JP4799251B2 - Similar case search device, similar case search method and program thereof - Google Patents

Description

  The present invention relates to a similar case search apparatus, a similar case search method, and a program for searching for cases similar to cases in medical images.

  Conventionally, doctors have performed diagnostic interpretation by displaying images stored in a database on the screen using a workstation for interpretation of a medical system. There is still. For example, doctors thought that it might be stomach cancer, but there is a wonder that it may be a stomach ulcer. In that case, comparison is performed by referring to images of similar cases in the past. For example, in the past case, we diagnosed gastric cancer in the radiology department, but in the definitive diagnosis, we searched for cases that were gastric ulcers (diagnostic discrepancy example), displayed the image on the screen, and compared it with the image currently being interpreted I do.

  However, since the database provided in the conventional system is configured to simply search for the target image based on the patient number and examination number, there is a discrepancy between the previous diagnosis and the diagnosis in the medical record in order to find a diagnosis discrepancy example. The case was searched and the examination number / patient number was input from the workstation, and the searched image was displayed. In addition, when searching for past examples using memory, it was necessary to remember the examination number and patient number.

  Therefore, the radiology interpretation results, the requested department diagnosis results, the CAD (Computer Aided Diagnose: Computer Diagnosis Support) diagnosis results and the final definitive diagnosis, and whether there is a discrepancy, If there are discrepancies in the diagnosis results, save them as reference cases, and if you want to check past cases compared to the images that the doctor is currently interpreting, you only need to search for images of limited reference cases What has been made is proposed (for example, Patent Document 1).

  In addition, a medical diagnosis involves diagnosing both soft and hard tissues and observing images taken with multiple modalities to analyze the characteristics and status of specific anatomical structures. The diagnostic performance can be improved. However, the interpretation results of images taken with each modality may be different. Therefore, these interpretation results are collated using the CAR (Computer Aided Collation) method, and the contents of interpretation of two or more doctors are collated and displayed in such a way that different locations can be recognized in the case of different interpretation results by doctors. A technique for paying attention to a place has been proposed (for example, Patent Document 2).

On the other hand, in recent years, there has been proposed a system that assists doctors in interpretation by allowing similar cases of a diagnosis target case to be searched and referenced using images having similar feature amounts. In addition, in such an image search, there has been proposed a method in which a similarity is calculated using a feature amount of an image and displayed from cases having a high similarity.
JP-A-5-101122 JP 2004-199691 A

  However, the apparent similarity depends on the subjective factors of the doctor, so when learning the search engine (classifier) used to calculate the similarity, it is not possible to decide what to use as the correct reference very difficult.

  Also, images that appear similar but have a high possibility of different diseases are stored in association with the definitive diagnosis, and when the doctor interprets the images, check them with reference to the stored images. The desire to do so is higher than images with a high accuracy rate.

  Therefore, in the present invention, in order to solve such a problem, a similar case search device that can search for similar cases that strengthen diagnosis support for cases that are not good for each doctor or cases with a high misdiagnosis rate, A similar case search method and its program are proposed.

The similar case search device of the present invention includes a case storage means for storing a plurality of case data including a case image and a diagnosis result of the case image,
The feature amount obtained from the case image in the plurality of case data stored in the case storage means and the diagnosis result of the case image are learned, and the case data with similar diagnosis results are in the same class A classifier that divides the case data into a plurality of classes so as to belong to
A feature amount obtained from a diagnosis target image is input to the classifier to obtain a class to which the diagnosis target image belongs, and a case similar to the diagnosis target image from among case images of the case data belonging to the class A similar case retrieval means for retrieving an image and setting the case data of the similar case image as similar case data similar to the case of the diagnosis target image;
When similar case data is retrieved using the similar case retrieval means, the specific case data is obtained from the specific case image of the specific case data so that the specific case data is searched more frequently than other case data learned by the classifier. It is characterized by further comprising an additional record learning means for causing the classifier to additionally record the obtained feature amount and the diagnosis result of the specific case image.

The similar case retrieval method of the present invention is a feature obtained from the case image in a plurality of case data stored in case storage means for storing a plurality of case data including a case image and a diagnosis result of the case image. A classifier generating step of learning a quantity and a diagnosis result of the case image, and generating a classifier in which the case data is divided into a plurality of classes so that the case data having similar diagnosis results belong to the same class; ,
A feature amount obtained from a diagnosis target image is input to the classifier to obtain a class to which the diagnosis target image belongs, and a case similar to the diagnosis target image from among case images of the case data belonging to the class A similar case search step of searching for an image and setting the case data of the similar case image as similar case data similar to the case of the diagnosis target image;
When searching for the similar case data, the feature amount obtained from the specific case image of the specific case data and the specific value are searched so that the specific case data is searched more frequently than other case data learned by the classifier. And a postscript learning step for allowing the classifier to additionally learn the diagnosis result of the case image.

The program of the present invention is a computer,
Case storage means for storing a plurality of case data including case images and diagnosis results of the case images,
The feature amount obtained from the case image in the plurality of case data stored in the case storage means and the diagnosis result of the case image are learned, and the case data with similar diagnosis results are in the same class A classifier that divides the case data into a plurality of classes so as to belong to
A feature amount obtained from a diagnosis target image is input to the classifier to obtain a class to which the diagnosis target image belongs, and a case similar to the diagnosis target image from among case images of the case data belonging to the class A similar case retrieval means for retrieving an image and setting the case data of the similar case image as similar case data similar to the case of the diagnosis target image;
When similar case data is retrieved using the similar case retrieval means, the specific case data is obtained from the specific case image of the specific case data so that the specific case data is searched more frequently than other case data learned by the classifier. The obtained feature amount and the diagnosis result of the specific case image are made to function as additional recording learning means for additionally learning the classifier.

  “Diagnosis result” refers to a result obtained by diagnosing a medical image. For example, information input by a doctor diagnosing an image, information on a definitive diagnosis obtained by pathological examination, or medical image using CAD or the like. Information on the results of automatic analysis.

  “Class” refers to a group of case data divided into several groups based on a certain standard.

  “Dividing the case data into a plurality of classes so that the case data with similar diagnosis results belong to the same class” is, for example, classifying the same diagnosis results such as disease names into the same class, It is desirable to classify the same class so that case data of one disease name can be included, but it is also possible to classify the data so that case data of a plurality of disease names can be included.

  The similar case search means may use case data having high similarity between the feature amount of the diagnosis target image and the feature amount of the case image of the case data as the similar case data.

  The similar case retrieval unit may further retrieve a case image similar to the diagnosis target image from case images of case data belonging to a class adjacent to the class to which the diagnosis target image belongs.

  The additional record learning means may be configured to make the classifier learn by making the weight of the specific case data larger than the other case data.

  The specific case data may be case data having a high probability that the diagnosis of the specific case image is a misdiagnosis.

  In the present invention, a classifier learned so that case data with similar diagnosis results are in the same class is used to determine the class to which the image to be diagnosed belongs, and to detect similar cases from the case data included in the class To do. At this time, it is possible to perform accurate interpretation by paying attention to the specific case by additionally learning the specific case so that the detection frequency of the specific case such as a case where the doctor's misdiagnosis rate is high.

  In addition, by detecting case data with high similarity between the feature amount of the diagnosis target image and the case image feature amount of the case data as similar case data, it is possible to detect similar cases in which the image features are similar but different As a result, it is possible to accurately determine which of the different diseases the image to be diagnosed is.

  By searching for case images of case data belonging to a class to which a diagnosis target image belongs and a class adjacent to the class, many cases having similar image characteristics but different disease names can be detected.

  Case data that has a high probability of being misdiagnosed as a specific case image can be used to search for similar cases at a high frequency so that doctors and hospitals can check frequently misdiagnosis cases to prevent misdiagnosis. become able to do.

  Embodiments of a similar case retrieval apparatus according to the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of a medical system provided with a similar case retrieval apparatus according to an embodiment of the present invention. As shown in FIG. 1, a medical system 1 according to the present embodiment includes a modality 2 that captures a medical image, an image server 3 that stores a medical image obtained by capturing an affected area of a subject, and a workstation 4 on which a doctor interprets an image. The report server 5 stores a report created by the interpretation doctor interpreting the medical image displayed on the screen of the workstation 4, the electronic medical record server 6 that stores the electronic medical record, and the network 7.

  Modality 2 includes patients such as CR (Computed Radiography), CT (Computed Tomography), MRI (Magnetic Resonance Imaging), US (Ultrasound imaging), etc. It is a device or system for recording the image of the image as digital data.

  The image server 3 receives a large amount of medical images taken by each modality 2 installed in the radiology department via the network 7 and receives the medical images together with patient information on the subject, photographing date, information on the taken modality, and the like. Save in DICOM compliant format. In addition, the image server 3 is installed with database management software and has a function of searching for stored medical images using various information attached to the medical images.

The workstation 4 has a function of reading out a medical image to be interpreted from medical images stored in the image server 3 via the network 7 and displaying it on the screen of the image display device. In order for an interpreter such as a doctor to interpret a medical image, the workstation 4 is preferably provided with a display device such as a high-definition CRT.
The report server 5 stores the result of the medical image read by the image interpretation doctor with the image interpretation device 4 as an image interpretation report.

  The electronic medical record server 6 stores the electronic medical record of each patient. In the electronic medical record, medical date and time, input person, medical record, various findings, test results, instruction record, insurance score claim items, and the like are recorded. Furthermore, order information of various examinations such as image photographing is also recorded, and medical images obtained by photographing each patient stored in the image server 3 and pathological result information such as biological examinations can be obtained via the ordering system.

  Further, a program for searching for similar cases is installed in the workstation 4 so as to function as the similar case searching apparatus 40. On the screen of the workstation 4, case data 100 similar to the diagnosis target image is displayed together with the diagnosis target image. Note that the program for searching for similar cases may be on another server and started from a workstation.

  As shown in FIG. 2, the similar case search device 40 includes a case storage unit 41 that stores a plurality of case data 100 and a classifier 42 that learns the plurality of case data 100 and classifies the case data into a plurality of classes. And a classifier 42 to obtain a class to which the diagnosis target image belongs, and search for case data 140 similar to the case of the diagnosis image from the case data in this class, and at the time of the search Additional recording learning means 44 is provided for allowing the classifier to perform additional recording learning so that the frequency of searching for specific case data increases.

  The case data 100 includes a case image 110 and a diagnosis result 120 of the case image 110. The case image 110 of the case data 100 is selected from medical images stored in the image server 3. The diagnosis result 120 includes an interpretation report of the case image 110 stored in the report server 5, a medical record of the electronic medical record corresponding to the case image 110 stored in the electronic medical record server 6 (including definitive diagnosis and various findings), etc. It is. The selection of the case image 110 is considered to be effective for comparison with the diagnosis target image at the time of diagnosis with reference to the interpretation report of the case image 110, the medical record of the electronic medical record corresponding to the case image 110, and various findings. A thing is selected and stored in the case storage means 41.

  The classifier 42 uses, for example, a histogram, a gray value, an average value of pixel values, a centroid, an entropy, an evaluation value, an edge, from the case image 110 of each case data 100 used for learning stored in the case storage unit 41. Various feature quantities such as contrast are calculated, and classification is performed with the disease name obtained by the definitive diagnosis in the diagnosis result 110 as the correct answer. The classifier 42 weights specific data such as a decision tree, a Bayes discriminator, a linear discriminator such as LDA (Linear Discriminant Analysis), and a nonlinear discriminator such as SVM (Support Vector Machine). Any classifier that can perform learning may be used. In this embodiment, a case where a decision tree is used as a classifier will be described in detail below.

Decision tree, as shown in FIG. 3, a technique to continue to classify data as tree branches on the basis of the various conditions, as a condition for classifying, for example, entropy as the evaluation formula and a minimum At a certain point, the threshold of each feature amount is searched and divided so that the class is divided into two. In the example of FIG. 3, although it is finally divided into classes 1, 2, 3, 4 and 5, the “root” (see FIG. 3A) uses the feature quantity i (see FIG. 3B). When the entropy is to search for a place where the minimum Te, can be divided into sets of sets and classes 2 and 4 of the final divided class 1, 3 and 5. Similarly, if each feature is divided using feature quantities, it is finally divided into classes 1, 2, 3, 4, and 5. Specifically, for example, MacArthur, S., Brodley CE, and Broderick, L., “Interactive Content-Based Image Retrieval Using Relevance Feedback”, Computer Vision and Image Understanding, 88, pp 55-75, 2002. 1) can be used. Reference 1 includes bronchiectasis, mixed connective tissue disease, idiopathic organizing pneumonia, lymphocytic interstitial pneumonia, cancerous lymphangiopathy, etc. Classification can be made for various disease names. Ideally, it should be divided so that the case data 100 of a single disease name belongs to a leaf (class), but learning that the case data 100 of one disease belongs and belongs to multiple leaves Often results. For example, each mountain in FIG. 3 (b) is case data corresponding to a disease name, but if this is classified, it is divided by a dashed line, and the case data 100 of one disease is divided and belongs to a plurality of leaves. Become. Usually, there are many cases where there are about 2.5 diseases per leaf. The decision tree can be generated using a commercially available software tool such as C5.0.

  The similar case search means 43 calculates a feature amount from the diagnosis target image 130, determines which leaf it belongs to using the learned decision tree, and the diagnosis target image and the leaf into which the diagnosis target image is classified For example, the similarity between the case data 100 and the case image 110 is calculated using the distance in the feature amount space between the case image 110 and the diagnostic case image 130, for example. Specifically, the search is performed using the N-neighbor method to search for cases with high similarity among the case data 100 with a short distance in order from the top to obtain similar case data 140. Alternatively, not only those belonging to the same leaf (class) as the diagnosis target image 130 but also case data 100 having a close distance in the feature amount space from adjacent classes may be used as cases having high similarity.

Postscript learning means 44, when searching the similar case data, e.g., data (specific case data) 150 of the particular case, the classifier 42 is searched previously learned other case data 100 good Ri have frequently As described above, the specific case data 150 is additionally recorded in the classifier 42.

When additional writing learning is performed, learning is performed with the cost for the specific case data 150 set higher than other cases. For example, the “differential misclassification costs” provided in the commercial tool C5.0 is changed, and the specific case data 150 is weighted to be learned. Alternatively, when using the entropy minimum as the evaluation formula, it is possible to obtain the same effect by increasing the number of samples desired to increase the weight virtually.

  As the specific case data 150, the case data 100 that is easily misdiagnosed by a doctor, the case data 100 interpreted by an interpreting doctor called a famous doctor, and the like may be selected so that the similar case search means 43 may increase the detection frequency. . Further, the specific case data 150 may be selected from the case data 100 stored in the case storage means 41. However, the case data 100 in which a doctor is likely to make a diagnosis mistakenly, the interpretation of an interpreting doctor called a master doctor. The case data 100 thus obtained may be added to the case storage means 41 for learning.

  Next, learning of a classifier, a search method for searching for similar images similar to the diagnosis target image using the learned classifier, and a postscript learning method for the classifier will be described based on the flowchart of FIG.

  First, the classifier 42 provided in the workstation 4 uses the case data 100 stored in the case storage means 41 to determine that the classifier 42 is such that case data with the same disease name is in the same class. Tree learning is performed. Furthermore, the case data 100 stored in the case storage means 41 is classified into each class of the learned classifier 42 (S100).

  When the interpretation doctor performs interpretation, first, a list of medical images to be diagnosed is displayed on the screen of the workstation 4. When the interpretation doctor designates the diagnosis target image 130 from the list, the image server 3 is displayed. The diagnosis target image 130 is retrieved from the list and displayed on the display device (S101). The image interpretation doctor creates an image interpretation report by observing the displayed diagnosis target image 130, but there is a case where the diagnosis is lost.

  If the interpretation doctor is unsure of interpretation, the similar case retrieval unit 43 retrieves the case data 100 similar to the diagnosis target image 130. When the search of the similar case data 150 is instructed, the workstation 4 determines which class the diagnosis target image 130 belongs to by using the classifier 42 in the similar case search means 43, and the diagnosis target image The case image 110 having a high similarity to the diagnosis target image 130 is searched from the case data 100 belonging to the same class as 130 (S102), and the case data 100 of the case image 110 having a high similarity is displayed as the similar case data 140 on the screen. Displayed above (S103). Since the same class includes case images 110 of a plurality of diseases, a plurality of similar case data 140 having different disease names are retrieved even in the same class, and the doctor searches for the case images 110 of the similar case data 140. By comparing, it is possible to accurately determine which disease. Further, if similar case images 110 are searched from adjacent classes, it is possible to search a large number of similar case data 140 having similar disease names but different disease names.

  FIG. 4 shows an example in which the top five similar case data are displayed. The interpretation doctor compares the case image 110 of the similar case data 140 displayed on the screen with the diagnosis target image 130 and also refers to the diagnosis result 120 of the similar case data 140. The interpretation doctor performs interpretation with reference to the similar case data 140, and inputs the interpretation result as an interpretation report (S104). When the input of the interpretation report is completed, the interpretation report is transmitted to the report server 5 and stored (S105).

  Therefore, the specific case data 150 for additional learning is created from the images and the interpretation reports read by each doctor in daily routines, and stored in a predetermined storage device. Since the medical image and the report are normally stored in the image server 3 or the report server 5, the medical image that has been interpreted may be retrieved from the server and extracted as specific case data 150 at a later date.

  Or, when a definitive diagnosis (pathological examination result) is given for each case image, the interpretation result recorded in the interpretation report is compared with the definitive diagnosis result, and the interpretation result is incorrect May be extracted as specific case data 150. This comparison may be performed manually, or the result of the definitive diagnosis of the electronic medical record and the interpretation result of the interpretation report may be automatically compared to extract cases with different results as specific case data.

  In this way, a case where the diagnosis result of a specific doctor is different from the diagnosis result of the interpretation report is designated as specific case data 150 (S106), and in addition to the case data 100 of the case storage means 41 (S107), the additional record learning means 44 Addition learning (relearning) of the decision tree is performed (S100). In the postscript learning, learning is performed with the cost for the specific case data 150 set higher than other cases. In this way, by performing additional learning of what has been misdiagnosed by a specific doctor, by searching for similar case data 140 using the re-learned classifier 42, the misdiagnosed cases learned by weighting are ranked higher. It is possible to search for similar cases that are more likely to be ranked, and it is possible to realize more enhanced diagnosis support for cases that are difficult for doctors to use.

  Further, in this embodiment, the case where weighting is applied to the case where the diagnosis of the interpretation report is different from the definitive diagnosis (the medical record and pathological diagnosis) and the relearning is performed, but the case where the weight is applied is limited to the above example. Not what you want.

  For example, in order to create a search engine that has learned the interpretation results of an interpreting doctor called a famous doctor, the diagnostic results of an interpretation report of a specific doctor may be re-learned, or multiple doctors may display details. A similar case with a large number of requests may be learned as a case where it is easy to make a mistake.

In addition, although an example in which re-learning is limited to a specific doctor has been shown, it is not necessary to limit to a specific doctor, and misdiagnosed cases can be regarded as cases that are easily mistaken, and weighted learning is also possible. Good.
In the above description, information obtained from an interpretation report, an electronic medical record, or the like has been described as a diagnosis result. However, a CAD result may be included.

  Further, in the above description, the case where the class is divided by the disease name has been described, but the classification is performed by determining the range in which the diagnosis results are similar, including information other than the disease name such as the disease name and the malignancy level and progress of the disease. You may do it.

  As described above in detail, by making it easy to search for a specific case as a similar case, it is possible to prompt the doctor to call out the specific case. If this specific case has a high misdiagnosis rate or is diagnosed by a well-known physician, it is possible to improve the interpretation accuracy with reference to these means.

Block diagram showing schematic configuration of medical system Block diagram showing schematic configuration of similar case search device Illustration for explaining a decision tree An example of how to display similar cases Flow chart showing a method for searching similar images similar to the diagnosis target image using the learned classifier and the flow of additional learning of the classifier

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Medical system 2 Modality 3 Image server 4 Work station 5 Report server 6 Electronic medical record server 7 Network 40 Similar case search device 41 Case memory means 42 Classifier 43 Similar case search means 44 Additional record learning means
100 case data
110 case images
120 diagnostic results
130 Images to be diagnosed
140 Similar case data
150 Specific case data

Claims (5)

Case storage means for storing a plurality of case data including case images and diagnosis results of the case images,
The feature amount obtained from the case image in the plurality of case data stored in the case storage means and the disease name included in the diagnosis result of the case image are learned as correct answers, and the entropy of the feature amount is minimized. Then, a classifier classified according to each disease name ,
For specific case data among the case data stored in the case storage means, additional writing learning means for making the classifier to perform additional writing with a greater weight than other case data stored in the case storage means; ,
The feature amount obtained from the diagnosis target image is input to the classifier, the class to which the diagnosis target image belongs is obtained, and the feature amount of the diagnosis target image is selected from the case images of the case data belonging to the class. Find the case image with a high similarity feature value as a case image that is similar to the diagnosis target image, similar case search to similar case data similar case data of the similar case image to cases of the diagnosis target image similar case search apparatus characterized by comprising a means. The similar case retrieval means further retrieves a case image similar to the diagnosis target image from case images of case data belonging to a class adjacent to the class to which the diagnosis target image belongs in the feature amount space. The similar case retrieval apparatus according to claim 1 . The specific case data is similar case retrieval apparatus according to claim 1 or 2, wherein the probability diagnosis was misdiagnosed particular case image has a high case data. Case storage means for storing a plurality of case data including case images and diagnosis results of the case images ,
A classifier generating means for generating a classifier obtained by learning a plurality of case data stored in the case storage means;
Additional learning means for causing the classifier to perform additional learning;
A similar case retrieval method of a similar case retrieval device comprising similar case retrieval means for retrieving similar case data similar to a case of a diagnosis target image,
The classifier generation means learns the feature amount obtained from the case image in the plurality of case data stored in the case storage means and the disease name included in the diagnosis result of the case image as a correct answer, A classifier generating step for generating a classifier classified according to each disease name where the entropy of the feature amount is minimized ;
The specific case data of the case data stored in the case storage means is added to the classifier with a higher weight than the other case data stored in the case storage means by the additional record learning means. A postscript learning step to learn,
By the similar case search means, and inputs the feature quantity obtained from the diagnostic object image on the classifier, to obtain the class of the diagnosis target image belongs, said from among case image of the case data belonging to the class A case image having a feature amount having a high similarity to the feature amount of the diagnosis target image is searched as a case image similar to the diagnosis target image, and the case data of the similar case image is similar to the case of the diagnosis target image. similar case search method characterized by comprising: a similar case retrieval step of the case data. Computer
Case storage means for storing a plurality of case data including case images and diagnosis results of the case images,
The feature amount obtained from the case image in the plurality of case data stored in the case storage means and the disease name included in the diagnosis result of the case image are learned as correct answers, and the entropy of the feature amount is minimized. Then, a classifier classified according to each disease name ,
For specific case data among the case data stored in the case storage means, additional writing learning means for making the classifier to perform additional writing with a greater weight than other case data stored in the case storage means; ,
The feature amount obtained from the diagnosis target image is input to the classifier, the class to which the diagnosis target image belongs is obtained, and the feature amount of the diagnosis target image is selected from the case images of the case data belonging to the class. Find the case image with a high similarity feature value as a case image that is similar to the diagnosis target image, similar case search to similar case data similar case data of the similar case image to cases of the diagnosis target image similar case search program to function as a means.

Images