Diagnosis support program, computer readable recording medium recorded with diagnosis support program, diagnosis support device and diagnosis support method

[Technical field to which the invention belongs] The present invention relates to a diagnostic support technique for supporting diagnosis by interpreting an image, and more particularly to a technique for improving diagnostic accuracy.

[0002] Conventional Techniques Traditionally, interpretation of CT (Computed Tomography) images, MRI (Magnetic Resonance Imaging) images, etc. (hereinafter referred to as "CT images") has been cultivated by doctors such as radiologists over a long period of time. It has been done by subjective judgment based on the experience that has been gained. However, misdiagnosis due to oversight or misunderstanding cannot be avoided by image diagnosis based only on subjective judgment. In order to avoid such misdiagnosis, various measures have been taken, such as reading CT images by multiple doctors, but many problems such as time constraints remain.

[0003] On the other hand, at present, the digitization of CT images is rapidly progressing, and almost all CT images from simple photographs to angiography images are being digitized. Then, in order to make all the digitized CT images useful for diagnosis, PACS (Picture Archiving and Communication System) was developed to rapidly transmit and store images.

[0004] [Problems to be Solved by the Invention] By the way, in the interpretation of CT images, it is known that the diagnostic accuracy is improved by referring to CT images taken in the past having similar cases. There is. However, in the conventional PACS, since only digital images are accumulated and referenced, it is extremely difficult to select an appropriate reference image from a large amount of accumulated CT images. For this reason, despite the development of PACS, the utilization of CT images taken in the past was insufficient, and diagnosis was still performed based on the subjective judgment of doctors, making it difficult to improve diagnostic accuracy. ..

[0005] Therefore, in view of the above-mentioned conventional problems, the present invention makes it possible to search for similar cases using the feature amount of the image when interpreting the CT image, and to improve the diagnostic accuracy. The purpose is to provide technology.

[Means for Solving the Problems] Therefore, in the diagnostic support technique according to the present invention, the image feature amount of the lesion position detected from the diagnostic image is extracted, and the reference image and the feature amount are accumulated. It is characterized by searching for reference images that are image-similar based on the extracted features from the database.

[0007] According to such a configuration, a reference image similar to the image is searched from the reference images stored in the database based on the feature amount of the lesion position detected from the diagnostic image. Therefore, a doctor who interprets the diagnostic image and makes a diagnosis can easily refer to a past case similar to the case of the lesion appearing in the diagnostic image. In addition, it is desirable to register the diagnostic image and its feature amount in the database. In this way, the diagnosed diagnostic image and its feature amount are registered in the database, and are searched as the reference image and its feature amount in the subsequent diagnosis. Therefore, as the interpretation of the diagnostic image progresses, the reference materials are gradually enriched.

[0008] Further, the feature amount of the reference image stored in the database is collated with the feature amount of the diagnostic image to calculate the image similarity, and the image similarity is calculated.<u style="single">A plurality of reference images are selected from those having a high degree of similarity according to a predetermined rule.</u>Here, it is desirable that the similarity is calculated in consideration of the weight set for each organ, and it is desirable that the weight is set in a table configured to be changeable.

[0009] According to such a configuration, the similarity between the diagnostic image and the reference image is defined by the image similarity, and the reference image having the strongest similarity is used in order.<u style="single">Be selected</u>It will be. Therefore, when interpreting the diagnostic image, the image that has a strong possibility of becoming a reference image is referred to, so that unnecessary reference is prevented and the diagnostic efficiency is improved. Here, if the similarity is calculated in consideration of the weighting set for each organ, the similarity is calculated according to the characteristics of the organs, and the calculation accuracy of the similarity is improved. Further, if weighting is set in the table that is configured to be changeable, for example, correction according to the characteristics peculiar to the CT device becomes possible, and the calculation accuracy of the similarity is further improved.

[0010] In addition to this, it is desirable to display the findings associated with the searched reference image. In this way, in addition to the reference image, the findings can be referred to, and even when the patient's disease name is unclear, the disease name can be diagnosed from the findings of similar cases. .. Furthermore, when detecting the lesion position from the diagnostic image, it is desirable to detect the lesion position of the designated organ. In addition, as image features, it is desirable to extract wide-area features, local features, and common features for all lesion positions on the diagnostic image.

[0011] According to such a configuration, the detection process of the organ not to be diagnosed is not performed, and the process speed is improved. Further, if a wide-area feature amount, a local feature amount, and a common feature amount are detected as image feature amounts, it is possible to extract the feature amount according to the characteristics of the disease, for example. Tumor oversight is prevented.

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described in detail below with reference to the accompanying drawings. FIG. 1 shows the configuration of a diagnostic support device that embodies the diagnostic support technology according to the present invention. The diagnostic support device is built on a computer system equipped with at least a central processing unit (CPU) and memory, and is operated by a program loaded in the memory.

[0013] The diagnostic support device includes an image database 10, a feature database 12, a finding database 14, a lesion position detection unit 16, a lesion feature extraction unit 18, a lesion feature matching unit 20, and a disease name probability calculation unit. 22 and is composed of. In the following description, the database will be referred to as DB. Image data such as CT images and MRI images are stored in the image DB 10 in a format compliant with DICOM (Digital Imaging and Communications in Medicine). Here, DICOM is a standard for medical digital images and communication developed in the United States, and is also a standard adopted in Japan. Then, in order to correlate the image DB10, the feature DB12, and the finding DB14, at least the patient ID, the examination ID, and the image file name are registered in the index of the image DB10, as shown in FIG. Information such as the inspection date and the target site (organ name) is registered in the DICOM header of the image file.

[0014] In the feature DB 12, the lesion feature amount extracted from the image data is accumulated. The lesion feature amount is information that enables the retrieval of image data using image similarity, and is, for example, statistics on the size (volume or area), shape (sphericity or circularity), and brightness of the lesion. Quantities (means, deviations, etc.) and texture statistics (spatial frequency decomposition, Fourier transform, wavelet transform, etc.) are used. The lesion feature amount is collected as feature data associated with the patient ID and the examination ID. Then, as shown in FIG. 3, at least the target site, the inspection date, and the feature data are registered in the index of the feature DB12. Here, if the feature data is divided so as to have a one-to-one correspondence with the image data, it takes a long time to access the file at the time of searching. Therefore, for example, it is desirable to unify the feature data on a monthly basis. In this case, only the year and month of the inspection are registered on the inspection date. In addition, the search efficiency can be improved by separating the feature data for each target part.

[0015] Findings DB14 stores findings that are the result of interpretation of CT images. Then, as shown in FIG. 4, at least the patient ID, the examination ID, the name of the doctor in charge, and the findings are registered in the index of the findings DB14. The findings include at least the name of the disease diagnosed by interpreting the CT image. The lesion position detection unit 16 detects the lesion position in the organ or site to be diagnosed (hereinafter referred to as organ) from the CT image. The lesion position is detected by, for example, the technique shown in Noboru Niki "Lung Cancer CT Screening Support System", Journal of Japanese Society of Radiological Technology, Vol.56 No.3, March (2000), pp.337-340.

[0016] The lesion feature amount extraction unit 18 extracts the lesion feature amount that expresses the image similarity as a statistic for all the lesion positions detected by the lesion position detection unit 16. For the extraction of lesion features, for example, Masaki Kondo et al., "Classification of tumor shadows into aerobic and solid types by three-dimensional chest X-ray CT images and their application to benign / malignant differentiation" Technical Report of IEICE MI2000-16 ( 2000-05), pp.27-32 by the technique shown.

[0017] In the lesion feature amount collation unit 20, the lesion feature amount extracted by the lesion feature amount extraction unit 18 and the lesion feature amount accumulated in the feature DB 12 are collated with each other for the organ to be diagnosed, and a scale similar to an image. The similarity representing is calculated. The disease name probability calculation unit 22 calculates the probability of the disease name of the organ to be diagnosed based on the similarity calculated by the lesion feature amount matching unit 20. Then, the name of the disease and its probability are displayed on a display device (not shown) to support the diagnosis by a doctor.

[0018] Next, an outline of the diagnostic support device having such a configuration will be described. When tomography is performed by the X-ray CT apparatus, the CT image is accumulated in the image DB 10 and the lesion position is detected by the lesion position detection unit 16. When the lesion position is detected, the lesion feature amount is extracted for all the lesion positions. The extracted lesion feature amount is registered in the feature DB12, and the lesion feature amount matching unit 20 calculates the similarity of the CT images of the past cases accumulated in the feature DB12. Then, the disease name probability calculation unit 22 calculates the probability of the disease name according to the similarity of the CT images, and the probability is shown to the doctor together with the disease name in the order of probability.

[0019] Therefore, a doctor who makes a diagnosis by interpreting a CT image can easily refer to a past case similar to the case to be diagnosed, and by referring to the findings, the subjectivity is eliminated. You will be able to make an objective diagnosis. At this time, since the disease name of the case to be diagnosed and its probability are also displayed, the disease name can be diagnosed by referring to the displayed disease name and its probability even when the patient's disease name is not clear. In this way, the accuracy of diagnosis by a doctor can be improved.

[0020] FIGS. 5 to 7 show a flowchart showing the control contents of the diagnostic support device. It should be noted that such control is executed after, for example, a CT image is taken by an X-ray CT apparatus, an organ is designated, and a patient ID and an examination ID are input. In FIG. 5, which shows the main routine, in step 1 (abbreviated as S1 in the figure; the same applies hereinafter), the captured CT image (hereinafter referred to as diagnostic image) is registered in the image DB 10. At this time, the patient ID, the examination ID, and the image file name are additionally registered in the index of the image DB 10 in order to identify the image in association with the patient ID and the examination ID. In addition, it should be noted.<u style="single">The process in step 1 corresponds to a part of the database registration means and the database registration process.</u>[0021] In step 2, the lesion position in the designated organ is detected from the diagnostic image by the function provided by the lesion position detection unit 16. That is, the diagnostic image is narrowed down by the luminance value or the CT value, and filtering processing such as morphology is performed, and the contour of the organ is detected. Next, a difference is calculated that indicates how much the contour of the detected organ deviates from the normal range. The extent of a normal organ is defined by considering it as the inner contour of the ribs, for example in the case of the lungs. Then, the portion where the calculated difference is equal to or larger than the predetermined width is regarded as the lesion position.

[0022] In addition,<u style="single">The process in step 2 corresponds to the lesion position detecting means and the lesion position detecting step.</u>In step 3, the subroutine for extracting the lesion feature amount shown in FIG. 6 is called in order to extract the lesion feature amount for all the detected lesion positions. The process in step 3, that is,<u style="single">The entire process in FIG. 6 corresponds to the feature amount extraction means and the feature amount extraction step.</u>[0023] The process in step 2 corresponds to the lesion position detection function, the lesion position detection means, and the lesion position detection step. In step 3, the subroutine for extracting the lesion feature amount shown in FIG. 6 is called in order to extract the lesion feature amount for all the detected lesion positions. The process in step 3, that is, the entire process in FIG. 6, corresponds to the feature amount extraction function, the feature amount extraction means, and the feature amount extraction step.

[0024] In step 4, the extracted lesion feature amount is registered in the feature DB12. At this time, the lesion feature amount is incorporated into the feature data corresponding to the target site and the examination date and unified in order to improve the search efficiency. However, when the inspection dates are different, or when the corresponding feature data is not registered, the target site, the inspection date, and the feature data are additionally registered in the index of the feature DB12. In addition, it should be noted.<u style="single">The process in step 4 corresponds to a part of the database registration means and the database registration process.</u>[0025] In step 5, the lesion feature amount of the diagnostic image and the lesion feature amount accumulated in the feature DB 12 are collated and accumulated for the organ to be diagnosed by the function provided by the lesion feature amount matching unit 20. The similarity of the CT image (hereinafter referred to as "reference image") is calculated. That is, as shown in FIG. 8, the lesion feature amounts of the diagnostic image and the reference image are, for example, volume, average brightness, brightness deviation, sphericity, and texture in order from the first element for each image unit (inspection unit). They are arranged in a vector like statistics. Each element is normalized in the range 0 to 1 so that it can be compared at the same level. For example, the volume can be normalized because the volume of the lesion is always smaller than the volume of the lung when the target site is the lung.

Then, the two feature vectors A and B are set to A = (f).₁ f₂ f₃ ) B = (g₁ g₂ g₃ ) And the weighting vector W is W = (w)₁ w₂ w₃ ) Then, the similarity S is, for example,<img file="JP4021179B2_D0001.tif" />It is calculated as follows. Here, E represents a vector in which each component is 0, and | W | represents the sum of the components of the weighting vector W. Further, as shown in FIG. 9, it is desirable that the weighting vector W is set in a table format for each organ.

[0027] It is desirable that each component of the weighting vector W can be freely changed by the user. Further, the optimum weighting value may be set by learning the correspondence between the lesion feature amount and the findings in advance by the neural network. here,<u style="single">The process in step 5 corresponds to the similarity calculation means and the similarity calculation step.</u>[0028] In step 6, a subroutine for calculating the disease name probability shown in FIG. 7 is called in order to calculate the probability of the disease name of the lesion portion appearing in the diagnostic image. In step 7, the disease name and its probability are displayed, for example, by a screen as shown in FIG. In addition, it should be noted.<u style="single">The process in step 7 corresponds to the finding display means and the finding display step.</u>[0029] In step 8, the findings, which are the final diagnosis by the physician, are registered in Findings DB14. At this time, the doctor can make an objective diagnosis without subjectivity by referring to the disease name and its probability estimated from the diagnostic image, the reference image and its findings in addition to the diagnostic image. The findings registered here will be used as the findings of the reference image in the subsequent diagnosis, and as the number of registered findings increases, the materials for diagnostic support will be enhanced.

Since the reference image, feature data and findings are not accumulated at the start of operation of the diagnostic support device, the reference image, feature data and findings representing typical cases are registered in each DB as an initial state. It is desirable to keep it. FIG. 6 shows a subroutine for extracting lesion features. The lesion feature amount is extracted by the function provided by the lesion feature amount extraction unit 18.

[0031] In step 11, a wide area feature amount over a wide area of the lesion is extracted. That is, for example, spatial frequency decomposition, Fourier transform, and wavelet transform are applied to the entire detected lesion, and a texture statistic as a wide-area feature is extracted. In step 12, the local feature amount at which the lesion remains locally is extracted. That is, for example, sphericality (three-dimensional) or circularity (two-dimensional) is extracted from the detected local lesion.

[0032] In step 13, common feature amounts common to the lesions are extracted. That is, for example, statistics (mean, deviation, etc.) of size (volume or area) and brightness are extracted for the detected lesion. FIG. 7 shows a subroutine for calculating the disease name probability. The calculation of the disease name probability is performed by the function provided by the disease name probability calculation unit 22.

[0033] In step 21, among the feature data stored in the feature DB 12, the similarity is calculated, and as shown in FIG. 11, the rank, the test ID, the similarity, and the disease name are grouped and similar. Arranged in order of degree. At this time, the disease name can be obtained from the findings by searching the findings DB14 using the patient ID and the examination ID included in the feature data as keys.

[0034] In step 22, for example, the top 10 feature data or the feature data having a similarity of 50% or more is selected from the feature data arranged in order of similarity as a reference case useful for diagnosing the diagnostic image. Will be done. In step 23, for all selected reference cases, the disease name probability, which is the average of the similarity for each disease name, is calculated.

[0035] In step 24, the disease names and their probabilities are displayed in the order of disease name probabilities. If the correspondence between the lesion features and the findings has already been learned in the neural network, the disease name and its probability can be calculated directly from the lesion features. According to the diagnostic support device described above, the lesion position is detected from the diagnostic image of the designated organ, and the lesion feature amount at all the detected lesion positions is extracted. The extracted lesion feature amount is collated with the feature data for the same organ accumulated in the feature DB12, and the similarity of cases that can be a reference image is calculated. Then, the disease name and its probability are calculated based on the calculated similarity, and this is displayed on the display device.

[0036] Therefore, a doctor who interprets the diagnostic image to make a diagnosis can refer to a past case similar to the case of the lesion appearing in the diagnostic image, and makes an objective diagnosis without subjectivity. Can be done. Since the objective diagnosis is possible, for example, even a doctor who lacks experience in interpreting images can drastically reduce the possibility of misdiagnosis and improve the diagnosis accuracy. In addition, by utilizing the reference case, the disease name can be estimated even when the disease name is still not obtained from the diagnostic image alone, and the diagnostic accuracy can be improved.

[0037] Further, the diagnostic image for which the diagnosis has been completed is registered in the database together with the lesion feature amount and the findings, and is utilized as a reference image in the subsequent diagnosis. Therefore, as the interpretation of the diagnostic image progresses, the reference materials are gradually enriched, and the diagnostic accuracy can be further improved. Although the diagnostic support device in the present embodiment is built on a stand-alone computer system, it may be built on a client / server model connected via a network. In this case, reference cases can be accumulated on a nationwide or global scale, which contributes to the improvement of medical technology and is extremely useful from the standpoint of the public interest.

[0038] A program that realizes such a function can be recorded on a computer-readable recording medium such as a magnetic tape, a magnetic disk, a magnetic drum, an IC card, a CD-ROM, or a DVD-ROM. The diagnostic support program according to the invention can be distributed on the market. Then, a person who has acquired such a recording medium can easily construct a diagnostic support device according to the present invention by using a general computer system.

[0039] Further, if the diagnostic support program according to the present invention is registered on the server connected to the Internet, the diagnostic support device according to the present invention can be downloaded by downloading the program via a telecommunication line. Can be easily constructed.

(Appendix 1) A lesion position detection function that detects a lesion position from a diagnostic image, a feature amount extraction function that extracts an image feature amount of a lesion position detected by the lesion position detection function, a reference image, and a reference image. A diagnostic support program for realizing a reference image search function for a computer to search for a reference image that is image-similar based on the feature amount extracted by the feature amount extraction function from a database in which the feature amount is accumulated. ..

(Appendix 2) A lesion position detection function that detects a lesion position from a diagnostic image, a feature amount extraction function that extracts an image feature amount of a lesion position detected by the lesion position detection function, a reference image, and a reference image. A diagnostic support program for realizing a reference image search function for a computer to search for a reference image that is image-like based on the feature amount extracted by the feature amount extraction function from a database in which the feature amount is accumulated. A computer-readable recording medium that records images.

(Appendix 3) A computer-readable recording medium on which the diagnostic support program described in Appendix 2 is recorded, which is provided with a database registration function for registering the diagnostic image and its feature amount in the database.

(Appendix 4) The reference image search function is provided with a similarity calculation function for calculating image similarity by collating the feature amount of the reference image stored in the database with the feature amount of the diagnostic image. , A computer-readable recording medium on which the diagnostic support program according to Appendix 2 or Appendix 3, wherein the reference images are searched in the order of similarity calculated by the similarity calculation function.

(Appendix 5) The computer-readable recording medium on which the diagnostic support program described in Appendix 4 is recorded, wherein the similarity calculation function calculates the similarity in consideration of the weighting set for each organ. ..

(Appendix 6) A computer-readable recording medium on which the diagnostic support program described in Appendix 5 is recorded, wherein the weighting is set in a table configured to be changeable.

[0046] (Supplementary note 7) The description in any one of Supplementary note 2 to Supplementary note 6, wherein the finding display function for displaying the findings associated with the reference image searched by the reference image search function is provided. A computer-readable recording medium on which a diagnostic support program is recorded.

(Appendix 8) The lesion position detection function is a computer reading that records a diagnostic support program according to any one of Appendix 2 to Appendix 7, characterized in that the lesion position of a designated organ is detected. Possible recording medium.

(Appendix 9) The feature amount extraction function is characterized by extracting a wide range feature amount, a local feature amount, and a common feature amount for all lesion positions in the diagnostic image. A computer-readable recording medium on which the diagnostic support program described in any one of Appendix 2 to Appendix 8 is recorded.

(Appendix 10) A lesion position detecting means for detecting a lesion position from a diagnostic image, a feature amount extracting means for extracting an image feature amount of a lesion position detected by the lesion position detecting means, a reference image, and a reference image. It is characterized in that it is configured to include a reference image search means for searching an image-like reference image based on the feature amount extracted by the feature amount extraction means from a database in which the feature amount is accumulated. Diagnostic support device.

(Appendix 11) The diagnostic support device according to Appendix 10, further comprising a database registration means for registering the diagnostic image and its feature amount in the database.

(Appendix 12) The reference image search means includes a similarity calculation means for calculating image similarity by collating the feature amount of the reference image stored in the database with the feature amount of the diagnostic image. The diagnostic support device according to Appendix 10 or Appendix 11, wherein the reference images are searched in the order of similarity calculated by the similarity calculation means.

(Appendix 13) The diagnostic support device according to Appendix 12, wherein the similarity calculation means calculates the similarity in consideration of weights set for each organ.

(Appendix 14) The diagnostic support device according to Appendix 13, wherein the weighting is set in a table that is configured to be changeable.

(Appendix 15) The description in any one of Appendix 10 to Appendix 14, characterized in that the findings display means for displaying the findings associated with the reference image searched by the reference image search means is provided. Diagnostic support device.

[Appendix 16] The diagnostic support device according to any one of Supplementary note 10 to Supplementary note 15, wherein the lesion position detecting means detects a lesion position of a designated organ.

(Appendix 17) The feature amount extraction means is characterized by extracting a wide range feature amount, a local feature amount, and a common feature amount for all lesion positions in the diagnostic image. The diagnostic support device according to any one of Appendix 10 to Appendix 16.

(Appendix 18) A lesion position detection step of detecting a lesion position from a diagnostic image, a feature amount extraction step of extracting an image feature amount of a lesion position detected by the lesion position detection step, a reference image, and a reference image. Diagnostic support including a reference image search step for searching a reference image that is image-similar based on the feature amount extracted by the feature amount extraction step from a database in which the feature amount is accumulated. Method.

(Supplementary Note 19) The diagnostic support method according to Supplementary note 18, further comprising a database registration step of registering the diagnostic image and its feature amount in the database.

(Appendix 20) The reference image search step includes a similarity calculation step of collating the feature amount of the reference image stored in the database with the feature amount of the diagnostic image to calculate the image similarity. The diagnostic support method according to Appendix 18 or Appendix 19, wherein the reference images are searched in the order of similarity calculated by the similarity calculation step.

(Appendix 21) The diagnostic support method according to Appendix 20, wherein the similarity calculation step calculates the similarity in consideration of weights set for each organ.

[0061] (Appendix 22) The diagnostic support method according to Appendix 21, wherein the weighting is set in a table that is configured to be changeable.

[0062] (Supplementary note 23) The description in any one of Supplementary note 18 to Supplementary note 22, wherein the finding display step for displaying the findings associated with the reference image searched by the reference image search step is provided. Diagnosis support method.

[0063] (Appendix 24) The diagnostic support method according to any one of Appendix 18 to Appendix 23, wherein the lesion position detection step detects a lesion position of a designated organ.

(Appendix 25) The feature amount extraction step is characterized by extracting a wide range feature amount, a local feature amount, and a common feature amount for all the lesion positions in the diagnostic image. The diagnostic support method described in any one of Appendix 18 to Appendix 24.

[Effect of the Invention] As described above, according to the diagnostic support technique according to the present invention, a doctor who interprets a diagnostic image to make a diagnosis has a past similar to the case of a lesion appearing in the diagnostic image. You will be able to easily refer to the cases of. Therefore, it is possible to make an objective diagnosis without subjectivity. Since the objective diagnosis is possible, for example, even a doctor who lacks experience in interpreting images can drastically reduce the possibility of misdiagnosis and improve the diagnosis accuracy. In addition, by utilizing the reference case, the disease name can be estimated even when the disease name is not clear from the diagnostic image alone, and the diagnostic accuracy can be improved.

BRIEF DESCRIPTION OF THE DRAWINGS [Fig. 1] Fig. 1 is a configuration diagram of a diagnostic support device according to the present invention.

FIG. 2 is an explanatory diagram of an index of an image DB.

FIG. 3 is an explanatory diagram of an index of a feature DB.

FIG. 4 is an explanatory diagram of a finding DB index.

FIG. 5 is a flowchart of a main routine showing control contents.

FIG. 6 is a flowchart of a subroutine for extracting lesion feature amounts.

FIG. 7 is a flowchart of a subroutine for calculating a disease name probability.

FIG. 8 is an explanatory diagram of a method for calculating the degree of similarity.

FIG. 9 is an explanatory diagram of a weighted vector table.

FIG. 10 is an explanatory diagram of a screen displaying a disease name and its probability.

FIG. 11 is an explanatory diagram of reference cases arranged in order of similarity.

[Explanation of sign] 10 Image DB12 Feature DB14 Findings DB16 Lesion position detection unit 18 Lesion feature amount extraction unit 20 Lesion feature amount matching unit 22 Disease name probability calculation unit