JP2007020829A - Method and system for analyzing image data - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and system for analysis which supports discrimination of a tissue by analyzing a time-series multi-level image composed of time-lapse change of a plurality of measured quantities. <P>SOLUTION: A similar area is extracted by a method for analysis having: a step for setting a unified map for normalizing a plurality of images; a step for deforming the time-series multi-level image based on the unified map; a step of setting a distance to a vector composed of the measured quantities at a plurality of times related to respective pixels in the respective pixels of the deformed time-series multi-level image; and a step of clustering the pixels by the distance. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an information system and a medical image diagnostic apparatus equipped with a medical image data analysis method.

  With the development of medical image diagnostic apparatuses, physical quantities representing various physical phenomena have been obtained over time. In particular, in Magnetic Resonance Imaging (MRI), a large number of physical quantities have been measured in order to realize high-precision diagnosis. For example, physical quantities include nuclear spin density, longitudinal relaxation time T1, transverse relaxation time T2, apparent transverse relaxation time T2 * due to magnetic field inhomogeneity, diffusion coefficient, and apparent diffusion coefficient representing the restricted diffusion of tissues and cells ( Examples include Apparent Diffusion Coefficient (ADC), flow rate such as blood flow, and chemical shift indicating the difference between substances such as metabolites. Including secondary information such as a combination of these and changes in the physical quantity due to administration of a contrast medium or the like, a very large amount of measurement has been acquired. In addition, it has become possible to acquire the physical quantities of nuclei called other nuclei such as carbon 13 and phosphorus 31 other than hydrogen nuclei that have been mainly measured. Considering these combinations, the amount of measurement obtained is enormous. In the time direction, a method for acquiring data at high speed, such as Echo Planar Imaging (EPI), has been developed, and images can be acquired at intervals of 1 second. Furthermore, coupled with the development of a high-speed volume data acquisition method, the amount of information that can be obtained in the time direction has become enormous.

  An example of a diagnostic technique using images representing a plurality of measurement amounts is diagnosis of ischemia. In ischemia diagnosis, images of nuclear spin density, T2, ADC, and Perfusion (blood flow volume, blood volume, average arrival time using contrast medium) are measured, and tissue damage is judged in an integrated manner. Another example is a definitive diagnosis of cancer. In the definitive diagnosis of cancer, images such as nuclear spin density, T1, T2, ADC, and metabolite concentration due to chemical shift are measured, and the malignancy and tissue properties are comprehensively determined.

  On the other hand, an example of a diagnostic technique using a time-series image although it is a single measurement amount is a diagnosis of ischemia. In the ischemia diagnosis, changes in the T2 image after contrast medium administration are tracked in order to measure the above-described blood flow volume, blood volume, average arrival time, and the like. Another example is Functional MRI (fMRI) for observing brain activation sites. fMRI measures and analyzes the relationship between sensory and motor stimuli and T2 * changes over time.

  However, a measurement / analysis technique from time-series images of a plurality of measurement quantities has not been established yet. One reason is the dramatic increase in labor required for diagnosis. This is because, as described above, the physical quantity to be observed, its change, and the data amount of one image are increasing, and it is almost impossible to see all of them in a short time. Such an increase in the labor required for the diagnosis may be overlooked or misdiagnosed, and therefore, it is desired to reduce it as much as possible. In order to solve this problem, several methods for analyzing medical image data have been proposed.

  Non-Patent Documents 1 and 2 propose a method of clustering pixels using nuclear spin density, T2, ADC, and Perfusion in order to discriminate the cerebral ischemic region. Non-Patent Document 3 proposes a method of clustering pixels using nuclear spin density, T2, and ADC in order to discriminate tumor tissue. In the case of clustering pixels, an appropriate distance is introduced into a multidimensional space having a plurality of measured physical quantities as axes, and pixels having close distances are collected as the same organization using, for example, the k-means method. However, in these methods, analysis in the time direction is not performed, and tissue discrimination at a temporary point is repeatedly performed.

  Non-Patent Document 4 proposes a method of performing multidimensional analysis using nuclear spin density, T1, and T2 in order to discriminate atherosclerotic tissue. In this method, clustering is performed in consideration of not only the physical quantity of each pixel but also the spatial distance of each pixel. That is, in Non-Patent Documents 1, 2, and 3, the spatial distance between pixels was not considered, but in Non-Patent Document 4, not only a physical quantity axis but also a multidimensional space with a spatial distance as an axis. An appropriate distance is introduced. Thereby, there is an effect that the tissue having the same property in the distant region is clustered as another.

  Non-Patent Document 5 suggests that ADC changes after ischemia and changes in the amount of lactic acid signal can be measured over time to depict the ischemic center and the ischemic peripheral part that can still be treated. This is due to the fact that in the ischemic region, ADC decreases to about half of normal tissue and the lactate signal increases.

Carano RAD, Takano K, Helmer KG, Tatlisumak T, Irie K, Petruccelli JD, Fisher M, Sotak CH. Determination of focal ischemic lesion volume in the rat brain using multispectral analysis.J Magn Reson Imaging 1998; 8: 1266-1278.

Carano RAD, Li F, Irie K, Helmer KG, Silva MD, Fisher M, Sotak CH.Multispectral analysis of the temporal evolution of cerebral ischemia in the rat brain.J Magn Reson Imaging 2000; 12: 842-858. Carano RAD, Ross AL, Ross J, Williams SP, Koeppen H, Schwall RH, Van Bruggen N, Quantification of tumor tissue populations by multispectral analysis.Magn Reson Med 2004; 51: 542-551. Itskovich VV, Samber DD, Mani V, Aguinaldo JGS, Fallon JT, Tang CY, Fuster V, Fayad ZA.Quantification of human atherosclerotic plaques using spatially enhanced cluster analysis of multicontrast-weighted magnetic resonance images.Magn Reson Med 2004; 52: 515 -523. Takegami T, Ebisu T, Bito Y, Hirata S, Yamamoto Y, Tanaka C, Naruse S, Mineura K. Mismatch between lactate and the apparent diffusion coefficient of water in progressive focal ischemia.NMR in Biomed 2001; 14: 5-11.

  In the prior art, a time-series multivalue image analysis method has not been proposed. Non-Patent Document 5 suggests that ADC changes after ischemia and changes in the amount of lactic acid signal are measured over time to depict the ischemic center and the ischemic peripheral area that can still be treated. No analysis method has been reported.

  Non-Patent Documents 1, 2, 3, and 4 propose a method for discriminating a tissue from multi-value images such as nuclear spin density, T1, T2, and ADC, but no analysis is performed in the time direction. In the method shown here, only the time is stopped and the organization is determined at each time point.

  The problem to be solved by the present invention is to provide a medical image data analysis method and system for analyzing a time-series multilevel image, which is a time-series image of a plurality of measurement amounts, and supporting tissue discrimination.

  The present invention is an image data analysis method for analyzing time series images of a plurality of measurement amounts, wherein the image extracting means extracts the time series images of the plurality of measurement quantities, and the image distance setting means comprises a plurality of In each pixel of the time-series image of the measurement amount, a step of setting a distance to a vector composed of the measurement amounts at a plurality of time points obtained when each pixel is fixed, and the clustering unit clusters the pixels according to the set distance. It consists of steps.

The present invention is an image data analysis method for analyzing time series images of a plurality of measurement amounts, wherein the image extraction means extracts a time series image of the plurality of measurement amounts and a unified map for standardizing the plurality of images. And a step in which the image deformation means deforms the time-series image based on the unified map, and an image distance setting means fixes each pixel in each of the deformed time-series images of the plurality of measurement amounts. And a step of clustering pixels according to the set distance by the clustering means.
Further, the step of calculating a function having the time axis as a defined area and the plurality of measured quantities as a range when the image weighting unit fixes each pixel in each of the transformed time series images of the measured quantities. It is characterized by having. Further, the image display means includes a step of displaying a result of classifying the pixels.

  The present invention is an image data analysis system for analyzing a time series image, an image extraction means for extracting a time series image of a plurality of measurement amounts, and a unified map for standardizing the plurality of images, and a time series Image deformation means for deforming an image based on a unified map, and image distance setting for setting a distance to a vector consisting of measurement amounts at multiple points obtained when each pixel is fixed in each pixel of the deformed time-series image And clustering means for clustering pixels according to the set distance.

  In addition, each pixel of the modified time series image of the plurality of measurement amounts has an image weighting unit that calculates a function having the time axis as a domain and the plurality of measurement amounts as a range when each pixel is fixed. It is characterized by. An image data analysis system comprising image display means for displaying pixel classing results.

  According to the medical image data analysis method of the present invention, it is possible to collect a region that takes the same time-dependent change from a plurality of time-varying measurement values as a cluster, thereby reducing labor for viewing all images in detail. Is possible. It is possible to support the tissue discrimination by observing the cluster result and observing the change of the measured value with time. In addition, it is possible to reduce the possibility of missing a small area that changes with time different from others. It is also possible to quantitatively evaluate the possibility that a certain region belongs to a known tissue. That is, when a cancer property is expressed by a set of physical quantities, it is possible to evaluate how close the measured region is to the cancer property.

  Hereinafter, the best mode for carrying out the present invention will be described.

A data analysis method for a time-series multilevel image according to the present invention will be described with reference to the drawings. FIG. 1 shows a flowchart of time-series multivalue data analysis of the present invention.
First, step 11 will be described. The unified map B (x, y, z) represents a spatial standard for analysis. x, y, and z represent spatial coordinates, and B (x, y, z) is a function that maps a physical quantity for each coordinate. Such a unified map is for handling data of a plurality of measurements in a spatially uniform manner. For example, such a unified map is necessary when using a measurement method in which different spatial distortions are included in the data, or when it is desired to derive a standard tendency from different patients. In addition, such a unified map setting may be shown explicitly, such as a standard brain map, or may be shown explicitly, such as using an image to determine the measurement position in a series of measurements. There may be no.
Although the pixel is represented by (x, y, z), it may be (x, y) in a two-dimensional space.

  Next, in step 12, the time-series multivalued image S (m, tm, x, y, z) is deformed according to the set unified map B (x, y, z). Here, the time-series multivalued image is measurement data obtained by a certain measuring method using a measuring apparatus. x, y, and z represent spatial coordinates, m represents a physical quantity to be measured, and tm represents a measurement time. The time tm represents an absolute time or an elapsed time from a reference event such as the start of an ischemic attack. When the measurement time required to acquire measurement data is long, an operation such as representing the measurement start time is performed.

  Since the deformation of the time-series multi-value image differs depending on the measurement method, it is common to perform the deformation for each measurement amount m. The transformation includes (1) affine transformation to match spatial differences, (2) interpolation processing to match the number of pixel values and differences in resolution, and (3) luminance conversion to match signal amount unevenness. , Etc. are used. For example, (1) may be as follows. This is a case where B (x, y, z) is a Spine Echo (SE) image and S (1, t1, x, y, z) is an image acquired by Echo Planar Imaging (EPI). In such a case, in EPI, a spatial displacement occurs due to high-speed switching of the gradient magnetic field. In order to correct this misalignment, deformation processing is performed such as calculating the image center of gravity and performing parallel movement, tracking feature points in the image, and performing rotational movement and enlargement / reduction. For example, (2) may be as follows. This is a case where B (x, y, z) is an SE image and S (2, t2, x, y, z) is an image acquired by Spectroscopic Imaging (SI). In such a case, the resolution of the image acquired by SI is low and the number of pixels is small. For this reason, spatial interpolation processing such as linear interpolation is performed to match the resolution and the number of pixels. For example, (3) may be as follows. B (x, y, z) is an SE image using a normal volume coil, and S (3, t3, x, y, z) is an SI image measured using a local surface coil to increase SN It is. Since the surface coil has a strong sensitivity in the vicinity, luminance conversion is performed to suppress the spatial unevenness of the signal.

  In addition, when analyzing tissue changes due to anticancer agents, the time series image S (1, t1, x, y, z) is used to measure lactic acid, using the deformation amount of the time series image S (1, t1, x, y, z) for measuring the nuclear spin density. 2, t2, x, y, z) may be deformed. In this way, there may be a case where deformation cannot be performed independently for each measurement amount. Further, as in this example, even if the measurement amount is the same, the deformation amount may change with time.

  In step 13, the time-series images S (m, tm, x, y, z) after the transformation are collected, and the time-series multi-value images S (m , Tm, x, y, z). As internal processing of the analysis program, modified S (m, tm, x, y, z) is used without using a new memory for S (m, tm, x, y, z). May be processed.

  FIG. 2 is a schematic diagram of this time-series multivalued image S (m, tm, x, y, z). A time-series multilevel image is represented by a function S (m, tm, x, y, z). Here, m represents each measurement amount of the multi-valued image, and takes values m = 1, 2,. M is the number of measured quantities. tm represents a measurement time point for each measurement quantity m, and is a discrete value corresponding to the measurement quantity m. (x, y, z) represents a spatial position. The graph in the upper left of the drawing represents a space with x on the horizontal axis and y on the vertical axis, and each cell represents a pixel. Each pixel is assigned M time-series measurement values that are measurement quantities. In the drawing, a case where three types of time-series measurement values are assigned to a certain pixel (x0, y0, z0) is shown. The first row on the right shows m1 time series data. The value of S (m1, t, x0, y0, z0) is composed of measurement time points t11, t12, t13, t14, t15, t16, and t17. The second and third stages correspond to the case where two types of measurement values are obtained by the same measurement. For example, signal amounts of various metabolites can be simultaneously measured in Spectroscopic Imaging (SI) images. Such a case corresponds to a case where time series signals of lactic acid, choline, creatine, etc. are taken.

Step 14 in FIG. 1 is a weighting of (m, tm). This specifies the importance of each measured quantity. For example, when the ADC and lactic acid signal amounts are two measured amounts, it may be difficult to compare without weighting. For example, the ADC is about 0.8 × 10 ⁻⁹ m ² / s for normal brain parenchyma. The amount of signal of lactic acid is different in arbitrary units, and it is difficult to compare values with different units in the same column. On the other hand, if each measured quantity is weighted, it can be handled in an integrated manner. Of course, this step is not necessary. For example, as will be described later, it can be incorporated into the definition of distance.

Moreover, weighting can be set not only for each measurement amount m but also for the measurement time point tm. For example, when the measurement time is taken fine, the weight is low, and when the measurement time is coarse, the weight is high. As a result, when measurement is performed at unequal intervals, it is possible to appropriately weight each measurement point and improve analysis accuracy.
In step 15, the time-series multivalued image S (m, tm, x, y, z) is weighted using the (m, tm) weighting set in step 14, and the time-series multivalued image W is processed. (m, tm, x, y, z). Of course, if no weighting is performed in step 14, this step is not necessary.

  In step 16, the time series multi-value vector W (m, tm, x, y, z) when each pixel (x, y, z) is fixed is converted into a Σtm (m = 1, 2,... M) dimensional space. Set the distance as if it were a point at. FIG. 3 shows a schematic diagram of the present multidimensional space. Each arrow represents a set of the measurement amount m and the measurement time tm. The number of dimensions is Σtm (m = 1, 2,... M) at all measurement points. A vector W (m, tm, x0, y0, z0) corresponding to each pixel (x0, y0, z0) fixed can be considered as a point in this multidimensional space. Set the distance in this space. As the distance, for example, the Euclidean distance, the Manhattan distance, and the Mahalanobis distance can be used.

  In step 17, the pixel (x, y, z) is clustered using the set distance in the multidimensional space. As a clustering method, non-hierarchical clustering methods such as k-means, k-nearest neighbor, and fuzzy c-clustering, and hierarchical clustering such as nearest neighbor method, median method, and group average method are used. In addition, since the tissue properties may differ greatly, such a clustering method may be applied in several stages.

  Finally, in step 18, the clustering result of the pixel (x, y, z) is displayed. Normally, this display is performed using the unified map B (x, y, z). Of course, the clustering result may be inversely converted into each time-series image and superimposed. Further, when batch processing is performed, display is not necessarily required, and the analysis result may be saved.

  In this embodiment, a multi-dimensional space is considered with the time series of measurement points related to a plurality of measurement quantities as axes. However, it is also possible to set a distance by including space axes x, y, and z as axes. . That is, the points are clustered in a multi-dimensional space of 3 + Σtm (m = 1, 2,... M) dimensions as the number of dimensions. Thereby, for example, pixels that are spatially close can be set to be as close as possible in a multidimensional space. Note that there is a method of applying a spatial smoothing filter function to a time-series multi-valued image as a method for making it easy to put such spatially close pixels in the same cluster. It goes without saying that either method may be used.

  FIG. 4 shows a functional configuration diagram of an analysis system equipped with a medical image data analysis method using this algorithm. This analysis system consists of three main means. The first means is image data input / output means for inputting / outputting image data to / from this analysis program. This includes an image extraction means for inputting image data from an image database or an image diagnostic apparatus in which images are stored, and an image storage means for saving the analyzed image in the image database. The second means is an image display means for showing an image to the user. This includes a unified map display means for displaying a unified map, a time series multi-value image display means for displaying a time-series multi-value image as original data to be analyzed, and an analysis result display means for displaying the analysis result. The third is image analysis means. For this purpose, as described in the analysis algorithm, a time series multi-value image transforming means into a unified map, a time series multi-value image weighting means, a time series multi-value image distance setting means, a time series multi-value image, Clustering means for each pixel is included.

  First, in accordance with a user input, the image extraction means extracts a predetermined unified map and a time-series multi-value image from the image database. Here, as the unified map, one of time-series multi-value images or a calculated image such as an average may be used. At this time, the user does not directly select the unified map, but the image extracting means selects or calculates the unified map according to a predetermined method. In addition, when an image is directly captured from the image diagnostic apparatus, the image extraction unit extracts a measurement image. The image data structure is generally the DICOM format, but may be a data structure unique to the apparatus. Further, in the case of a time-series multi-value image, it may be performed with a data structure in which one image data is used, or may be handled as a set of a plurality of image data. The time information for representing the time series may be an absolute time or an elapsed time from a certain event, and the data may be included in the image data structure or held as separate data. Absent. However, if it is held as separate data, a means for reading them is required.

  The extracted unified map and time-series multi-value image are displayed on the computer screen by the unified map display means and the time-series multi-value image display means, respectively. If the unified map is not explicitly specified here, there is no need to display it.

  The unified map and the time-series multi-value image are passed to the image analysis means. Here, the user selects an analysis method and parameters according to the displayed image, and executes the analysis. The image analysis means displays a user analysis method, a dialog for prompting parameter input, and the like as necessary, and executes analysis processing according to the algorithm. Note that the analysis processing here is not limited to the above-described algorithm, and other analysis processing as described in other embodiments can be performed.

  The resulting image, time series graph, etc. are passed to the analysis result display means and displayed on the computer screen. The user performs the analysis again if necessary or saves the analysis result. When storage of the analysis result is designated, the image storage unit stores the analysis result in the image database. Here, the storage destination of the analysis result is not limited to the image database, but may be printed, stored in e-mail, or stored in a normal file system.

  FIG. 5 shows a screen configuration example of the medical image data analysis method. In this example, the screen configuration is roughly divided into three. The upper part is an interface for selecting an analysis target image and allowing the user to select an analysis method and an analysis parameter on the image analysis control panel. For example, a button for selecting a unified map and a time-series multi-valued image is arranged at the left end. Next to that, buttons and the like are arranged so that the presence / absence of a weighting process required for the analysis process and the method when the weighting process is required can be selected. A button for setting a distance between time-series multi-valued images is arranged on the right side. On the right side are buttons and columns for selecting a clustering method and inputting parameters thereof. Note that these may not be buttons but input means in a menu format or an input field format.

  The middle row is an image display panel obtained by extracting a file designated by the user via the image analysis control panel. The unified map is displayed on the left, and the time-series multivalued image is displayed on the right. In the time-series multivalued image, different measurement amounts are shown on the upper and lower sides, and the left and right show time changes. In this example, the upper part is ADC and the lower part is a signal amount of lactic acid, and the whiter the higher the respective values are. In addition, the contour line of the unified map is superimposed and displayed. In this example, labor is saved, but a setting panel for controlling the contrast of these images may be arranged. Needless to say, the binary and time directions of ADC and lactic acid need not be three.

The lower panel is a panel that displays the analysis results. On the left side, each pixel is clustered, and pixels belonging to the same cluster are displayed in the same color. The line graph on the right side shows the change over time of ADC and lactic acid in each pixel. The line color is set according to the cluster.
In addition, this example shows a typical screen display, and needless to say, it is not necessary to have such an arrangement or configuration.

  FIG. 6 shows a typical modification of the algorithm of this embodiment. The difference between FIG. 1 and FIG. Other processing steps are the same.

  Filtering in the space (x, y, z) direction and time tm direction avoids a decrease in analysis accuracy due to a low SN ratio. Depending on the amount of measurement, the signal-to-noise ratio is low, and if the data is analyzed as it is, the determination result may be greatly affected by noise. In order to suppress this, a spatially smoothing filter may be applied, or a moving average in the time direction may be calculated. This processing is performed by providing filter means in the image analysis means of FIG.

  The calculation of the statistical value in the time tm direction is used when the axis to be analyzed is replaced with a representative value to increase the analysis accuracy. In addition, when the number of measurement points is extremely different depending on the measurement amount, thinning processing for the measurement points may be performed. Further, a difference between measured values at the time of measurement may be calculated. This also has the effect of making it easier to extract changes over time.

  Furthermore, it is also possible to calculate a calculation image from images of several measurement amounts m. For example, as in the case where an ADC image is calculated from a plurality of diffusion-weighted imaging (DWI) images, fitting with an exponential decay function may be performed to combine the plurality of images into one. In this case, since the number M of measurement quantities and the number of measurement points tm change, a process such as reducing the number is necessary when calculating the number of dimensions in a later process.

  Note that the processing shown in step 69 is not necessarily performed at this position in the flow. The calculation processing in the spatial direction and the time direction that can be performed independently with the measurement amount m can be performed before deformation. In addition, image processing between a plurality of images can also be performed before transformation if each of them is already spatially consistent. In Example 4 to be described later, the order is partially changed.

  It is also possible to quantitatively evaluate how close a certain pixel is to a tissue whose change over time of a plurality of measurement quantities is already known. For this calculation, the distance set in step 16 or 66 may be used.

  By such an analysis method, there is an advantage that a plurality of time-series images that are usually difficult to grasp can be simplified. In particular, since pixels that show the same temporal change are in the same cluster, there is an effect that it is easy to determine what type of tissue it is by looking at only representative values. Even when the tissue cannot be completely discriminated only by this analysis method, the original time-series image can be observed in detail based on the discrimination result. In addition, there is an effect that it is easy to extract a tissue or the like that changes with the passage of time that is unlikely to be overlooked under the great effort of viewing all of a plurality of time-series images.

  In the second embodiment, step 14 and subsequent steps are changed in the algorithm of the first embodiment. This algorithm considers each pixel as a point in the function space from measurement time to multiple values, rather than considering each pixel as a point in the measurement time as an axis in the multidimensional space.

This algorithm will be described with reference to FIG.
The description starts from step 74 different from the first embodiment. The steps so far are the same as those of the first embodiment, particularly FIG. In step 74, first, the measured value is fitted with the function F (m, t) for each pixel (x, y, z). In the simplest case, as the function F, a step function starting from each measurement value and the measurement time point is used. However, the end time of the last measurement is, for example, the average value of the intervals at the measurement time points, or if the measurement time points are equal intervals, the interval is used. The function F may be a function specified by a line graph connecting each measurement value and the measurement time point. A function F defined by a certain parameter may be used. For example, when a model that attenuates with a certain time constant is considered to be applied to a certain measurement quantity, the measured value is parameter-fitted with an attenuation function, and the attenuation function is set to F (m, t). This fitting may be independent for each measurement amount m, or may be performed based on a plurality of measurement amounts. This choice depends on the model applied.

  Next, at step 75, weighting of the measurement amount is set in the same manner as at step 14 in FIG. However, the weighting regarding the time tm is not necessarily required here. If weighting is desired in the time direction, it can also be performed at a location where the distance of the function space is set in step 77 later.

  In step 76, the multi-value time function image F (m, t, x, y, z) is weighted according to the weighting on the measurement quantity m, and the multi-value time function image G (m, t, x, y, z) is obtained. calculate. This is because, as described in the first embodiment, a plurality of measurement amounts are integrated and handled. This weighting may be performed with repeated changes until a desired clustering result is obtained.

  In step 77, a multi-value time function G (m, t, x, y, z) when each pixel (x, y, z) is fixed, a function having time as a domain and an m-dimensional value as a range. Think of it as a point in space. In this function space, we define distances. As this distance, for example, a square integral distance or an integral distance can be used. The square integration distance is a function in which, when two functions are given, the difference is squared and integrated in the time direction to calculate the square root of the integral value, which is used as the distance between the two functions.

  In step 78, the pixel (x, y, z) is clustered using the set function space distance. As a clustering method, non-hierarchical clustering methods such as k-means, k-nearest neighbor, and fuzzy c-clustering, and hierarchical clustering such as nearest neighbor method, median method, and group average method are used. In addition, since the tissue properties may differ greatly, such a clustering method may be applied in several stages.

  Finally, in step 79, the clustering result of the pixel (x, y, z) is displayed. Normally, this display is performed using the unified map B (x, y, z). Of course, the clustering result may be inversely converted into each time-series image and superimposed.

  By such an analysis method, there is an advantage that a plurality of time-series images that are usually difficult to grasp can be simplified. In particular, since pixels that show the same temporal change are in the same cluster, there is an effect that it is easy to determine what type of tissue it is by looking at only representative values. Even when the tissue cannot be completely discriminated only by this analysis method, the original time-series image can be observed in detail based on the discrimination result. In addition, there is an effect that it is easy to extract a tissue or the like that changes with the passage of time that is unlikely to be overlooked under the great effort of viewing all of a plurality of time-series images.

  The difference from the first embodiment is that, since the fitting to the function F is used, the same handling is possible even at the measurement time of unequal intervals, and it is resistant to noise. Thereby, depending on measurement data, there exists an effect that an analysis precision can be improved rather than the method of Example 1. FIG.

  The present embodiment relates to a system equipped with the medical image data analysis method described in the first and second embodiments. FIG. 8 shows a schematic configuration diagram of this system. The upper part of FIG. 8 shows a data analysis system 81 in which the present analysis method is mounted. A data analysis system includes a CPU that performs arithmetic processing, a memory that temporarily stores programs and data for high-speed arithmetic processing, a storage device that stores programs and data, an input device that accepts input from the user, and performs arithmetic operations for the user It consists of an output device that shows the results, a medical image diagnostic device, and a communication device that communicates with other computers. A program describing the medical image data analysis method is stored in a storage device. In addition, measurement data, measurement conditions, and other data acquired from the medical image diagnostic apparatus via the communication device are stored in the storage device. In response to a start command for the program from the user, the program is stored in the memory. Based on this program, the CPU performs data calculation processing, outputs an input screen for prompting user input such as parameter setting to the output device if necessary, and accepts input from the user by the input device. The analysis result calculated by the CPU is output to the output device.

  The lower part of FIG. 8 represents the medical image diagnostic apparatus 82. Here, an MRI apparatus will be described as an example. The MRI includes a static magnetic field generating magnet 83, a gradient magnetic field coil 84, an RF coil 85, a computer 86, a synthesizer, a modulator, an amplifier, and an AD converter. The high-frequency magnetic field that excites the nuclear spin of the measurement object 87 is generated by supplying the current to the RF coil 85 by shaping the waveform of the high-frequency generated by the synthesizer with a modulator and power amplification. The gradient magnetic field generating coil 84 supplied with a current from the gradient magnetic field power source generates a gradient magnetic field and modulates a magnetic resonance signal from the measurement target 87. The modulated signal is received by the RF coil 85, amplified by an amplifier, acquired by an AD converter, and input to a computer 86. The acquired data is processed and stored in a computer. The computer performs control so that each device operates at a timing and intensity programmed in advance. The stored data is transmitted to the data analysis system via the communication device at a predetermined timing.

  In this embodiment, the data analysis system 81 and the medical image diagnostic apparatus 82 are described as separate systems. Of course, by making the data analysis system 81 and the computer 86 the same, the medical image diagnostic apparatus itself can be a data analysis system.

  In this embodiment, the MRI is used for explanation, but it goes without saying that the present invention can also be applied to other medical diagnostic apparatuses such as X-ray CT, PET, and ultrasonic diagnosis. It is also possible to connect the data analysis system to a plurality of medical image diagnostic apparatuses so that data acquired by the plurality of apparatuses can be analyzed. In this case, the measurement amounts of Examples 1 and 2 are obtained by a plurality of apparatuses.

In the present embodiment, the present invention will be described while showing an MRI measurement protocol, a time-series multi-value image acquired, an analysis method and an analysis result thereof.
FIG. 9 shows this measurement protocol. The upper graph shows the flow of measurement, with time [minutes] on the horizontal axis. From the start of measurement, measurement preparation such as positioning is first performed, and an SE image is measured. Next, after 30 minutes, ADC measurement and SI measurement are alternately repeated three times. The imaging sequence for ADC measurement and the imaging sequence for SI measurement are shown below. In the sequence, the horizontal axis represents time, and the vertical axis represents the strength of each RF, slice gradient magnetic field Gs, readout gradient magnetic field Gr, and encode gradient magnetic field Ge. Here, in ADC measurement, a plurality of DWI measurements are performed while changing the diffusion gradient magnetic field shown in steps in the middle Gs, Gr, and Ge. The ADC image is calculated by performing exponential fitting from a plurality of DWI images. In SI measurement, data is repeatedly measured while changing the stepwise encoding gradient magnetic field of Ge.

  The MRI apparatus shown in the lower part of FIG. 8 operates at the timing and intensity described in this sequence, and acquires measurement data. According to this measurement protocol, an ADC image and a plurality of metabolite images obtained by SI are obtained as time-series images at three time points. Hereinafter, for simplicity, an image obtained by SI measurement will be described as one type of lactic acid image. That is, the time-series multivalued image obtained by this measurement protocol takes two values, ADC and lactic acid signal, and has three measurement points in the time direction.

  In this example, the time course of ADC and lactic acid after cerebral ischemia was measured using this protocol.

  FIG. 10 shows an analysis flowchart to which the present invention is applied. For comparison, description will be made with reference to the steps of FIG.

Step 101 corresponds to the unified map setting in step 61. The definition area of the unified map is set to a signal area that is 10% or more of the maximum value of the SE image.
Step 102 corresponds to step 69 in which one measurement amount image is calculated from a plurality of measurement amount images.

  Steps 103, 104, and 105 correspond to the transformation step 68 into the unified map. Since SI images have low pixel values and low spatial resolution, they are matched to the unified map using zero filling in the frequency space.

  Step 106 corresponds to steps 63 and 66. This time, it is a binary multi-value image of ADC and lactic acid, and each measurement is made at 3 time points, so it becomes a time-series multi-valued image taking 6-dimensional values.

  Step 107 corresponds to steps 64 and 65. ADC was weighted so that it was normalized by the normal value of the brain parenchyma and lactic acid by the maximum value of the measurement time. Step 108 corresponds to the clustering of step 47. Here, the Euclidean distance was used as the distance, and the k-means method was used as the clustering method. As the number k of clustering, 5 which is an assumed type of organization was used. Of course, the Mahalanobis distance or the like can be used as the distance, and the number of clusters can be set while determining the clusters.

  FIG. 11 shows the analysis result. The top row shows the difference between clusters in the unified map. The legend on the right contains the user's interpretation of the data that saw it. The graph below shows the time transition of ADC on the left and the time transition of lactic acid on the right. Each line graph represents one pixel. The user's interpretation in the legend was entered by observing this graph. Here, blood flows from an area considered normal brain parenchyma, an area considered ventricular and cerebrospinal fluid, an area considered to be an ischemic center, an area considered to be an ischemic peripheral area, and an area that is not ischemic. There is a possibility that it can be divided into areas where the coming reperfusion is considered to occur.

  Thus, according to the present analysis method, there is an effect that the ischemic region can be discriminated in detail. In particular, it is difficult to extract a reperfusion region or a region considered to be an ischemic peripheral portion only by looking at a plurality of time-series images. It is considered to be a tissue discrimination method that is easy for the first time by this measurement protocol and the time-series multi-value image analysis of the present invention.

  It is possible to analyze medical image data and reduce the diagnostic effort.

The algorithm of the medical image data analysis method of this invention. The schematic diagram showing a time-sequential multi-value image. The schematic diagram which considered each pixel of a time-sequential multi-value image as a point in multidimensional space. The functional block diagram of the analysis system carrying a medical image data analysis method. Screen display example of an analysis system equipped with a medical image data analysis method. Another algorithm of the medical image data analysis method of the present invention. Another algorithm of the medical image data analysis method of the present invention. An example of the system carrying the medical image data analysis method of this invention. An example of a measurement protocol for time-series multivalued images. Analysis algorithm for time-series multivalued images measured with this measurement protocol. Analysis results of time-series multivalued images measured with this measurement protocol.

Explanation of symbols

81: Data analysis system, 82: Medical diagnostic device, 86: Computer.

Claims (8)

In the image data analysis method for analyzing time series images of multiple measurement quantities,
An image extracting means for extracting time series images of a plurality of measurement amounts;
An image distance setting means, in each pixel of the time series image of the plurality of measurement amounts, setting a distance to a vector consisting of measurement amounts at a plurality of time points obtained when each pixel is fixed;
A clustering means comprises a step of clustering the pixels according to the set distance. In the image data analysis method for analyzing time series images of multiple measurement quantities,
A step of extracting a time-series image of a plurality of measurement amounts and a unified map for standardizing the plurality of images;
An image transformation means for transforming the time-series image based on the unified map;
An image distance setting means, in each pixel of the transformed time series image of a plurality of measurement quantities, setting a distance to a vector consisting of measurement quantities at a plurality of time points obtained when each pixel is fixed;
A clustering means comprises a step of clustering the pixels according to the set distance.   2. The image data analysis method according to claim 1, wherein when the image weighting unit fixes each pixel in each of the transformed time-series images of the plurality of measurement amounts, a plurality of time axes are defined in the domain. And a step of calculating a function having a measured amount as a range.   4. The image data analysis method according to claim 1, wherein the image display means includes a step of displaying a result of classifying the pixels.   5. The image data analysis method according to claim 1, wherein the filter means performs any one or more of spatial smoothing of the time-series image or calculation of moving average in the time direction. Image data analysis method. An image data analysis system for analyzing time series images,
Image extracting means for extracting time series images of a plurality of measurement amounts and a unified map for standardizing the plurality of images;
Image deformation means for deforming the time-series image based on the unified map;
Image distance setting means for setting a distance to a vector composed of measurement amounts at a plurality of time points obtained when each pixel is fixed in each pixel of the transformed time-series image;
Clustering means for clustering the pixels according to the set distance, an image data analysis system.   7. The image data analysis system according to claim 6, wherein when each pixel is fixed in each pixel of the transformed time series image of the plurality of measurement amounts, the time axis is defined as a range and the plurality of measurement amounts are defined as value ranges. An image data analysis system comprising image weighting means for calculating a function to perform. 8. The image data analysis system according to claim 6, further comprising image display means for displaying a result of classifying the pixels.

Images