JP2006235971A - Image referencing apparatus and method for classifying and searching for image - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image referencing apparatus and a method for classifying and searching for images making it possible to perform classification and searches of high objectivity and compatibility in such a way as to reduce human workload and facilitate access to desired images. <P>SOLUTION: SOM is repeatedly performed wherein a plurality of DICOM original images are used as input values and self-mapping of an anatomical structure as the phase characteristic of each DICOM original image is carried out through predetermined weighting. In this way, the DICOM original images are classified using an arrangement having such a characteristic as arranging the images in such a manner that the more closely the phase structures of anatomical structure information resemble each other, the more closely they are positioned to one another. Also, a search space is organized using a digest image that corresponds to each output unit available through SOM, and searches for the DICOM original images using the digest image are conducted. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an image reference device (viewer) and an image classification search method that are incorporated in a medical imaging device represented by, for example, a magnetic resonance imaging device, or realized by a medical workstation or the like.

  In recent years, with the advancement of medical imaging equipment, the amount of information has increased dramatically in terms of the quality and quantity of medical images, and it is possible to acquire several tens to hundreds of slice images in one shot. it can. In addition, various types of image diagnosis have been performed with the introduction of sophisticated medical image equipment. For example, recent cerebral docks detect asymptomatic cerebral infarction and cerebral white matter lesions at an early stage, and promote lifestyle changes that are closely related to those cerebral diseases, so that subsequent symptomatic cerebral infarction One of the purposes is to prevent the occurrence. The majority of asymptomatic cerebral infarction is called lacunar infarction, which is mainly a small infarct under 1 cm formed in the penetrating region, about 3-10 mm in diameter, irregular, heterogeneous, low signal area in T1-weighted horizontal section, T2 It is defined as a high signal area in the emphasized horizontal section. Such lacunar infarction is considered to be deeply related to the occurrence of stroke in the hypertensive group. If lacunar infarction is observed on MR images and if it has hypertension, the risk of future stroke Since there is a report that the sex is about 10 times higher, it is considered that there is a possibility that the occurrence of serious stroke that can occur afterwards can be prevented by appropriate blood pressure management and lifestyle improvement guidance.

  MR images of the brain used for brain dogs and the like are the most frequently screened parts. Brain screening must begin with a careful search for lesions. In practice, each step shown in FIG. 13 is almost unconsciously repeated, and several differential diagnoses are considered, and interpretation is performed while asking himself / herself whether there is a finding suitable for it. In particular, in order to find minute lesions such as lacunar infarction, it is important to compare the current image with images taken in the past and examine changes over time. Is required.

  As described above, in recent image diagnosis, the burden on a doctor who is required to interpret an image while selecting an image of interest in accordance with the purpose of diagnosis from the captured image is enormous. In daily medical care, it is important to convey the judgment regarding the similarity of the target image to the doctor in an easy-to-understand format in order to reduce the burden of the doctor's interpretation work and promote efficient interpretation work. In general, in a conventional image management system, two approaches of search and classification are adopted. There are two methods for searching for an image: a method for searching by text information (keywords and explanations) added to the image, and a method for searching by the similarity of image content information (features such as colors and textures). Similarity search using image content information can be searched by features such as the color of the entire image, the pattern, the color, shape and position of the object in the image, and the length, inclination, and position of the straight line in the image. is there. On the other hand, image classification includes a method of performing classification using text information added to an image and a method of performing classification using image content information. There is a similar scene classification as a classification using image content information, and clustering is possible according to the color of the entire image or the block division area of the image and the feature amount of the texture.

  However, the conventional image management system has the following problems, for example.

  For example, whether to use text information or image content information in an image search is arbitrarily selected by each hospital or doctor. Therefore, the criteria used for image classification and retrieval in the conventional system vary among hospitals and doctors, and the objectivity and compatibility are poor.

  Further, the conventional classification executed using text information or image content information has a low degree of freedom, and it is difficult to manage an enormous number of images. For example, in classification / search using text information, images with the same header information are classified / searched as the same image even if there is an anatomical structural difference. Accordingly, it may be difficult to access a desired diagnostic image, and if a plurality of diagnostic images retrieved simultaneously are used artificially, a great deal of labor is required.

In addition, as a well-known document relevant to this application, there exist the following, for example.
JP 2004-174217 A JP 2004-174218 A JP 2004-174219 A JP 2004-174220 A

  The present invention has been made in view of the above circumstances, and can perform classification / retrieval with high objectivity and compatibility, and classification / retrieval that can easily access a desired image with less human work load. An object is to provide an image reference device and an image classification search method.

  In order to achieve the above object, the present invention takes the following measures.

  According to a first aspect of the present invention, storage means for storing a plurality of first images relating to a predetermined part of a subject and the plurality of first images as inputs, and an anatomy of the predetermined part by a predetermined weighting factor. By executing a self-mapping with a target structure as a phase characteristic of each of the first images, and outputting a plurality of output units each composed of the classified first images. A second image corresponding to each output unit is generated by imaging a self-mapping unit for classifying the first image and the predetermined weighting factor used for the self-mapping, and the second image Image generating means for calculating a correlation coefficient indicating the similarity between images, dividing means for dividing the plurality of output units obtained into at least two groups based on the calculated correlation coefficient; Self-mapping A determination means for determining whether or not to repeatedly execute the self-mapping based on the calculated correlation coefficient and the result of the division, and the determination means repeatedly executes the self-mapping. When it is determined, the self-mapping, the generation of the second image, the calculation of the correlation coefficient, the division, and the determination are repeated.

  According to a second aspect of the present invention, a plurality of first images relating to a predetermined part of a subject and the plurality of first images are input, and an anatomical structure of the predetermined part is determined by a predetermined weighting factor. As the phase characteristic of the first image, the predetermined weighting factor used for the self-mapping that outputs a plurality of output units each composed of the classified first images is imaged, so that the self-mapping A connection information table associating each second image of the plurality of output units obtained by the above, the second image corresponding to each output unit, and the first image constituting each output unit; Storage means for storing, selection means for selecting one of the second images corresponding to each output unit, and selection from the storage means based on the connection information table An image reference device comprising: search means for searching for the first image constituting the output unit corresponding to the second image that has been searched; and output means for outputting the searched first image. is there.

  According to a third aspect of the present invention, a plurality of first images relating to a predetermined part of the subject are input, and an anatomical structure of the predetermined part is set as a phase characteristic of each first image by a predetermined weighting factor. Performing a self-mapping to output a plurality of output units each composed of the classified first images, thereby classifying the plurality of input first images; A second image corresponding to each output unit is generated by imaging the predetermined weight coefficient used for self-mapping, and a correlation coefficient indicating a similarity between the second images is calculated. A second step, a third step of dividing the obtained plurality of output units into at least two groups based on the calculated correlation coefficient; and the phase calculated as a result of the self-mapping. Number of relationships, previous A fourth step of determining whether or not to repeatedly execute the self-mapping based on a result of the division; and if it is determined to repeatedly execute the self-mapping in the fourth step, the self-mapping, And a fifth step of repeating the generation of the second image, the calculation of the correlation coefficient, the division, and the determination.

  As described above, according to the present invention, an image reference device and an image classification search capable of executing classification / retrieval with high objectivity and compatibility, and classification / retrieval that can easily access a desired image with less human work load. A method can be realized.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, components having substantially the same function and configuration are denoted by the same reference numerals, and redundant description will be given only when necessary.

  FIG. 1 is a block diagram of an image reference device 1 according to this embodiment. As shown in the figure, the image reference device 1 includes a digest image storage unit 2, a DICOM original image storage unit 3, an operation unit 4, a display unit 5, a transmission / reception unit 6, a learning / classification unit 7, and a search space organization unit 8. The control unit 9 is provided.

  The digest image storage unit 2 stores the digest image generated by the learning / classification unit 7 and a label image generated by linearly combining a plurality of digest images. The generation of the digest image and the image management using the digest image and the label image will be described in detail later.

  The DICOM original image storage unit 3 stores an original image (DICOM original image) acquired by a medical image device in accordance with the DICOM standard and used as input information for processing in the image reference device 1. In the present embodiment, for the sake of specific explanation, the DICOM original image stored in the DICOM original image storage unit 3 is a plurality of brain images for each subject (patient) acquired by the magnetic resonance imaging apparatus. Suppose there is.

  The operation unit 4 provides various instructions from the operator, conditions, a region of interest (ROI) setting instruction, a DICOM original image classification and digest image generation described later, instructions for organizing a search space using the digest image, and the like. A trackball, various switches, a mouse, a keyboard, and the like for incorporation into the device 1 are included.

  The display unit 5 displays a DICOM original image, a digest image, a label image, and the like in a predetermined form. The display unit 5 displays a connection information table (described later) indicating the relationship between the DICOM original image and the digest image in a predetermined form.

  The transmission / reception unit 6 transmits / receives information including image data to / from other apparatuses via the network.

  The learning / classification unit 7 uses a plurality of DICOM original images stored in the DICOM original image storage unit 3 as input values, and uses a predetermined weight Wij as an anatomical structure as a phase characteristic of each DICOM original image. (Hereinafter also referred to as “SOM”). Thereby, the inputted plurality of DICOM original images are classified based on the anatomical structure.

  The search space organizing unit 8 generates a digest image using the weighting coefficient in each SOM executed in the learning / classifying unit 7. In addition, the search space organizing unit 8 organizes the search space in which the DICOM original images are arranged using the digest image.

  The control unit 9 dynamically or statically controls each unit constituting the image reference measure. In particular, the control unit 9 performs the digest image storage unit 2, the DICOM original image storage unit 3, the operation unit 4, the display unit 5, the learning unit in the DICOM original image classification / digest image generation and search image organization (both described later). The classification unit 7 and the search space organization unit 8 are controlled in an integrated manner.

  FIG. 2 is a block diagram for explaining the configuration of the learning / classifying unit 7 and the search space organizing unit 8. As shown in the figure, the learning / classification unit 7 includes a self-learning unit 11, a partial problem dividing unit 13, a partial problem learning unit 15, and a relearning determination unit 17. The search space organization unit 8 includes a search space management unit 19, a digest image generation unit 21, a connection information table generation unit 23, an SOM management label generation unit 25, and an image correspondence unit 27.

  The self-learning unit 11 learns the phase structure of a predetermined part of the patient by executing SOM using a plurality of MR images as input values and mapping them to a plurality of output units. The SOM executed by the self-learning unit 11 will be described in detail later.

  The subproblem dividing unit 13 divides the mapping result (a plurality of output units) obtained by SOM by the self-learning unit 11 and the partial problem learning unit 15 into two subproblems. Here, the partial problem means a target of SOM that is further expanded hierarchically. This division is performed by comparing correlation coefficients indicating the similarity between digest images acquired for each SOM according to a predetermined rule.

  The subproblem learning unit 15 performs SOM using the DICOM original images belonging to the subproblems divided by the subproblem dividing unit 13 as input values, and maps them to a plurality of output units. Relearn. This relearning is repeatedly executed according to the determination result of the relearning determination unit 17.

  The relearning determination unit 17 compares the correlation coefficients between digest images obtained by the SOM in the partial problem learning unit 15 and determines whether or not to perform relearning using the SOM.

  The search space management unit 19 manages the organized search space in response to an instruction input from the operation unit 4. More specifically, management such as reading a DICOM original image (or digest image) associated with a digest image (or DICOM original image) selected by an operation from the operation unit 4 based on the connection information table is executed. To do.

  The digest image generation unit 21 generates, for each output unit, a digest image obtained by imaging a weighting coefficient obtained through SOM. Moreover, the digest image generation part 21 calculates the correlation coefficient between the produced | generated digest images.

  The connection information table creation unit 23 associates the digest image corresponding to each unit with a plurality of DICOM original images classified into each unit (which constitute each unit), and the header of the associated DICOM original image Based on the information, a connection information table indicating which digest image is connected to which DICOM original image is created.

  The SOM management label generation unit 25 generates a label image by linearly combining the digest images.

  The image corresponding unit 27 writes a digest image corresponding to the private tag of each DICOM original image. The written digest image is used as a SOM management label in search processing described later.

  In FIG. 1, the digest image storage unit 2 and the DICOM original image storage unit 3 have different configurations. However, the configuration may be realized by a single hardware without being limited thereto. . Similarly, in FIG. 1, the self-learning unit 11 and the sub-problem learning unit 15 are separate components for easy understanding, but the single learning unit is not limited to this. May be.

(Self-Organizing Map)
Next, SOM will be described. The SOM preserves the phase of input data and performs topological mapping, and does not require an explicit teacher in the learning process.

  As shown in FIG. 3, a general SOM is composed of two layers: an input layer including an input unit and a mapping layer including an output unit. A typical SOM learning algorithm is as follows.

(1) Let w _ij (1 ≦ i ≦ n) be a weighting factor from input unit i to output unit j at time t. The unit weighting factor is initialized with a random number, and the initial range near node j is set large.

(2) Let x _i (1 ≦ i ≦ n) be an input to node i at time t.

(3) the Euclidean distance d _j between the input data and the output node j is calculated by the following formula (1).

(4) the Euclidean distance d _j searches the output unit to be minimized.

(5) The weighting factor for the unit included in the vicinity defined by N _c (t) is updated by the following equation (2).
w _ij (t + 1) = w _ij (t) + α (t) (x _i (t) −w _ij (t)) (2)
Here, α (t) is the learning rate coefficient (0 <α <1), and N _c (t) is the size of the neighboring region, and decreases with time.

(6) The processes (2) to (5) are repeated.

  Specifically, the image reference device 1 according to the present embodiment executes the following processing using the SOM.

  (A) The self-learning unit 11 outputs a plurality of DICOM original images (for example, m images) as input values and outputs an anatomical structure as a phase characteristic of each DICOM original image with a predetermined weight Wij as shown in FIG. Self-mapping using a one-dimensional SOM with 5 units.

  (B) The digest image generation unit 21 generates a digest image corresponding to each output unit by imaging the weighting factor used in the SOM in (a). Moreover, the digest image generation part 21 calculates the correlation coefficient showing the similarity between adjacent digest images.

  (C) The sub-problem dividing unit 13 compares the correlation coefficients between adjacent digest images, and divides the mapping result in (a) into two sub-problems at the boundary of the unit having the smallest correlation coefficient ( That is, it is divided into two groups A1 and A2.

  (D) As shown in FIG. 4, the sub-problem learning unit 15 takes a plurality of DICOM original images (for example, n images) belonging to the group A1 divided by the sub-problem dividing unit 13 as input values, and uses a predetermined weight Wij. The anatomical structure is self-mapped as a phase characteristic of each DICOM original image using a one-dimensional SOM having five output units. The same processing is performed for a plurality of DICOM original images (mn) belonging to the group A2.

  (E) The digest image generation unit 21 generates a digest image corresponding to each output unit by imaging the weighting factor used in the SOM in (d) above. Moreover, the digest image generation part 21 calculates the correlation coefficient showing the similarity between the produced | generated digest images.

  (F) Whether the relearning determination unit 17 compares the correlation coefficients between the digest images corresponding to the respective output units obtained by the SOM for the group A1, and further performs the relearning using the hierarchical SOM. Judge whether or not. In this determination, the following relearning rule focusing on the correlation coefficient between map layer units (between digest images) is adopted. Similar processing is executed for the group A2.

<Rule 1> The unit is divided 1: 4 at the lowest boundary of the correlation coefficient (see FIG. 5A).

• 1 unit: Considered independent and does not relearn.

-4 units: Re-learning is performed regardless of the correlation coefficient.

<Rule 2> The unit is divided into 2: 3 at the lowest boundary of the correlation coefficient (see FIG. 5B).

・ When the learning is divided into 2: 3 in the first learning, re-learning is always performed.

• 2 units: If the correlation coefficient between two units is “0.95” or more, 2 units are regarded as independent and relearning is not performed.

If the correlation coefficient between the two units is “0.95” or less, relearning is performed.

・ 3 units: If all correlation coefficients are “0.95” or more, 3 units are regarded as independent and relearning is not performed.

-If there is "0.95" or less among all correlation coefficients, relearning is performed.

<Rule 3>
・ Decide “end” in interactive interaction with the operator.

-Re-learn as a partial problem until a unit that does not fire appears.

  (G) When it is determined by the relearning determination unit 17 that further hierarchical relearning is to be performed, the processes (c) to (f) are repeated. On the other hand, if the re-learning determination unit 17 determines that the hierarchical re-learning is not performed, the DICOM original image classification / digest image generation process is terminated.

  The plurality of DICOM original images finally acquired by the series of processes (a) to (g) are finally classified by a predetermined arrangement as shown in FIG. The sequence of the DICOM original images obtained as a result of this is “arranged closer to each other as the phase structure of the anatomical structure information is similar”, “the phase structure of the anatomical structure information is dissimilar. It has two characteristics, “they are arranged far away from each other.” Since the SOM that self-maps the anatomical phase structure is used as the standard for this arrangement, it can be said that the standard follows an extremely objective standard.

  In this way, the SOM executed by the image reference device 1 is repeatedly executed according to the determination result. As shown in FIG. 6, the process developed by repeating SOM can be regarded as a hierarchical development. Hereinafter, the hierarchical expansion by SOM repetition is assumed to be “higher” as the number of repetitions decreases, and “lower” as the number of repetitions increases.

  The digest image corresponding to each output unit generated using the weighting coefficient in each SOM reflects the contents of each DICOM original image constituting each output unit as some information, and therefore belongs to the output unit. It has the characteristic of representing a plurality of DICOM original images. Therefore, the operator can indirectly acquire the information of each DICOM original image by observing the digest image without observing all the DICOM original images of each output unit. it can. By utilizing the characteristics of the digest image, the present image reference device 1 is designed to facilitate the retrieval of the DICOM original image. This will be described in detail later.

  In the SOM executed in the above (a) and (d), the number of output units obtained as a mapping result is five. This is because the image reference device 1 according to the present embodiment employs a one-dimensional SOM that considers up to the second vicinity (for example, in FIG. 4, the output units 1 and 5 are viewed from the output unit 3). Second neighborhood). However, it is possible to adopt a configuration in which a one-dimensional SOM (where n is a natural number other than 2) that considers up to the maximum n-th neighborhood without being bound by this.

  In the above (c), the sub-problem dividing unit 13 divides the five output units into two sub-problems according to the correlation coefficient between the digest images. This employs a binary tree which is the most basic example in clustering. However, without being bound by this, it may be further divided into a plurality of partial problems using the correlation coefficient between digest images.

  In (f) above, it is determined whether or not to perform relearning based on the correlation coefficient value between digest images being “0.95”. This value “0.95” is in accordance with conventional usage. Accordingly, a configuration may be adopted in which the re-learning determination is performed using another reference value without being limited to this example.

(Dicom original image classification and digest image generation)
Next, classification of a DICOM original image and generation of a digest image realized using SOM will be described.

  FIG. 7 is a flowchart showing the flow of each process executed in the DICOM original image classification / digest image generation by the image reference apparatus 1. As shown in the figure, first, in response to a “self-learning / classification start” operation from the operation unit 4, a plurality of DICOM original images acquired by the magnetic resonance imaging apparatus are passed through the search space management unit 19. Are input to the self-learning unit 11 (step S1).

  Next, the self-learning unit 11 performs self-mapping using a one-dimensional SOM with five map layer units (output units) (step S2). The digest image generation unit 21 images the SOM weighting coefficient to generate a digest image corresponding to each output unit (step S3), and calculates a correlation coefficient between the generated digest images (step S4).

  Next, the sub-problem dividing unit 13 compares the correlation coefficients between the adjacent digest images, and divides the mapping result in step S2 into two sub-problems at the boundary of the unit having the smallest correlation coefficient (two Dividing into groups A1 and A2, step S5). Further, the sub-problem dividing unit 13 determines whether or not the executed SOM is the first time (step S6). If the sub-problem dividing unit 13 determines that the SOM is the first time, the sub-problem learning unit 15 uses, for each divided group, a map layer unit (with a DICOM original image constituting each input value) The output unit) performs self-mapping using 5 units of one-dimensional SOM (step S2), and repeats the processes of steps S3 to S5 (steps S3 to S5).

  On the other hand, when it is determined that the SOM executed by the partial problem dividing unit 13 is not the first time (that is, when the SOM is executed by the partial problem learning unit 15), the re-learning determining unit 17 It is determined whether or not to end the repetition (step S7). If it is determined that the SOM repetition is to be terminated, the DICOM original image classification / digest image generation is terminated. When it is determined that the repetition of SOM is not completed, the relearning determination unit 17 compares the correlation coefficients between the digest images corresponding to the output units according to the rule (f), and further hierarchically It is determined whether or not to perform re-learning using a simple SOM (steps S8, S9, and S10), re-learning is executed in the partial problem learning unit 15 (step S2), and the processes of steps S3 to S5 are repeated. (Steps S3 to S5).

  Next, when it is determined that there is an output unit that does not fire, the relearning determination unit 17 determines the number of units and the digest image for the target problem (Step S13), and performs DICOM original image classification / digest image generation processing. finish.

(Organization of search space using digest images)
Next, organization of the search space using the generated digest image will be described. Here, the search space is conceptually an abstract space in which a plurality of DICOM original images exist, and various types of information such as examination ID, series ID, imaging region, enhancement type, imaging direction, slice number, etc. are coordinate axes. It is what. On the other hand, the search space actually means a memory area of the DICOM original image storage unit 3 where a plurality of DICOM original images with various tag information are present. The search space organization using the digest image is stored in the DICOM original image storage unit 3 by organizing (systemizing) a plurality of DICOM original images using the digest image corresponding to each SOM. The DICOM original image can be easily retrieved.

  FIG. 8 is a flowchart showing a flow of processing executed by the image reference device 1 in organizing the search space. As shown in the figure, the image correspondence unit 27 associates a digest image corresponding to each output unit obtained by each SOM with a plurality of DICOM original images classified into each unit (step) A). For this association, the image correspondence unit 27 attaches a digest image corresponding to the classified unit to the private tag in the header information of each DICOM original image stored in the DICOM original image storage unit 3, for example. Executed. Therefore, when the learning / classification unit 7 executes up to k layers of SOM, k DIGITAL images are added to each DICOM original image.

  Next, the connection information table creation unit 23 creates a connection information table based on the header information of the DICOM original image associated by the image correspondence unit 27 (step B).

  FIG. 9 is a diagram for explaining an example of a connection table created by the connection information table creation unit 23. For example, it is assumed that an MR image having the characteristics described in the first stage of the original image information table 30 is input as a DICOM original image, and the above-described DICOM original image classification and digest image generation processing is executed. In such a case, the connection information table creation unit 23, for example, a connection information table 31 (digest image) that correlates a digest image and a DICOM original image with each other based on the header information of the DICOM original image associated with the image correspondence unit 27. And a connection information table 32 (a table managed with a DICOM original image as a label).

  The connection information table 31 and the connection information table 32 are examples, and the actual digest image and the like are pasted to the connection information table actually created by the connection information table creating unit 23 as shown in FIG. There is no need. Therefore, any form may be used as long as the digest image and the DICOM original image are associated with each other. For example, the connection information table 31 shown in FIG. 9 and the connection information table 33 shown in FIG. It may be configured to create.

  Next, the SOM management label generation unit 25 linearly combines the five digest images generated in each SOM (weighted addition process), and generates a label image for managing and identifying the digest image (step C). This label image is automatically stored in the digest image storage unit 2, for example.

  The SOM executed by the image reference apparatus 1 self-maps anatomical structure information as phase characteristics of each DICOM original image. Therefore, as the SOM process progresses (that is, as the SOM hierarchy progresses), the input DICOM original image is naturally classified according to the shooting direction, image enhancement type, part, and the like. Therefore, the label image obtained by linear combination of the digest images corresponding to each SOM symbolizes the image attributes such as the photographing direction, the image enhancement type, the region, etc., and which DICOM original image includes which attribute. Since it is a symbol of whether or not it is input, it is possible to manage the digest image more easily by using the label image.

  FIG. 11 is a conceptual diagram for explaining a result obtained by the search space organization process using the digest image. As shown in the figure, the DICOM original image and the digest image are associated by the connection information table, and the label image and the digest image are associated by linear combination. The operator uses the label image composed of the digest image targeted for the head (or the label image composed of the T2-weighted image and the digest image targeted, the label image composed of the digest image targeted for the pre-payment). The DIOCM original image associated with the digest image that constitutes the label image is for the head (or for the T2-weighted image and for the forehead). Judgment can be made immediately.

  The access to the search space organized by this process may take any form. For example, the operator first selects a desired attribute related to the image (for example, “T2-weighted image related to the pre-head cut”) by the label image, and is organized in an interface as shown in FIG. The structure which presents search space may be sufficient. In the figure, when one of a plurality of digest images displayed on the upper right of the screen is selected, a thumbnail image of the DICOM original image associated with the digest image is displayed on the lower right of the screen by the connection information table. By selecting a desired DICOM original image among the thumbnail images, the DICOM original image can be displayed as a full image (that is, an image that has not been compressed or thinned out) on the left side of the screen. In this way, by using thumbnail images as necessary, it is possible to more quickly select a diagnostic image from a large number of DICOM original images corresponding to a digest image.

  According to the configuration described above, the following effects can be obtained.

  According to this image reference apparatus, SOM for self-mapping an anatomical structure as a phase characteristic of each DICOM original image with a predetermined weight Wij is repeatedly executed with a plurality of DICOM original images as input values. As a result, the array of the plurality of DICOM original images finally obtained is arranged so that the phase structure of the anatomical structure information is similar, and “the phase structure of the anatomical structure information is the same. The dissimilarity is placed farther away. ” Accordingly, it is possible to automatically classify a large amount of unsorted and unclassified DICOM original images according to an objective standard, thereby reducing an artificial burden. In addition, since a new image arrangement (image classification) based on the anatomical phase structure can be provided, there is no variation compared to the artificial arrangement (classification), and compatibility is high. As a result, it is possible to contribute to improving the quality of medical practice.

  In particular, an array of a plurality of DICOM original images finally obtained is arranged so that the topological structure of the anatomical structure information is similar. Therefore, the operator can determine the quantitative nature of the similarity determined by the system and the similarity (order) between the search results.

  In addition, the present image reference device organizes a search space using digest images corresponding to each output unit, and enables retrieval of DICOM original images using the digest images. Therefore, the search target can be visually confirmed using the digest image as an index, and the desired DICOM original image can be accessed appropriately and quickly while checking the progress of the search. As a result, it is possible to reduce the work burden on the interpreting doctor and the like, and to prevent a mistake in selecting a diagnostic image, and to contribute to improving the quality of medical practice.

  Further, according to the present image reference device, a digest image can be set as a search target. Therefore, the number of direct accesses to the DICOM original image that is personal information can be reduced, and the safety of privacy can be ensured.

  In recent years, with the advancement of medical imaging equipment, the amount of information has increased dramatically in terms of the quality and quantity of medical images, and several tens to hundreds of slice images can be acquired in one shot. be able to. On the other hand, there is an increasing need to interpret images while selecting images of interest according to the purpose of diagnosis from those images, and to compare and interpret current images and past images, and the burden on doctors is enormous. I'm following. For example, it is important to examine changes over time in order to find minute lesions such as early cancer. In such a current situation, by performing thumbnail display using the digest image, it is possible to efficiently display the target image for comparative interpretation, thereby reducing the burden on the doctor. The digest image generation method that forms the basis of the invention has the possibility of automatically classifying and organizing unclassified and unorganized multimedia DBs, and therefore has the possibility of becoming a basis for multimedia data retrieval.

  Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. Specific examples of modifications are as follows.

  (1) Each function according to the present embodiment can also be realized by installing a program for executing the processing in a computer such as a workstation and developing the program on a memory. Programs that can cause the computer to execute the method are stored and distributed in a recording medium such as a magnetic disk (floppy (registered trademark) disk, hard disk, etc.), optical disk (CD-ROM, DVD, etc.), semiconductor memory, etc. Is also possible. At this time, at least one of the configuration of the learning / classifying unit 7 and the search space organizing unit 8 shown in FIG. 2 is constructed in the memory area of the medical workstation by the developed program.

  (2) In the above-described embodiment, an example has been described in which an original image using DICOM is used as a standard for images and communication, and header information (accompanying information) is used. However, the DICOM standard is not essential, and if it is another standard, the same effect can be obtained by using the incidental information regarding the image attached according to the standard.

  (3) In the above-described embodiment, the description has been given using a plurality of brain images (MR images) of a predetermined patient as an image to be classified and searched for specific description. However, the present invention is not limited to this, and is standardized from DICOM original images collected by other modalities such as an X-ray CT apparatus, DICOM original images in which images collected by a plurality of modalities are mixed, and anatomical viewpoints. Alternatively, a schematic image, an animation image, or the like may be used as an image to be classified and searched. Moreover, the structure which makes a classification / search object the thing from which a different imaging | photography site | part, emphasis kind, and imaging | photography direction coexist may be sufficient.

  In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

  As described above, according to the present invention, an image reference device and an image classification search capable of executing classification / retrieval with high objectivity and compatibility, and classification / retrieval that can easily access a desired image with less human work load. A method can be realized.

FIG. 1 is a block diagram for explaining the configuration of an image reference device 1 according to this embodiment. FIG. 2 is a block diagram for explaining the configuration of the learning / classifying unit 7 and the search space organizing unit 8. FIG. 3 is a diagram for explaining a general SOM. FIG. 4 is a diagram for explaining the SOM executed by the image reference apparatus 1. FIG. 5 is a conceptual diagram for explaining rules in the determination executed by the relearning determination unit 17. FIG. 6 is a diagram illustrating an example of a final result obtained by processing using the SOM of the image reference device 1 according to the present embodiment. FIG. 7 is a flowchart showing the flow of each process executed in the DICOM original image classification / digest image generation by the image reference apparatus 1. FIG. 8 is a flowchart showing a flow of processing executed by the image reference device 1 in organizing the search space. FIG. 9 is a diagram for explaining an example of a connection table created by the connection information table creation unit 23. FIG. 10 is a diagram for explaining another example of the connection table created by the connection information table creation unit 23. FIG. 11 is a conceptual diagram for explaining a result obtained by search space organization processing using a digest image. FIG. 12 is a diagram illustrating an example of an interface for presenting an organized search space. FIG. 13 is a diagram showing a typical processing flow in conventional image management.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Image reference apparatus 2 ... Digest image memory | storage part 3 ... DICOM original image memory | storage part 4 ... Operation part 5 ... Display part 6 ... Transmission / reception part 7 ... Learning / classification part 8 ... Search space organization part 11 ... Self-learning part 13 ... Sub-problem dividing unit 15 ... sub-problem learning unit 17 ... re-learning determining unit 18 ... search space organizing unit 19 ... search space managing unit 21 ... digest image generating unit 23 ... connection information table generating unit 25 ... SOM management label generating unit 27 ... Image corresponding part

Claims (7)

Storage means for storing a plurality of first images relating to a predetermined part of the subject;
The plurality of first images are input, and self-mapping is performed with the anatomical structure of the predetermined site as a phase characteristic of each first image by a predetermined weighting factor, and each of the classified first images is classified. Self-mapping means for classifying the plurality of input first images by outputting a plurality of output units composed of one image;
A second image corresponding to each output unit is generated by imaging the predetermined weight coefficient used for the self-mapping, and a correlation coefficient indicating a similarity between the second images is calculated. Image generating means for
Dividing means for dividing the obtained plurality of output units into at least two groups based on the calculated correlation coefficient;
A judgment means for judging whether or not to repeatedly execute the self-mapping based on the result of the self-mapping, the calculated correlation coefficient, and the result of the division;
If the determination means determines that the self-mapping is repeatedly performed, the self-mapping, the generation of the second image, the calculation of the correlation coefficient, the division, and the determination are repeated.
An image reference device comprising: Connection information creating means for creating a connection information table in which the second image corresponding to each output unit and the first image constituting each output unit are associated with each other;
Selecting means for selecting any one of the second images corresponding to each output unit;
Search means for searching for the first image constituting the output unit corresponding to the second image selected from the storage means based on the connection information table;
Output means for outputting the searched first image;
The image reference device according to claim 1, further comprising: A plurality of first images relating to a predetermined portion of the subject;
The plurality of first images are input, and the anatomical structure of the predetermined portion is constituted by the first images classified as phase characteristics of the first images by a predetermined weighting factor. A second image of each of the plurality of output units obtained by the self-mapping by imaging the predetermined weight coefficient used for self-mapping that outputs the plurality of output units;
A connection information table in which the second image corresponding to each output unit is associated with the first image constituting each output unit;
Storage means for storing
Selecting means for selecting any one of the second images corresponding to each output unit;
Search means for searching for the first image constituting the output unit corresponding to the second image selected from the storage means based on the connection information table;
Output means for outputting the searched first image;
An image reference device comprising:   The first image is an image acquired by a magnetic resonance imaging apparatus, an X-ray CT apparatus, an ultrasonic diagnostic apparatus, a nuclear medicine diagnostic apparatus, an X-ray diagnostic apparatus, or other medical imaging equipment. The image reference device according to any one of 1 to 3. A plurality of first images related to a predetermined part of the subject are input, and self-mapping is performed with the anatomical structure of the predetermined part as a phase characteristic of each first image by a predetermined weighting factor. A first step of classifying the input plurality of first images by outputting a plurality of output units composed of the classified first images;
A second image corresponding to each output unit is generated by imaging the predetermined weight coefficient used for the self-mapping, and a correlation coefficient indicating a similarity between the second images is calculated. A second step of:
A third step of dividing the obtained plurality of output units into at least two groups based on the calculated correlation coefficient;
A fourth step of determining whether or not to repeatedly execute the self-mapping based on the result of the self-mapping, the calculated correlation coefficient, and the result of the division;
If it is determined that the self-mapping is repeatedly executed in the fourth step, a fifth step of repeating the self-mapping, generation of the second image, calculation of the correlation coefficient, division, and the determination; ,
An image classification search method comprising: A sixth step of creating a connection information table in which the second image corresponding to each output unit and the first image constituting each output unit are associated with each other;
A seventh step for selecting one of the second images corresponding to each output unit;
An eighth step of searching for the first image constituting the output unit corresponding to the second image selected from the storage means based on the connection information table;
A ninth step of outputting the searched first image;
The image classification search method according to claim 5, further comprising:   The first image is an image acquired by a magnetic resonance imaging apparatus, an X-ray CT apparatus, an ultrasonic diagnostic apparatus, a nuclear medicine diagnostic apparatus, an X-ray diagnostic apparatus, or other medical imaging equipment. 5. The image classification search method according to 5 or 6.

Images