further improved using the global shape constraint of processing the coordinates of the contour parts with high certainty. The global shape constraint is based on learning data obtained by manually evaluating the corresponding ventricular contours of the discrete heart.

[0006] This method also includes a step of testing the consistency of the contour of the ventricle in the living body extracted as described above. The process is to check the consistency of the extracted contour based on a comparison between the parameters determined by the learning data and the contour.

In the present invention, the learning data is preferably obtained from continuous learning image frames created for one or more cardiac cycles for each of the separate hearts. The learning image frame is also composed of a plurality of pixels, and each pixel is arranged on the vertical axis and the horizontal axis of the learning image frame, and is associated with a density value vector for the position of the pixel through a series of image frames. The ventricular contour of each other heart is manually extracted from each learning image frame. One classification is specified per pixel for a series of training image frames for each heart. The specified classification is as follows. (a) pixels of the learning image frame inside the contour of the manual extraction throughout the whole cardiac cycle; (b) pixels of the learning image frame outside the contour of the manual extraction throughout the whole cardiac cycle; (c) one or more images Pixels that were inside the manually extracted contours in one part of the cardiac cycle that was the frame, and outside the contours in other parts of the cardiac cycle.

When the last classification corresponds to (c), the learning image frame in which the position of the pixel related to the outline has changed with respect to the outline determines the classification of the pixel of each learning image frame. The classification differs depending on the condition regarding the position of each pixel, that is, whether the pixel is inside or outside the contour, or the frame where the position of the pixel has changed (if any) with respect to the contour.

The method also includes the step of specifying the mean, covariance, or classification probability of each pixel in the training image frame for each separate heart. These are based on the density values of the pixels in the series of learning image frames and the classification specified for the pixels. The mean, covariance or classification probability is included in the prior information obtained from the training data.

In the process of improving the initial estimation, a process of determining boundary motion data on the contour of a separate heart, and estimating a component indicating a motion constraint based on the boundary motion data for each learning image frame in one cardiac cycle. It is preferable that a step of performing the above is included.

In one preferred form of the invention, the in-vivo ventricle is the left ventricle, and the image frames are created during the systole and diastole of the cardiac cycle, and the systole and diastole image frames are generated. Are processed separately to extract the contour of the in-vivo ventricle.

It should also be noted that, besides targeting the left ventricle, this method can be applied to other internal organs in the body, the right ventricle, and the left and right atrium.

[0013] The method also includes the step of determining an end-systolic image frame of the in-vivo heart. This is done by a classification method using a frame-by-frame histogram of the density values in the region of interest throughout the series of image frames. This classification method uses a histogram of density values to determine an end-systolic image frame. This image frame usually shows the smallest area where bright pixels are distributed in the region of interest.

The present invention further includes a method of analyzing and evaluating the physiological function of internal organs in a living body for diagnostic purposes using image data. The method determines the contour and its changes through a series of image frames. Assess the physiology of internal organs by observing changes in contours in different image frames in a series of image frames, e.g., whether the ventricular valves are leaking or the muscles, including the heart wall, are weak. It can be confirmed that the outline evaluation based on a single image usually cannot provide the doctor with the same information as the present invention.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS First, an image equipment utilizing the present invention will be described. A typical conventional X-ray imaging apparatus 10 is shown in FIG. Also shown is the equipment required to process X-ray images created by equipment consistent with the present invention and extract and display contours of internal organs. In the X-ray imaging apparatus 10, the X-ray source 12 uses the current supplied from the X-ray source control device 14 as energy, and the X-ray control device 14 supplies the level of the current and voltage sent to the X-ray source 12 and supplies the current. Determine the time interval to be performed. In response to signals sent from the X-ray controller 14, the X-ray source 12 generates an X-ray beam 18. The x-ray beam passes through the chest of patient 20 and enters detector 16. X-ray source 12 and detector 16 are in line along a common long axis, and the X-ray beam passes through a region of interest in the chest of patient 20. The organ to be imaged is usually at the center of the region of interest, ie at the center of the beam 18. After passing through the chest of the patient 20, the beam 18 enters an image intensifier 22 in the detector 16, which converts the x-ray beam into a visible image. A radiopaque contrast agent, such as an iodine compound, is injected into the internal organs to distinguish the internal organs located in the region of interest from the surrounding soft tissues and to more clearly depict the organs. Absorb or block energy. As a result, in the image created by the image intensifier 22, the interior of the internal organ is displayed much brighter than the surrounding background. A partially silver-plated mirror 24 is on the long axis of the detector 16 and directs a portion of the image light from the image intensifier 22 transversely to the long axis of the detector to the lens of a cine camera 26 (optional). Reflect toward. The remaining light of the image passes through the partially silvered mirror 24 and is transmitted along the long axis of the detector 16 to the lens of the video camera 28. In the embodiment shown, video camera 28 generates an analog signal resulting from scanning an image of image intensifier 22. Alternatively, the image of the image intensifier 22 can be projected to a video camera (not shown) mounted outside the detector. The analog signal has a voltage corresponding to each pixel of the image, and the value indicates the density value of the pixel. The analog signal is input to an analog-to-digital converter (AD CONV.) 30. This converter converts the voltage indicating the density value of each pixel into a corresponding digital value. It should be noted that some other video cameras emit a digital output signal directly, and if that video camera is used, the analog-to-digital converter 30 can be omitted. In the preferred embodiment, the range of density values for each pixel of the digital image data is from 0 to 225 (in fact, the darkest and lightest pixels are in a much smaller range). It should be noted that the digital image data created by the X-ray imaging device 10 at time intervals determined by the X-ray source control device 14 is referred to as an "image frame" in this description and in the claims. In an embodiment, a plurality of successive image frames of the internal organs are created at time intervals by setting the X-ray source controller 14 so that the X-ray source 12 is repeatedly turned on for a very short time. In the application of the invention described below, the region of interest imaged by the x-ray beam 18 includes the heart of the patient 20, more specifically the left ventricle. By processing successive digital image frames of the left ventricle created for more than one cardiac cycle, embodiments of the present invention automatically and continuously extract the endocardial surface of the left ventricle at multiple points in the cardiac cycle. Is done. The digital image data of each image frame is stored in the image storage device 32. The image storage device 32 is preferably configured with a large-capacity hard drive or a non-volatile storage medium. Image storage device 32 is bidirectionally coupled to processor 34.
The processor 34 is preferably constituted by a personal computer such as a desktop type or a workstation. A keyboard 36 and a mouse 38 (or other pointing device) are coupled to the processor 34 so that an operator can enter data and / or instructions for controlling the software of the processor 34. This software is for performing the method of the present invention, which processes digital image data stored as image frames in the image storage device 32 and creates contours of the left ventricle at different times during the cardiac cycle. Is to be able to do. Thus, extracted by the method of the present invention,
The outline of the left ventricle on the image frame is displayed on a display 40 connected to the processor 34 for an operator or other user. Further, the outline definition data of the image frame extracted by the processor 34 can be stored in the image storage device 32 in the opposite direction, and can be analyzed or evaluated at any time after the extraction.

Here, a partial cross-sectional view of the human heart shown in FIG. 2 will be considered. This corresponds to a typical projection angle during ventricular angiogram recording, and the shape of the heart is determined by the outer surface 62. Prior to imaging of the left ventricle 64 of the heart 60, a radiopaque contrast agent is injected into the left ventricle. This is to create an image frame including a relatively bright region in the left ventricle 64 with an X-ray apparatus. Although the left ventricle is distinguished in the image at this stage, the endocardium of the left ventricle 64 (or (Internal surface) It turns out that the white shadow surrounded by the outline of 66 cannot be clearly described. However, according to the method, an image frame created by X-rays can be processed, and a contour substantially close to the endocardium in the left ventricle of the patient can be obtained for each image frame. In the embodiment of the present invention, the type of image data to be processed is X
Although the image frames created by the lines are shown, it should be noted that the method is applicable to other types of image data, including data from ultrasound images, nuclear magnetic resonance images. However, it is difficult to interpret the image and extract the contour of the left ventricle in an image frame created by these various techniques. The reason is that the outline of the left ventricle (or other internal organs to be examined) is usually not sharply defined. In particular, with respect to the left ventricle as an example of an internal organ of interest, the location of the apex 72, i.e., the lower part of the contour, is usually in the contrast silhouette of the image frame, the upper part of the left ventricle, the adjacent aortic valve 68 which opens into the aorta. , Is rendered much less sharply than the mitral valve 76, which opens to the left atrium 74. The shape of the left ventricle 64 changes throughout the cardiac cycle, and its cross-sectional view also changes to a minimum at end systole and a maximum at end diastole. The cross section and shape defined by the contour of the endocardial surface
It changes as part of the heart wall of systole and diastole. By assessing image frames for one or more cardiac cycles for left ventricular changes frame by frame, the physician may have noticed that the mitral valve leaked or the myocardium (muscle) of the left ventricular wall had a reduced range of motion and weakened. Can diagnose internal organ problems in the patient's heart. By observing the contour of the heart throughout the cardiac cycle, a physician can quickly detect a physiological dysfunction of the heart. The physician will look carefully to see if there is a problem that does not change the profile from frame to frame as compared to normal heart function. And, for example, if the range of movement is small and there is weak muscle on a part of the wall of the left ventricle, comparing the wall of that part with other parts, studying the relative change of the contour of the left ventricle The situation will be clear to some physicians. This is because portions containing weakened muscles with a reduced range of endocardial movement fail in contraction in several image frames during normal or active systole. Similarly, evaluating the contour through multiple cardiac cycles reveals mitral valve leakage. The ability to automatically extract the contour of the left ventricle immediately after image creation allows the physician to more quickly assess the condition of the heart when performing the relevant medical procedure. This method involves one or more cardiac cycles
Extract the contours of the left ventricle and other ventricles, atria (or other internal organs) of the heart with at least as accurate as a specialist in this type of image evaluation, making this task much faster than humans. Is what you do.

Next, an outline of the method according to the present invention will be described. An outline of the process of automatically extracting the contour of the left ventricle is shown in FIG. As described above, this sequence of digital image frames contains input data for the process of determining the contour of the left ventricle. This input data corresponds to the left ventricular angiogram 100. In block 102,
With respect to the contour of the left ventricle, an initial region is estimated at intervals over one cardiac cycle or more. This initial estimate is a function of the pixel-by-pixel density values of each image frame created during the entire cardiac cycle. In order to give shape to the initial estimation, pixels are classified by a function of the position and density value of each pixel with reference to parameters determined by manually evaluating learning data in advance using Bayesian classification. The learning data described in detail later is image data obtained by imaging left ventricles of a plurality of other hearts. By classifying each pixel of the patient's image frame as being inside or outside the left ventricular contour, an estimated initial left ventricular region is obtained at block 104. The next step is block 106. In particular, the initial left ventricular region estimated in block 104 is further improved by applying dynamic constraints to changes in the contour of the region between successive image frames. This dynamic constraint is also obtained by determining a component from the learning data. That is, for each left ventricle of the heart of the other individual used to create the training data, determining the extent of displacement of a point on the boundary or contour of the left ventricle between two consecutive points in time. Obtained by This component determines the limits on left ventricular contour motion between successive image frames used in making an initial estimate of the left ventricular region for the patient. A further refined and estimated left ventricular region is created by applying a kinetic constraint, which is block 108. As described above, the contour of the apex 72 of the left ventricle is the most difficult part of the endocardium to depict in image data. Usually, a radiopaque contrast agent is injected into the middle left ventricle so that it can be compared on an X-ray image.
There is a physical shape characteristic that the lower apical area often does not mix well with blood. In addition, if the patient's heart has good contractile function, the radiopaque contrast agent may have a physical shape characteristic that it is expelled from the left ventricle before reaching the apex. Therefore, the process of block 110 is performed to more reliably extract a blurred outline of the left ventricle in the region centered on the apex.
At this time, the contour of the region with the apex 72 is determined as a function of the contour other than the region with the apex 72 by applying the comprehensive shape constraint. In this process, learning data collected in advance is used to determine a comprehensive shape constraint. The manually extracted left ventricular contour of another individual's heart is evaluated to confirm the relationship between the contour of the upper left ventricle 2/3 and the contour near the apex. Block using this relationship
The estimated left ventricular region of 108 is evaluated, and the process of block 110 estimates the boundary of the left ventricle LV as described in block 112. The estimated left ventricular boundary in block 112 should represent the contour of the left ventricle fairly accurately, but further in step 114 the estimated left ventricular boundary of each image frame is within the expected range of the contour. Check whether or not.
The process of block 114 simply checks for any abnormalities during the process of determining the estimated contour, and re-evaluates the boundaries for the parameters obtained from the training data for the left ventricle of the separate heart. Make sure.
If the boundaries for each image frame pass this consistency test, then an acceptable left ventricular boundary is obtained at block 116. This allows a physician or other healthcare professional to diagnose the condition of the heart or to observe for other purposes. Because boundaries are estimated for each image frame, changes between image frames are immediately evident, providing meaningful information for a physician to more accurately diagnose a particular disease.

The above method will be described in more detail.
There are some initial assumptions when implementing the method outlined above in detail. First, end-systolic image frames are manually recognized or left ventricular angiogram data.
It must be determined by either automatic recognition by processing image frames containing 100. Secondly,
It is assumed that the borderline shows a monotonous movement during systole and diastole of the cardiac cycle. Third, in order to determine the direction of the left ventricle in the image frame, the aortic valve surface of the heart must be recognizable. Fourth, the left ventricle must be located approximately in the center of the image frame. Lastly, in the process of estimating the initial left ventricle region, it is assumed that a region outside the left ventricle is not included even if the region is outside the region inside the left ventricle. However, with further experience in applying this method, most of these assumptions can be relaxed or eliminated. In addition to the method of confirming that the created left ventricular contour is not regarded as an error because it does not meet the conditions when considering the premise, there is also a premise that it can be eliminated by performing additional processing.

The position of each pixel in the image frame is determined by the coordinates on the vertical and horizontal axes. The density value vector of each pixel is related to the position of each vertical axis and horizontal axis throughout the series of image frames. Further, in this method, a "feature vector" is used for each pixel. This “feature vector” is a density value vector through a series of image frames, and has additional elements indicating the position of the vertical axis, the horizontal axis, and the distance from the center of the region of interest. The learning data determines a set of feature vectors for each classification. In the example, the training data is normalized, 12 image frames for systole and diastole
18 image frames are created. The end systolic image frame can be determined manually, but is shown in FIG.
It is preferable that the end-systolic image frame is automatically determined as shown in FIG. The left ventricular angiogram of block 100 is input to block 120, which extracts a region of interest (ROI) using the feature vector of the pixel. In connection with this process, a simplified taxonomy is utilized at block 122. Taxonomy 122
Uses parameters obtained from learning data created for the left ventricle of a plurality of other hearts. The training data is processed to identify three regions from the manually extracted contour training image frames. And the pixels of the learning image frame are
It is classified into any one of these three areas. This 3
The two areas are: (1) a light background area (which is excluded from the area of interest because it indicates a mechanical shutter used in the imaging device); (2) a dark background area; and (3) a foreground area ( That is, a region inside the contour extracted manually. The contour set for the left ventricle of the heart for all image frames and for every discrete body is extracted by collecting all pixels classified as being in the foreground region in the training data. Based on this set of contours, the region of interest in the training data and the average of the density values of the pixels in each region are determined. As described in block 120, the classification method 122 uses the left ventricular contrast map. The region of interest of the data 100 can be determined. The process of block 120 is performed as follows. As shown in block 124,
All the pixels of the image frame included in the left ventricular angiogram data 100 are evaluated and extracted based on whether or not the pixel is within the region of interest by using the feature vector of the pixel by the classification method 122. Dark backgrounds are also defined by the taxonomy 122, as shown in block 121. This process is a left ventricular angiogram data
For each pixel contained in 100, the extraction process is repeated until all pixel positions have been processed, and for a series of images contained in the left ventricular angiogram data, a bright region of interest in block 124 and a block 121 A dark background is obtained by extraction. The next step 123 is to consider the density values of the pixels in the dark background in block 121 for each image contained in the left ventricular angiogram data 100. Obviously, the relative density values for each pixel of the image frame can vary significantly depending on differences in x-ray intensity, detector sensitivity, and other factors. Therefore, as shown in block 123, a histogram is created for the pixels in the dark area of each image frame. Each dark background histogram includes a separate sum that indicates the number of pixels having the same density value in the image frame. That is, it is the sum of the number of pixels for each different density value in the image frame. The dark background histogram defines the normalization factor shown in block 125. As described below, the normalizing element compensates for the deviation of the relative density value between the image frames. Region of interest in block 124 and block 125
The dark background normalization element is input to the process of identifying end-systolic image frames shown in block 126. At end systole, the left ventricle should show a minimal area. That is, an end-systolic image frame often coincides with an image frame in which the region having relatively bright pixels in the region of interest is the smallest. Therefore, a histogram is created by the pixels in the region of interest of each image frame. This histogram includes the total number of pixels having the same density value, which is calculated for each density value for the pixels in the region of interest. The histogram of the region of interest for each image frame is further normalized using the normalization factor of block 125 with respect to the deviation of the relative density values between each image frame. Each normalized region of interest histogram is
It is used as a feature vector (different from the feature vector obtained from the pixel vector) in the classification method 128 for determining the end-systolic image frame. Classification method 128 uses parameters obtained from training data generated by a plurality of separate hearts. From the manually extracted contour, the region of interest including the manually extracted contour, the dark background region by the classification method 122, and the ventricular volume calculation obtained from the density value histogram of the region of interest in each image frame of the training data, the end systole is calculated. The training data is processed to identify image frames. That is, as mentioned for block 123,
A dark background histogram is created for each image frame of the training data, defining the dark background normalization component of the training data just as described with respect to block 125. The region-of-interest histogram of each image frame of the learning data is normalized using the normalization element of the learning data. Finally, based on the identification of the normalized region of interest in each image frame of the training data and the end-systolic image frame for each heart in the training data, the classification method 128 performs left ventricular imaging as described at block 126. Figure data 100,
That is, an end-systolic image frame can be determined for the image data of the patient.

The process of estimating the initial left ventricular region in block 104 is shown in FIG. Prior to estimating the initial left ventricular region, the left ventricular angiogram data is pre-processed and normalized at block 132. Details of block 132 are shown in FIG. Referring to this figure, at block 160, "noise" is removed from the left ventricular angiogram data 100 by the morphological opening and closing process as shown at block 162. For a general rudimentary discussion of morphological opening and closing, see "Im
age Analysis Using Mathematical Morphology "by R.
M. Haralick, SR Sternberg, and X. Zhuang, IEEE
Trans.on Pattern Recognition and Machine Intelli
gence, Vol. 9, No. 4, pages 532-550. Morphological closure involves manipulating left ventricular angiogram data by brightening a relatively dark portion of a bright region of an image frame. This process is similar to leveling the lower part of the mortar with a trowel. To better understand this stage, each image frame is projected vertically onto a two-dimensional plane with each pixel
It is good to show the density value by height in the dimension direction. In this case, a small area where pixels having relatively low density values are gathered appears as a small dent in the three-dimensional direction within a large area where pixels having relatively high density values are gathered. The morphological closing process is used to increase the density value of this small area of pixels and to approach the density value of the surrounding bright pixels. Similarly, the morphological opening process lowers the relatively high density value of a pixel in a light part in a dark area so that the density value of that part approaches the low density value of a surrounding pixel. One of the features of these morphological opening and closing processes used to remove noise is that they do not significantly change the abrupt transition between relatively bright and dark areas. Earlier noise filtering techniques tend to blur such abrupt transitions. Block 16 by removing noise from left ventricular angiogram data
4 filtered left ventricular angiogram data is created. The left ventricular angiogram data 100 is image data for one cardiac cycle or more, and the number of actual image frames constituting the image data varies depending on the heart rate of the patient 20. Accordingly, at block 166, the filtered left ventricular angiogram data of block 164 is normalized with respect to heart rate. This normalization is performed by the conventional interpolation method shown in block 168 to obtain a predetermined and required number of fixed image frames. As shown in block 170, the purpose of the interpolation process is a total of 30 image frames (in the embodiment), which are thinned out so as to be substantially equally spaced in time during the cardiac cycle and omit extra image frames, that is, The purpose is to confirm that there are 12 systolic and 18 diastolic image frames. It is also possible to implement the present invention by normalizing the number of image frames different from the number used in the embodiment. The brightness deviation of the image frame resulting from the above image processing must be calculated by normalizing the maximum and minimum density values.
This process is indicated by block 172. The normalization process refers to the scaling of block 174 to determine if the range of density values applies to all of the series of image frames and is within the desired range (0 to 225 in the example). The left ventricular angiogram data 134 thus normalized is created in the process of block 132.

Returning to FIG. 5, the normalized left ventricular angiogram data is input to block 136, where the a priori information 156 is used to classify the pixels of each image frame. Here, the prior information 156 is obtained from the outline of the left ventricle manually extracted for the hearts of a plurality of different individuals, and is determined as follows. At block 138, a plurality of left ventricular angiogram data sets for another plurality of hearts are obtained as training data (created by the X-ray imaging apparatus 10 of FIG. 1 described above). The population selected for the purpose of testing this treatment consisted of about 70 individuals, of which 20 had a "normal" heart with no observed disease or functional problems. Each image frame in the left ventricular contrast map data 138,
And, for each individual heart in the population from which the training data was collected, such image data interpreting physicians manually extract the boundaries or left ventricular contours shown in block 140.
The left ventricular angiogram data for this heart is blocked.
The pre-processing and normalization of the patient's left ventricular angiogram data 100 at 132 is pre-processed at block 142. It should be understood that the entire process of generating prior information 156 from the training data is performed prior to the analysis of the left ventricular angiogram data 100 for the patient. The image data storage device 32 is used to store prior information in the process of block 136. By pre-processing all image frames of the training data, a training image frame of the normalized left ventricle contrast data shown at block 146 is created. On the other hand, in the block 144, the inside of the contour obtained in the block 140 is painted out and normalized to 30 learning image frames per cardiac cycle, thereby creating a manually extracted area in the block 148. An image frame including the manually extracted region of the systole in the cardiac cycle of block 148 is input to block 152. A classification is designated for each pixel in a series of image frames for each left ventricle, and classification is performed according to the position of the block 154. Similarly, all the pixels in the training image frame of the left ventricular angiogram data 138 of the other heart are processed in the expansion period, and the cardiac cycle portion is also classified by position. The step of specifying the pixel classification at block 152 is illustrated in FIG. This is a simplified diagram for explanation, and only three image frames 82, 84, 86 are shown (in the example, actually 12 systolic, 18 diastolic). There is an image frame). Although only four pixels 90a, 90b, 90c, and 90d are shown for each image frame, it is assumed that each image frame includes many pixels. In learning image frames, including those created during systole, some pixels, such as 90d, are always outside the manually extracted contours in all learning image frames. Similarly, in several or more image frames, such as 90a, some pixels are always inside the outline of manual extraction in the systolic image frame. Finally, the pixels of each training image frame, such as 90b and 90c,
It is inside the manually extracted contour in one or more image frames, and outside the manually extracted contour in the remaining systolic image frames. Table 1 shows how the relative position of the corresponding pixel in the systole of the cardiac cycle is applied when specifying a classification number by position for that pixel, the pixel for the contour of the manual extraction. Are shown in the table. “0” indicates that the pixel is outside the outline, and “1” indicates that the pixel is inside the outline.

[0022]

Table 1 Table 1 Pixel position with respect to frame outline 1st sheet 0 0 0 0 1 1 1 1 2nd sheet 0 0 1 1 0 0 1 1 3rd sheet 0 1 0 1 0 1 0 1 1 Classification number 0 1 2 3 4 5 6 7

If the corresponding pixel is always “0” outside the outline of the manual extraction in all image frames, the classification number is 0 from Table 1. Therefore, the pixel 90d has the classification number 0. Similarly, in this simplified example, the corresponding pixel is always the pixel inside the contour in the image frame.
The classification number 7 is designated for 90a. For other pixels, the classification is determined by the image frame whose corresponding pixel has changed its position from inside the contour to outside. In this way, the pixel 90c is inside the manually extracted outline in the first image frame, but outside the outline in the second and third image frames. Therefore, this pixel is classified as classification number 4. According to the same theory, the pixel 90b is inside the manually extracted contour for the first and second sheets, but is outside for the third sheet,
Therefore, the classification number is 6. The total number of classification numbers can be mathematically 2 ^N where ^N represents the number of image frames. However, not all classification numbers actually exist.
This is because the movement of the heart is limited. The frame 88 in FIG. 13 shows the classification by four possible positions for this simple example. Based on the above process, a classification by position is specified for pixels in each systolic learning image frame of all left ventricles of other hearts. Similar processing is applied to the learning image frame in the diastole. However, in the case of the diastole, the movement which is the reverse of the movement of the systole, so that the pixel (relative position “0”) that was initially outside the image frame moves to the inside of the manually extracted contour. Is classified in relation to the image frame. Accordingly, block 154 indicates that the position classification number has been determined for each pixel in each systolic and diastolic image frame. In the processing on the right side of the block 102 in FIG. 5, the image frames processed in the blocks 142 and 144 are the same. The pixels in each set of learning image frames are grouped based on the classification number according to the position assigned to the pixel, and parameters such as the average value, covariance, and classification probability constituting the a priori information 156 are specified in block 150. Is done. The average value is the average of the density values of all the vectors corresponding to the classification number at the same position, and the covariance is a measure of the degree of deviation from the average density value of each vector corresponding to the classification number at the same position. The prior classification probability is simply the ratio of the number of pixels corresponding to a certain classification number to the total number of classification numbers corresponding to all pixels. As shown in block 136, the a priori information obtained for systole is used to classify the pixels of the systolic image frame, and the a priori information obtained for diastole is used to classify the pixels of the diastolic image frame. The details of this process are shown in FIG. Referring to FIG. 7, the data relating to the normalized left ventricular angiogram in block 134 is subjected to a process of adding the coordinates of the vertical axis and the horizontal axis, which are the pixel positions, in block 180. The vector 182 is obtained. Each feature vector has
The density value of the pixel at the same position in each image frame is included. The position of a pixel is tracked by the vertical and horizontal coordinates (and distance from the center of the region of interest) of the pixel. In block 184, the classification number with the highest probability is found based on the prior information, and the specified classification number block 186 is obtained. Prior information in block 156 is obtained from learning data including the probability of each class P (C) and the conditional probability of class P (X / C).
In performing the process of block 184, the classification that maximizes the posterior function P (C / X) = P (X / C) * P (C) / P (X) is determined for the feature vector X at this stage. You. Further, each class C and each Xi of class C follow a multivariate normal distribution N (μc, Σc) (FIG. 13).
For a simplified example). Where,

[0024]

(Equation 1)

In the case of δ ² _ci , i = 1,2,3, automatic covariance is obtained, and in the case of δ ² _cij , _i3j = 1,2,3, cross variance is obtained. Once the classification number is specified, block 188 converts the classification number into a series of binary images, which are displayed in binary 1 if the pixel is inside the contour and if the pixel is outside the contour. An image frame displayed in binary 0 and displayed in black and white is created. The result is the initial left ventricle region estimated for each image frame, as shown in block 104.

Turning to FIG. 8, the details of block 106 show how the dynamic constraints between all pairs of frames improve the estimated left ventricular region. In performing this process, training data including boundary motion data (BMO) is input to block 190. The borderline dynamics data is used in the process of estimating components indicative of dynamic constraints at block 192 as described below. In the examples, when generating borderline dynamics data, care was taken to select a left ventricular angiogram of a heart with various diseases that are most commonly found in the general population in the area. In order not to make the components very diverse, it is desirable to obtain borderline dynamics data from the hearts of various regional or racial populations. Details of block 192 are shown in FIG. The borderline dynamics data is input to a block 200 process for normalizing borderline arc length and heart rate data.
Since the learning data is obtained from a plurality of sets of image data created for a plurality of different individuals, the training data is adjusted to various sizes and heart rates of the heart used for creating the boundary line dynamic data. Data normalization is required. The normalized boundary dynamics data 202 is input to block 204, where components indicative of boundary or contour dynamics are estimated. When the boundary dynamics data is created in block 190, the contours in the heart are manually extracted from the training image frames created over one cardiac cycle for the left ventricular angiograms of the other 182 different persons. The image frame is in the cardiac cycle
Interpolation is performed at 30 time points, i.e., 12 times in systole and 18 times in diastole. Using manually extracted contours or boundaries, two consecutive
The extent of the transition between the time points was measured at 100 equally spaced points around the contour for each training image frame comprising the normalized borderline dynamics data at block 202. The maximum transition of introversion and extroversion of contours between image frames is morphologically indicated by the two-dimensional component of block 204. Details of the process of estimating the components are shown in FIG. At block 224, the boundary motion data, which is the movement of the boundary between the two image frames i and j, is used to calculate 100 of the contraction period.
(The learning image frames that are very close in time during the systole and diastole of the cardiac cycle can be used, but the learning image frames that are long in time can also be used.) Is). Similarly block 226
In the above, 100 directional contraction components during the expansion period were extracted from the movement of the boundary between the two image frames i and j.
The process performed in blocks 224 and 226 causes the block
Boundary dynamic components, which are 100 directional components as shown in 228, are extracted. In block 230, the component that has the greatest motion in each direction is measured, thereby, as shown in block 194, the diastolic and systolic components K _IJ
And L _IJ are obtained. Components K _IJ and L _IJ represent kinetic constraints, as shown in block 194 of FIG. 9, to use the left ventricular contour information of image frame i to improve the left ventricular contour of image frame j. Used for I _t is one image frame at time t, I _s time s
If is a further image frames in, we can predict the boundaries of I _t from the boundary line of I _s.

[0027]

(Equation 2)

Here, L _ts and K _tS are time t and time s, respectively.
Components that indicate constraints on introvert and extrovert movements between the two provide an estimated left ventricular region of block 108 by improving the two image frames during diastole and systole. That is, with this component, one boundary line is given from another image frame of the cardiac cycle, and an annular left ventricular region where the left ventricle should exist can be determined. The kinetic constraints are repeatedly applied to the initial left ventricular region estimation of block 104 in block 196 of FIG. 8 to obtain the estimated left ventricular region of block 108.

Returning to FIG. 10, here, block 11 of FIG.
The process required to estimate the left ventricular boundary using global shape constraints is shown in FIG. In this process, the learning data is used again. As shown, the coordinates of the learning left ventricle boundary of block 210 and block 212 are shown.
Are manually input to block 214. Learning left ventricle boundary line coordinates are shown in Fig. 3
This is calculated or determined by applying the processes of block 102 "Estimate the left ventricle region based on the density over a period of time" and block 106 "Estimate the left ventricle region by dynamic constraints between all image sets". As described above, this determines the estimated contour for each learning image frame and determines the coordinates of the contour. At block 214, the regression coefficients are estimated by estimating the coordinates calculated for approximately the upper approximately two-thirds within each image frame of the training data in relation to the portion of the contour that includes the left ventricular apex. As mentioned above, the apical portion and region of the left ventricular contour cannot be accurately defined in the image frame, and the estimated left ventricular region in block 108 is not very accurate. Accordingly, a global shape constraint is used at block 110 to compensate for this inaccuracy of the apical region of the left ventricle. The regression coefficients in block 216 were generated for at least one cardiac cycle for another heart
From the left ventricular angiogram of 200 or more people, it was obtained by evaluating the contour extracted by the computer, the shape of the part with the apex of the manually extracted contour,
It is used to determine the mathematical relationship between the shape of the upper two-thirds of the contour extracted by the computer. To use the regression coefficients from block 216, block
At 218, the coordinates of the contour of the left ventricle are extracted from the left ventricle region estimated for the left ventricular angiogram data 100 at block 108. The boundary coordinates of the upper two-thirds of the contour are input to a block 222, and a regression process using a regression coefficient in a block 216 is applied. This determines the lower third of the contour, which is the region with the apex of the left ventricle. The result is block 112
Is the estimated left ventricle boundary. Once the estimated left ventricular boundaries have been obtained for each image frame at block 112, the results are preferably tested and the physician preferably determines whether the results are validly reliable as valid contour estimates. Details of the test for consistency of the estimated left ventricular contour are shown in FIG. In this process, the dynamic constraints obtained from the training data are used again in block 231. Applying the global shape constraint to determine the apical portion of the contour may have resulted in changes between image frames contrary to previously applied dynamic constraints. The result of applying the dynamic constraint again is the left ventricle boundary 232 in each image frame. These left ventricular boundaries are
Block 234 applies a global shape constraint, indicated by the curvature and distance from the center of the left ventricle, determined by the training data obtained from the evaluation of the discrete heart based on the manually extracted contours. Block 236 indicates that the resulting data satisfies the left ventricular boundary global shape constraint previously obtained from the training data. If you believe that the data about the resulting border is inconsistent,
A message appears on the display alerting the operator or other user that the boundaries are incorrect, as shown in block 238. In particular, in this process, a warning is issued when the apex is satisfied by the regression processing but the number of irregularities is not satisfied morphologically. if,
Even if the number of irregularities is correct, if the area indicated by the regression processing performed on the apex is too large, the estimated boundary is excluded because the reliability is too low. If no such alert is issued, the boundary data for the image frame of block 236 is
Accepted as 6. The allowed left ventricular border of each image frame thus obtained is at least as accurate as its corresponding manually extracted border. Because, the process of automatically determining the contours and boundaries is
This is because contours obtained from learning data manually extracted in advance by a technician specializing in this type of image interpretation are used. The automated process is for each image frame
After quickly obtaining the image data, the contour can be quickly extracted based on the learning data created in advance,
Therefore, it will be an important and valuable tool used by physicians in cardiology and other medical medical procedures. Although the embodiments of the present invention have been illustrated and described, they can be modified without departing from the spirit and viewpoint of the present invention.

[0030]

According to the present invention, the physician recognizes the expected shapes and movements of the individual ventricles and internal organs to be analyzed when evaluating the image, and utilizes the information obtained from the image as much as possible to make the internal organs available. Since the surface of the image is estimated and described, at least the same level of accuracy as image analysis performed by an image evaluation specialist of this kind can be performed more efficiently and more quickly than a human. Also, the accuracy with which the automation technology extracts the contours of internal organs is enhanced by more information than can be obtained from a single image. In addition, the analysis of multiple images provides more information needed to establish higher accuracy and can provide physicians with more information about the heart and other internal organs than current technology. The physician can more quickly assess the status of the heart and other internal organs when performing related medical procedures.

[Brief description of the drawings]

FIG. 1 is a schematic diagram showing an apparatus for creating an X-ray image of a built-in tissue and an instrument for creating an image for extracting a tissue contour according to the present invention.

FIG. 2 is a cross-sectional view of a human heart including a left ventricle.

FIG. 3 is a flowchart showing an entire process of extracting a contour of a left ventricle at a plurality of points in a cardiac cycle based on a plurality of images created through the cardiac cycle, applied to the present invention.

FIG. 4 is a flowchart illustrating a processing theory for determining an end-systolic image frame from continuous image frames of one cardiac cycle or longer.

FIG. 5 is a flowchart illustrating a process of estimating a left ventricular contour based on density values of pixels constituting an image frame.

FIG. 6 is a flowchart showing a process of preprocessing and normalization applied to image frame data.

FIG. 7 is a detail of the process of FIG. 5 relating to pixel classification for determining a binary image.

FIG. 8 is a flowchart illustrating a process for more accurately estimating the contour of the left ventricle based on the dynamic constraints obtained from the analysis of the separate left ventricle.

FIG. 9 is a flowchart illustrating a process for estimating a dynamic constraint associated with FIG. 8 in further detail.

FIG. 10 is a flowchart showing a process of estimating a contour near the apex of the left ventricle by applying a comprehensive shape constraint.

FIG. 11 is a flowchart showing details related to estimation of components used when determining a dynamic constraint.

FIG. 12 is a flowchart illustrating a process of testing the consistency of the estimated left ventricular contour.

FIG. 13 is a flowchart illustrating a method of specifying a classification for corresponding pixels in a plurality of image frames.

 ──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-2-22038 (JP, A) JP-A-4-146729 (JP, A) JP-A-58-130030 (JP, A)

Claims (19)

(57) [Claims] 1. A method for automatically extracting a contour of an internal organ officer based on digital image data of an area where the internal organ officer is located, wherein the image data is obtained when the internal organ officer wall is integrated with a physiological function of the internal organ officer. Represents a continuous image frame created in a period of completing one or more cycles of the repetitive motion, each frame of the continuous image frame has a plurality of pixels, (a) For an individual different from the extraction target, The outline is manually traced in advance for the learning data having an image of one or more cycles of the internal organs to be extracted, and the pixels included in the image frame exist inside the outline of the internal organs to be extracted. Using the parameters obtained in advance by classifying the pixel based on the probability, the image data to be extracted is compared by comparing the image data to be extracted. The region near the contour of the internal organs in a sequence of image frames comprising initial estimate. (b) Determine the permissible range of contour movement between consecutive frames in a series of image frames, based on boundary dynamic data obtained in advance by manually evaluating the target internal organs of other individuals. This improves the initial estimation for each of a series of image frames to extract the contour. A method for extracting an internal organ contour in a living body using a plurality of image frames of the internal organ. 2. In the step (b), a region estimated for each image frame is more accurately improved by applying a comprehensive shape constraint on a physical shape characteristic of an internal organ in a region delimited by a contour. 2. The method of claim 1, further comprising the step of:
The described method. 3. The process of evaluating the consistency of the outline of each image frame by comparing the outline of each image frame with constraints obtained from the learning data, and asking the operator whether the outline is out of the constraints. 3. The method of claim 2, further comprising the step of indicating.
The described method. 4. The process of obtaining an initial estimate of a region bounded by internal organs in individual image frames of a series of image frames, the method comprising the steps of: The method of claim 1, further comprising the step of normalizing to 5. The parameters are: (a) elaborately examine the contours of internal organs in each image frame of learning data relating to another individual and manually extract the parameters from the learning data; (b) Classify the individual pixels of each series of image frames according to the following: Throughout the cycle, the pixel is (i) always inside the contour. (ii) Always outside the contour. (iii) Transition from one side of the contour to the other. If the pixel transitions from one of the contours to the other as described above, the classification is expressed as a function of the image frame in which the transition occurred. (c) Concerning the target internal organ of another individual, the density value vector of each pixel of the image frame of the learning data is determined through one cycle. (d) For each pixel of the image frame of the learning data, determine an average value vector, a covariance vector, and a classification probability for the entire pixel including parameters obtained from the learning data as a function of the density value vector and the classification. 2. The method according to claim 1, wherein the method is estimated by: 6. An element of a density value vector of a pixel constituting image data of an internal organ whose contour is to be extracted changes with the movement of the internal organ wall, and for each pixel, it is determined from the density value vector and the learning data. The classification is specified based on the obtained parameters, and the process of replacing the classification number specified for each pixel with a binary value indicating a pixel inside the contour and all other pixels outside the contour is also performed. The method of claim 1, which is included. 7. The method according to claim 1, further comprising the step of removing noise from the image data by morphological opening and closing of the image data before the initial estimation of the region delimited by the contour. 8. A dynamic that restricts the movement of the contour between two images based on the maximum movement of the contour between the corresponding continuous learning image frames obtained from the learning data of the target internal organs of another individual. The method of claim 1 including using boundary dynamics data to estimate components for determining constraints. 9. The method of claim 8, further comprising the step of improving the initial estimation of the region delimited by the contour within the limits of motion at a plurality of points around the contour of each image frame. 10. To refine the region bounded by the contour, (a) use the training data to determine the position of the internal organs structural index within the first portion of the contour; Is determined to define a probability distribution as a function of another second portion of the contour that does not include the first portion of. (b) a relatively accurately defined contour portion corresponding to the second portion of the contour of the training image frame in each image frame, estimated in step (b) according to claim 1 for each of a series of image frames; Is extracted. (c) applying the regression coefficients in the regression process to the contour coordinates of the relatively accurately defined portion of the contour in the individual frames of the series of image frames, to define less accurately in each frame of the series of image frames; More precisely determine the other parts of the contour that have not been done. 3. The method according to claim 1, further comprising the steps of: 11. An in-vivo ventricle contour extraction method based on heart image data including continuous image frames created over one cardiac cycle of the heart, wherein the continuous image frames have a ventricle. It includes a region of interest located and is created consecutively at approximately equal intervals in time within the cardiac cycle, and each image frame is composed of a plurality of pixels having a density value. Using learning data obtained by manually determining the contour of the target ventricle, the probability of relying on the density values of pixels in successive image frames of the learning data to determine an initial estimate of the area bounded by the contour By indicating whether pixels in the image frame of the in-vivo heart are likely to be inside or outside the contour of the ventricle as a function of Picture frame To classify the pixels in. (b) For successive image frames, the initial estimation of the area delimited by the contour of the in-vivo ventricle is performed for a plurality of points spaced over the contour of the initial estimation, and It is improved by applying dynamic constraints that determine the limits of motion for each point above. A method for extracting an internal organ contour in a living body using a plurality of internal organ image frames. 12. An area delimited by the ventricular contour determined in the step (b) is processed by using coordinates of a more reliable portion of the contour using a comprehensive shape constraint based on learning data. 12. The method according to claim 11, further comprising the step of determining the coordinates of other unreliable parts. 13. With regard to the consistency of the in-vivo heart contour thus determined consistently, the contour is obtained from the learning data based on the manual extraction of the target ventricle contour of another heart. 12. The method of claim 11, further comprising testing by comparing with the obtained constraint. 14. The parameters obtained from the learning data, comprising: (a) creating a plurality of learning image frames for one heart cycle or more for each of the other hearts, wherein the learning image frames are composed of a plurality of pixels; Each pixel is arranged along the vertical and horizontal axes of the learning image frame, and has an associated density value vector. (b) Manually determine the contour for each other ventricle in each image frame. (c) Specify a classification for each pixel in the learning image frame of each of the other hearts. This classification observes more than one cardiac cycle,
The decision is made based on the following. (i) The pixels of the training image frame are inside the manually extracted contour throughout the cardiac cycle. (ii) The pixels of the training image frame are outside the manually extracted contours throughout the cardiac cycle. (iii) a learning image in which the pixel has changed position relative to the contour if the pixel is on one side of the manually extracted contour in part of the cardiac cycle and is on the opposite side of the contour in the rest of the cardiac cycle Based on frame. 12. The method according to claim 11, wherein the method comprises: 15. A method for designating an average value, a covariance, or a classification probability for pixels of the same classification in a training image frame based on a density value vector of each pixel and a classification specified for the pixel. 15. The method of claim 14, wherein said mean, covariance or classification probability is part of prior information obtained from training data. 16. The method for improving the initial estimation includes the step of determining boundary line dynamic data for a target ventricle contour of another heart, and the step of determining boundary line dynamic data for each learning image frame of one cardiac cycle. 12. The method of claim 11, comprising estimating a component indicative of a dynamic constraint based on the component. 17. The improvement process of the initial estimation includes a process of determining borderline dynamic data on the contour of another heart, and a process of defining a dynamic constraint based on the borderline dynamic data on each learning image frame of one cardiac cycle. 13. The method of claim 12, comprising estimating the indicated component. 18. The in-vivo ventricle is the left ventricle, and image frames are created throughout the systole and diastole of the cardiac cycle, and the systole and diastole image frames are processed separately to produce a live image. 12. The method of claim 11, wherein a contour of a ventricle in the body is extracted. 19. The method of claim 18, comprising determining an end-systolic image frame for the in-vivo heart using a classification method utilizing a histogram of the density values of each pixel.