JP2017067489A - Diagnosis support apparatus, method, and computer program - Google Patents

Abstract

The accuracy of diagnostic support using a machine learning algorithm is improved.
A diagnosis support apparatus that performs diagnosis support based on subject data that is data related to a subject includes a first discriminator C1 including a learned algorithm obtained by learning a first machine learning algorithm using a first data set. And a second discriminator C2 including a learned algorithm obtained by learning the first machine learning algorithm with a second data set different from the first data set, the output of the first discriminator C1 and the second And a determinator D that performs a determination according to the output of the second determinator C2. The first discriminator C accepts input of subject data and outputs a first discrimination result, the second discriminator C2 accepts input of subject data and outputs a second discrimination result, and the discriminator D In response to the first determination result and the second determination result, information for supporting the diagnosis of the subject is output.
[Selection] Figure 5

Description

  The present invention relates to a technique for providing diagnosis support based on subject data relating to a subject.

  A technique for performing a diagnosis support apparatus for a subject using a bone scintigram is known. In an existing diagnostic imaging support apparatus, a doctor uses a neural network (ANN) to determine whether a bone scintigram hotspot (a high-count region in which drugs are accumulated) is cancer metastasis, so that a doctor supports the diagnosis. (For example, Patent Document 1).

JP 2014-9945 A

Active shape models-tear training and application, T.A. F. Cootes, C.I. J. et al. Taylor, D.D. H. Cooper and J.M. Graham presented in Computer Vision and Image Understanding, Vol. 61, no. 1, pp. 38-59, 1995. Application of the Active Shape Model in a commercial medical device for bone densitometry, H.C. H. Thodberg and A.M. Rosholm presented in the Processings of the 12th British Machine Vision Conference, 43-52, 2001. Average brain models: A convergence study, Guimond A. Meunier J.M. Tillion J. -P presented in Computer Vision and Image Understanding, 77 (2), pp. 192-210, 2000.

  When a discriminator based on a machine learning algorithm such as ANN is used in a diagnosis support apparatus, if the state of a discriminator having a learned algorithm is different, different outputs may be made even for the same input. For example, even when the same learning algorithm is used, a discriminator trained using a data set and a discriminator trained using a different data set can obtain the same output when the same value is input. Not necessarily. Similarly, when the learning algorithms of the two discriminators are different from each other, the same output is not always output for the same input even when learning is performed with the same data set.

  In other words, when determining the state of a subject, the information output as diagnosis support information depends on the learning algorithm of the discriminator and the data set used for learning, and accuracy and stability as diagnosis support information There is a problem.

  Accordingly, an object of the present invention is to improve the accuracy in diagnosis support using a machine learning algorithm.

  Another object of the present invention is to provide information for stable diagnosis support in diagnosis support using a machine learning algorithm.

  A computer program according to an embodiment of the present invention is a computer program for a diagnosis support apparatus that receives subject data that is data relating to a subject and supports diagnosis of the subject, and the computer program is included in the diagnosis support apparatus. When executed, the first discriminator includes a learned algorithm obtained by learning the first machine learning algorithm with the first data set, and receives the input of the subject data and outputs the first discrimination result And a second discriminator including a learned algorithm obtained by learning the first discriminator and the first machine learning algorithm with a second data set different from the first data set. The second discriminator that receives the input of the subject data and outputs the second discrimination result, and the first discrimination Fruit and in response to the second determination result, to realize a determining unit for outputting the diagnosis support information is information for supporting a diagnosis of a subject, to the diagnosis support apparatus.

  A computer program according to another embodiment of the present invention is a computer program for a diagnosis support apparatus that receives subject data, which is data relating to a subject, and supports diagnosis of the subject, and the computer program is included in the diagnosis support apparatus. When executed, the first discriminator includes a learned algorithm obtained by learning the first machine learning algorithm with the first data set, and receives the input of the subject data and outputs the first discrimination result The first discriminator and a second machine learning algorithm different from the first machine learning algorithm are used as the first data set or a second data set different from the first data set. A second discriminator including a learned algorithm learned in step (b), which accepts input of the subject data Diagnostic support information which is information for supporting diagnosis of the subject according to the second discriminator that outputs the second discrimination result and the first discrimination result and the second discrimination result Is implemented in the diagnosis support apparatus.

  In a preferred embodiment, the first and second determination results may be probabilities as to whether or not the subject is in a predetermined state. The diagnosis support information may be information for supporting diagnosis, which is a probability of whether or not the subject is in the predetermined state or whether the subject is in the predetermined state.

  In a preferred embodiment, the first and second determination results may be whether or not the subject is in a predetermined state. The diagnosis support information may be information that supports a diagnosis as to whether or not the subject is in the predetermined state.

  In a preferred embodiment, the diagnosis support apparatus may further realize means for displaying the diagnosis support information and the first and second determination results on a display device.

  In a preferred embodiment, the diagnosis support apparatus may further realize means for displaying the diagnosis support information and the reliability of the diagnosis support information on a display device.

  In a preferred embodiment, the first machine learning algorithm may be an artificial neural network.

  In a preferred embodiment, the subject data may be data obtained from bone scintigrams.

  In a preferred embodiment, the diagnosis support apparatus may further realize accepting means for accepting a redetermination request. And when the said reception means receives a redetermination request | requirement, based on the said test subject data and the characteristic data which show the said test subject's characteristic, the said 1st discriminator and the said 2nd discriminator are new 1st and The second determination result may be output, and the determination device may output new diagnosis support information.

  According to still another embodiment of the present invention, a diagnosis support apparatus that accepts subject data that is data related to a subject and supports diagnosis of the subject has a learned algorithm trained in a predetermined data set by a machine learning algorithm A determination unit that outputs diagnosis support information that is information for supporting diagnosis of the subject according to a determination result of the classifier based on an input of the subject data, and a display device that displays the diagnosis support information Display means to be displayed, and receiving means for receiving a redetermination request. When the accepting unit accepts a re-determination request, the determining unit supports a new diagnosis of the subject according to the discrimination result of the discriminator based on the subject data and the feature data indicating the characteristics of the subject. Output information.

It is a block diagram which shows the structure of the image processing apparatus 1 which concerns on this embodiment. 3 is a flowchart showing an operation of the image processing apparatus 1. 4 is a flowchart showing the operation of the correction unit 20. 3 is a block diagram illustrating a configuration of a normalization unit 30. FIG. 3 is a block diagram illustrating a configuration of a first determination unit 140. FIG. 3 is a flowchart showing a bone scintigram dividing method by a normalization unit 30. It is a figure which shows an example of a bone scintigram and a skeleton part. It is an example of a health reference image set 4 is a flowchart illustrating a normalization method performed by a normalization unit 30. It is a figure which shows an example of the normalized bone scintigram set. It is a histogram which shows distribution of the count value of the bone scintigram before normalization. It is a histogram which shows distribution of the normalization count value of the bone scintigram after normalization. It is a figure which shows a scale range. It is a screen which shows a 1st display mode. It is a screen which shows a 2nd display mode.

  Hereinafter, a diagnosis support apparatus according to an embodiment of the present invention will be described with reference to the drawings. In the embodiment of the present invention described below, an image processing apparatus that generates diagnosis support information using image data of a subject as subject data will be described as an example of a diagnosis support device according to the present invention.

  FIG. 1 is a block diagram illustrating a configuration of an image processing apparatus 1 according to the present embodiment. As shown in FIG. 1, the image processing apparatus 1 includes an image processing apparatus main body 10, an input device 4 such as a keyboard, a touch panel, and a pointing device, and a display device 5 such as a liquid crystal display. The image processing apparatus main body 10 includes a correction unit 20, a normalization unit 30, a display control unit 60, and a storage unit 70. The display control unit 60 includes an adjustment unit 40, a conversion unit 50, and a generation unit 80.

  The image processing apparatus main body 10 is configured by, for example, a general-purpose computer system, and individual components or functions in the image processing apparatus main body 10 described below are realized by, for example, executing a computer program. The computer system has a memory and a microprocessor. The storage unit 70 may be realized by a memory. The computer program can be stored and distributed in a computer-readable recording medium.

  The storage unit 70 includes a captured image storage unit 15, an inspection data storage unit 17, a corrected image storage unit 25, a normalized image storage unit 35, a feature data storage unit 36, and an adjusted image storage unit 45. The captured image storage unit 15 stores a captured image that is a captured bone scintigram. The bone scintigram is an image obtained by imaging the distribution state of the radiopharmaceutical in the body of the subject. The captured image set is a pair of a captured image of a front image and a captured image of a rear image captured by a certain examination of a certain patient. In this embodiment, the value (pixel value) of each pixel in the bone scintigram is a count indicating the radiation intensity from a radioactive substance (radioisotope: radioisotope) injected into the human body before taking an image. Value. In addition, the captured image storage unit 15 stores imaging condition data including the inspection date and time, information indicating the front or rear surface, a reference visual field, and the like in association with the captured image. The examination data storage unit 17 is data indicating the characteristic amount of the subject, and for example, data obtained from an examination other than the examination using the image of the subject may be stored as examination data. The test data includes, for example, the subject's TNM classification (index of malignant tumor staging), tumor marker value, pathological test result, genetic test (risk gene expression) result, gender, age, family history, etc. Interview data may be included.

  FIG. 2 is a flowchart showing the operation of the image processing apparatus 1. This flow may be executed every time a captured image set is stored in the captured image storage unit 15 by an examination, or a plurality of captured image sets of a certain patient are stored in the captured image storage unit 15. It may be performed for examination or analysis.

  The correction unit 20 corrects the captured image stored in the captured image storage unit 15 in accordance with a predetermined standard, and stores the corrected image as a result in the corrected image storage unit 25 (S20).

  Thereafter, the normalization unit 30 detects features such as a skeletal region and a hot spot from the corrected image stored in the corrected image storage unit 25, and associates the resulting feature data with the corrected image to the feature data storage unit 36. Save to Further, the normalizing unit 30 normalizes the corrected image based on the feature data, and stores the normalized image as a result in the normalized image storage unit 35 (S30).

  Thereafter, the display control unit 60 generates a display screen by adjusting and arranging the normalized images stored in the normalized image storage unit 35, and causes the display device 5 to display the display screen (S60). End the flow.

  In the display control unit 60, the adjustment unit 40 adjusts the size of the normalized image stored in the normalized image storage unit 35 and stores it in the adjusted image storage unit 45 as a result. The conversion unit 50 converts the pixel value of the adjusted image stored in the adjusted image storage unit 45 into a gray scale density. The generation unit 80 generates a display screen by arranging the adjustment images in time series and causes the display device 5 to display the display screen.

  Note that the generation unit 80 may arrange the feature data stored in the feature data storage unit 36 in the display screen in association with the adjustment image.

  According to the image processing apparatus 1, it is possible to reduce the influence of differences in imaging conditions in a plurality of captured bone scintigrams. In addition, by displaying a plurality of bone scintigrams on the display device 5 according to a time series, it becomes easy for an image interpreter to determine temporal changes of the plurality of bone scintigrams.

  Details of the correction unit 20 will be described below.

  FIG. 3 is a flowchart showing the operation of the correction unit 20. In S20 described above, the correction unit 20 executes this flow.

  The correction unit 20 reads the captured image set from the captured image storage unit 15 and determines whether or not the read captured image set satisfies a predetermined standard (S2020).

  For example, the correction unit 20 calculates an evaluation value from the count values of all the pixels in the captured image set, and determines that the captured image set satisfies a predetermined standard if the evaluation value is within a standard range. The evaluation value is calculated based on, for example, the sum of the count values of all pixels in the previous image and the subsequent image. A more preferable evaluation value is represented by (sum of count values of all pixels of the previous image + sum of count values of all pixels of the subsequent image) / 2. The standard range is determined based on, for example, a captured image database that stores a plurality of captured image sets captured under standard imaging conditions. That is, if the evaluation value of the captured image set is within the standard range, the correcting unit 20 determines that the captured image set is captured under the standard imaging condition. The lower limit of the standard range is, for example, smaller than the minimum value of the evaluation value calculated from each captured image set in the captured image database, and the upper limit of the standard range is, for example, the maximum of the evaluation value calculated from each captured image set in the captured image database. Greater than value. In the standard range determined based on the actual captured image database, the lower limit is, for example, 1,200,000, and the upper limit is, for example, 2,200,000.

  When the captured image set satisfies the standard (S2020: Yes), the correcting unit 20 stores the captured image set in the corrected image storage unit 25 as a corrected image set without correcting (S2030), and ends this flow.

  When the captured image set does not satisfy the standard (S2020: No), the correction unit 20 calculates a scaling factor based on the evaluation value, and generates a corrected image set by multiplying each pixel of the captured image set by the scaling factor. And it preserve | saves to the correction image memory | storage part 25 (S2040), and complete | finishes this flow. The scaling factor is expressed by, for example, (reference evaluation value / evaluation value). By multiplying each pixel of the captured image set determined not to satisfy the standard by the scaling factor, the evaluation value of the modified image set as a result becomes the reference evaluation value. The reference evaluation value is a value within the standard range. The reference evaluation value determined based on the actual captured image database is 1,800,000, for example.

  The correction unit 20 reads the imaging condition data corresponding to the corrected image from the captured image storage unit 15 and stores it in the corrected image storage unit 25 in association with the corrected image.

  The correction unit 20 may determine the standard range and the reference evaluation value based on the statistic of the evaluation value calculated from each captured image set in the captured image database. For example, the correction unit 20 uses the average value and the standard deviation of the evaluation values calculated from each captured image set in the captured image database, sets (average value−standard deviation) as the lower limit of the standard range, and (average value + standard). Deviation) may be the upper limit of the standard range and (average value) may be the reference evaluation value.

  The above is the operation of the correction unit 20.

  As described above, the correction unit 20 can correct the captured image set whose count value is determined to be out of the standard so as to satisfy the standard. In other words, the count value of the captured image set that is not captured under the standard imaging condition can be corrected within the range of the count value of the captured image set that is captured under the standard imaging condition. Thereby, the accuracy of subsequent processing can be improved.

  Hereinafter, the normalization unit 30 will be described.

  The normalization unit 30 reads the corrected image set from the corrected image storage unit 25. In the following description, the corrected image may be called a bone scintigram or a patient image.

  FIG. 4 is a block diagram showing a configuration of the normalization unit 30. As shown in FIG. The normalization unit 30 includes an input image memory 105, a shape identification unit 110, an annotated image memory 115, a hot spot detection unit 120, a first hot spot memory 125, a hot spot feature extraction unit 130, a second It has a hot spot memory 135, a first determination unit 140, a third hot spot memory 145, a patient feature extraction unit 150, a patient feature memory 155, and a second determination unit 160.

  The input image memory 105 stores a bone scintigram set. The bone scintigram set obtained by one diagnosis has a front image captured from the front of the patient and a rear image captured from the back of the patient.

  The shape identifying unit 110 is connected to the input image memory 105 and identifies the anatomical structure of the skeleton included in the bone scintigram set. The shape identifying unit 110 has a predefined skeleton model. The skeletal model has at least one anatomical site (segment). Each site represents a general skeletal anatomical part. The skeletal model is adjusted to fit the skeleton of the bone scintigram set.

  The annotated image memory 115 is connected to the shape identifying unit 110 and stores an annotated image set. The annotated image set is based on information regarding a plurality of anatomical structures corresponding to a plurality of skeletal parts having different bone scintigram sets from the shape identification unit 110.

  The hot spot detection unit 120 is connected to the annotated image memory 115 and detects a high intensity region of the annotated image set. One form of hot spot detector 120 has a threshold scan that scans the bone scintigram set and identifies pixels that exceed a certain threshold. The preferred hot spot detector 120 has different thresholds for different anatomical sites defined by the shape identifier 110. In this case, the hot spot feature extraction unit 130 extracts at least one hot spot feature for each hot spot, and determines the shape and position of each hot spot.

  Another form of hot spot detection unit 120 may include an image normalization filtering unit that scans the bone scintigram set and identifies pixels that exceed a certain threshold. In this case, the hot spot feature extraction unit 130 extracts at least one hot spot feature for each hot spot, and determines the shape, texture, and geometric shape of each hot spot. Details of each feature in the preferred set of hot spot features will be described later as a specific example of the input of the first determination unit 140.

  Regarding the detected high intensity region, so-called “hot spot”, information generated as described above is stored in the first hot spot memory 125.

  The hot spot feature extraction unit 130 is connected to the first hot spot memory 125 and extracts a set of hot spot features of each hot spot detected by the hot spot detection unit 120. The extracted hot spot feature data is stored in the second hot spot memory 135.

  The first determination unit 140 determines whether or not the subject is in a predetermined state. For example, the first determination unit 140 is connected to the second hot spot memory 135 and determines whether each hot spot of the hot spot set is cancer metastasis (hereinafter, sometimes simply referred to as metastasis) ( Hot spot transfer determination). The first determination unit 140 is supplied with the feature of each hot spot of the hot spot set generated by the hot spot feature extraction unit 130. The hot spot transfer determination is based on the hot spot feature set data extracted by the hot spot feature extraction unit 130. The result of this determination is stored in the third hot spot memory 145.

  Here, the result of the hot spot transition determination may be a probability (0 to 1) indicating whether the hot spot is a transition, or a binarized value, that is, “yes, It may be a value corresponding to “is a transition” (for example, “1”) or a value corresponding to “no, a hot spot is not a transition” (for example, “0”). The determination result output by the first determination unit 140 may be output to the display device 5 as diagnosis support information for a doctor to make a diagnosis.

  The first determination unit 140 may be realized using a machine learning algorithm, for example. As the machine learning algorithm, for example, an artificial neural network (ANN), a support vector machine (SVM), a linear discriminant analysis (LDA), or the like can be used. The machine learning algorithm may be any learning method algorithm such as supervised learning, unsupervised learning, and semi-supervised learning. The 1st determination part 140 may be implement | achieved by the structure containing the learned algorithm which learned each machine learning algorithm by each method. For example, the first determination unit 140 may include a single classifier or a plurality of classifiers realized by a learned algorithm.

  The first determination unit 140 may include a region-specific determination unit 1400 for each anatomical region. Each hot spot in a certain part is sequentially processed by a part-specific determination unit 1400 configured to process the hot spot in the part.

  FIG. 5 is a block diagram illustrating an example of the configuration of the site-specific determination unit 1400 included in the first determination unit 140. The site-specific determination unit 1400 may include two or more discriminators C (C1 to C3) and one determination device D. In addition, in the example of FIG. 5, although the site | part determination part 1400 has three discriminators C, you may have two or four or more.

  Each discriminator C may output the probability of whether or not the subject is in a predetermined state, or may output whether or not the subject is in a predetermined state. For example, each discriminator C may determine whether or not a hot spot is a transition based on a predetermined input. Each discriminator C may output the probability (discrimination result) whether or not the calculated hot spot is a transition to the discriminator D. Each discriminator C may output a determination result as to whether or not the hot spot is transfer according to the probability.

  The determiner D outputs information that supports diagnosis of whether or not the subject is in a predetermined state according to the determination result of each determiner C. In the present embodiment, the determiner D determines whether or not the hot spot is a transition. The determination result of the determination device D is output as the determination result of the first determination unit 140. Since the determinator D makes a determination based on the determination results of the plurality of classifiers C, it is possible to make a determination that does not depend on the characteristics of the individual classifiers C, and the accuracy of diagnosis support is improved. Similarly, since the output of the determiner D is determined based on the outputs of the plurality of discriminators C, a stable output can be obtained without depending on the characteristics of the individual discriminators C.

  Here, each classifier C may be configured with a learned algorithm. The discriminators C1, C2, and C3 may be learned algorithms obtained by learning a common learning algorithm using different data sets. The data set may be, for example, the following first to third data sets. That is, the first data set is a data set including a lot of subject data diagnosed as normal (hot spots are not metastasized), and the second data set is subject data diagnosed as abnormal (hot spots are metastases). The data set including many and the third data set may be substantially the same number of the subject data diagnosed as normal and the subject data diagnosed as abnormal. For example, the discriminator C1 learns the first algorithm with the first data set, the learned algorithm, the discriminator C2 learns the first algorithm with the second data set, the discriminator C3 A learned algorithm obtained by learning the first algorithm using the third data set may be used. Thereby, the three discriminators C1 to C3 have different characteristics.

  Alternatively, any one of the discriminators C1, C2, and C3 may be a learning algorithm different from the other two, or all of the discriminators C1, C2, and C3 may be different learning algorithms. When learning algorithms are different, the data sets used for learning may be common. For example, the discriminator C1 shares a learned algorithm obtained by learning the first algorithm with any one data set (for example, the first data set), and the discriminator C2 uses a second algorithm different from the first algorithm in common. A learned algorithm learned from a data set or a learned algorithm obtained by further learning a third algorithm different from the first and second algorithms by the classifier C3 using a common data set may be used. Also in this case, the three discriminators C1 to C3 have different characteristics.

  The determiner D may determine whether the hot spot is a transition as follows. For example, if the output value of each discriminator C is 0 to 1, a statistical value (for example, average value, median value, etc.) based on them may be calculated and used as the output value of the discriminator D. Alternatively, the statistical value calculated by the determinator D may be further binarized to “0” indicating that there is no transfer at a predetermined threshold or “1” indicating that it is a transfer, and the output value of the determiner D may be used. . When performing these processes, the output of each discriminator C may be given a predetermined weight.

  Further, the determinator D binarizes the output values 0 to 1 of the respective discriminators C into “0” indicating that there is no transfer or “1” indicating that the transfer is performed with a predetermined threshold, The output of the determination device D may be set to “0” or “1” according to the number of “0” and “1”. At this time, the thresholds for binarizing the output values of the discriminators C may all be the same or different.

  Further, the determination device D may discretize into three or more classes, and one of the classes may be used as an output value. For example, when the determinator D divides a value of 0 to 1 into four classes with a predetermined threshold, values corresponding to “definitely normal”, “probably normal”, “probably transferred”, and “definitely transferred”, respectively. (For example, 0, 1, 2, 3) may be used as the output value.

  The first determination unit 140 may be configured with a single discriminator C. In this case, the determinator D is unnecessary, and the probability of being a hot spot output from the single discriminator C may be used as the output of the first determination unit 140 as it is. When the output of the discriminator C is a numerical value of 0 to 1, the first determination unit 140 binarizes the output value by the same process as the above-described discriminator D, or discretizes it into three or more classes. May be.

  As will be described later, when receiving a re-determination request from an interpreter or the like, the first determination unit 140 in addition to the hot spot feature set data in the second hot spot memory 135, the inspection in the inspection data storage unit 17 Data may also be accepted as input and re-determination may be performed. For example, in the case of re-determination, the determinator C of the site-specific determination unit 1400 receives input of inspection data together with hot spot feature set data, and re-determines whether the hot spot is a transfer. You may do it. At this time, the determinator D may determine whether or not the hot spot is a transition in the same manner as in a normal case. Alternatively, the site-specific determination unit 1400 may include a classifier C dedicated to examination data. The determiner D may determine whether or not the hot spot is a transition based on the output of the determiner C dedicated to inspection data, in addition to the determiner C used for normal determination.

  The patient feature extraction unit 150 is connected to the second hot spot memory 135 and the third hot spot memory 145, and the number of hot spots detected by the hot spot detection unit 120 and stored in the second hot spot memory 135. The patient feature set is extracted based on the output value of the hot spot transfer determination result from the first determination unit 140 stored in the third hot spot memory 145. The patient feature extraction unit 150 performs calculation using both the data from the hot spot feature extraction unit 130 and the data output from the first determination unit 140. The extracted patient feature data is stored in the patient feature memory 155. A suitable extracted feature will be described later as a specific example of the input of the second determination unit 160.

  The second determination unit 160 determines whether or not the subject is in a predetermined state. For example, the second determination unit 160 is connected to the patient feature memory 155, and based on the patient feature set data extracted by the patient feature extraction unit 150 and stored in the patient feature memory 155, at least one patient is identified. It is determined whether or not the patient has cancer metastasis (patient metastasis determination).

  The second determination unit 160 may have the same configuration as the first determination unit 140 or may be a part of the first determination unit 140.

  Hereinafter, a bone scintigram dividing method (segmentation) by the normalization unit 30 will be described.

  FIG. 6 is a flowchart showing a bone scintigram dividing method by the normalization unit 30. In S30 described above, the normalization unit 30 executes this flow.

  The bone scintigram division method uses, for example, an ASM (active shape model) approach, and the entire outline of the front and back images of the skeleton excluding the arms (upper limbs) and the lower parts (ends) of the legs (lower limbs). Includes performing extraction. Omitting these parts of the skeleton is not a problem because the lower arms and legs are very rare places for metastases and they are sometimes not acquired in the bone scanning routine. For illustration purposes, ASM is defined as a statistical method for finding objects in an image. The method is based on a statistical model based on a learning image set. In each learning image, the shape of the object is defined by target points that are determined manually by a human operator during the preparation or learning phase. A point distribution model is then calculated that is used to describe the general shape for the object. The general shape can be used to search for other images for a new example of an object type, such as a skeleton, as in the present embodiment. A method is provided for learning ASM that describes the anatomy of the human skeleton. The model includes the steps described below.

The first step, for example, divides the skeleton into 8 different learning sets (S205). The learning set is selected to correspond to an anatomical site that is particularly suitable for achieving consistent segmentation. Eight different learning sets are shown below.
1) The first learning set for the frontal image of the head and spine.
2) A second learning set for the frontal image of the ribs.
3) A third learning set for the frontal image of the arm.
4) A fourth learning set for the front image of the lower body.
5) A fifth learning set for the posterior image of the head and spine.
6) A sixth learning set for the posterior image of the ribs.
7) A seventh learning set related to the rear image of the arm.
8) The 8th learning set regarding the rear image of the lower body.

  Each learning set provides many case images (S210). Each image is created using a set of target points. Each target point is associated with a coordinate pair that represents a particular anatomical or biological point of the image. A coordinate pair is determined by manually identifying the corresponding anatomical / biological point (S215). In the front view, the following easily identifiable anatomical target points are used:

  Prior to capturing the learning set statistics, each set of target points is aligned to a different common coordinate system for each of the eight learning sets. This is realized by scaling, rotation, and interpretation of learning shapes that make the learning shapes as close as possible to each other, as described in Non-Patent Document 1. By examining the statistics of the learning set, a two-dimensional statistical point distribution model including the shape change observed in the learning set is obtained (S220). The statistical modeling of the change of the target point (shape) across the plurality of skeletons is performed as described in Non-Patent Document 1 and Non-Patent Document 2.

  The statistical model resulting from the shape change can be applied to the patient image to segment the skeleton. Starting with an average shape, new shapes within the range of possible changes in the shape model can be generated similar to the shape of the learning set so that the generated skeleton resembles the structure present in the patient image. . The frontal body segment divided using this method is, for example, skull, face neck, spine, upper sternum, lower sternum, right arm, left arm, right rib, left rib, right shoulder, left shoulder, pelvis, bladder, right femur, And the left femur. The posterior segment includes, for example, the skull, neck, upper spine, lower spine, spine, right arm, left arm, right rib, left rib, right shoulder blade, left shoulder blade, hip joint, lower pelvis, bladder, right femur, and left femur It may be.

  The first step in the search process may be to find a starting position for the average shape of the front image. For example, the highest point of the head may be selected. This is because it proves to be a robust starting position in the test and can be easily found by examining the intensity above a certain threshold in each horizontal line within the top of the image.

  Next, the search of the example of the skeleton shape model matching the skeleton in the patient image is executed according to the algorithm described in Non-Patent Document 1 and Non-Patent Document 2.

  Through the above operation, the normalization unit 30 acquires feature data including the region of the skeletal region identified from the bone scintigram set, the detected hot spot region, the hot spot metastasis determination result, the patient metastasis determination result, and the like. The feature data is stored in the feature data storage unit 36. Here, the output value of each discriminator C that is the basis for calculating the hot spot transfer determination result by the determining device D of the first determination unit 140 is also associated with the hot spot transfer determination result in association with the feature data storage unit 36. It may be stored in. Similarly, the output value of each discriminator C, which is the basis for calculating the patient metastasis determination result by the determination unit D of the second determination unit 160, is also stored in the feature data storage unit 36 in association with the patient metastasis determination result. You may be made to do.

  Hereinafter, another bone scintigram division method by the normalization unit 30 will be described.

  Another bone scintigram segmentation method is provided. The goal of another bone scintigram segmentation method is to recognize and extract contours from a plurality of different anatomical parts of the skeleton in the bone scintigram 300, similar to the bone scintigram segmentation method described above. FIG. 7 is a diagram showing an example of a bone scintigram and a skeleton part. As shown in this figure, the skeletal site is defined by overlaying a contour 320 on the patient image 310. The division method described here is a division represented by registration.

  The image registration method converts an image into a coordinate system of another image. Assume that these images represent entities of the same object class, here a skeleton. The transformed image is represented by the source image, while the untransformed image is represented by the target image. If the same image coordinates match the same geometric / anatomical position in the objects contained in the source and target images, it can be said that the coordinate systems of the source and target images match. The division by registration is to use a manually defined division for the source image and to register the source image to the target image when no division is defined. The source image segment is thereby converted to the target image, thus producing a target image segment.

  The segmentation of the source image in this form defines the reference healthy patient's anatomy and is manually drawn by the laboratory specialist as a set of polygons. FIG. 8 is an example of a health reference image set. The health reference image set is called “Atlas”. Here, the health reference image set includes a health reference image 400a and a health reference image 400b. The health reference image 400a shows a front image or a front image of the patient, while the health reference image 400b shows a back image or a rear image of the patient. Referring to the labels in the health reference image 400a and health reference image 400b, respectively, these regions were labeled with the front and rear skulls 401, 402, (2, 2) labeled (1,1). Labeled with anterior and posterior cervical vertebrae 403, 404, (3, 3) and anterior and posterior thoracic vertebrae 405, 406, (14) and anterior sternum 407, (4, 4). Anterior and posterior pelvis labeled with anterior and posterior lumbar vertebrae 409, 408, (11, 5) and anterior and posterior sacrum 411, 410, (15, 14), (5, 6, 7,6) labeled anterior and posterior left and right shoulder blades, (17,16) labeled anterior and posterior left and right clavicles, anterior labeled (7,8,9,8) Part and rear Left and right humerus, (9,10,11,10) labeled anterior and posterior left and right ribs and (12,13,12,13) labeled anterior and posterior left and right femurs 413 415, 412, and 414 are defined.

  The health reference image set is always used as a source image by the normalization unit 30, while the patient image to be examined serves as a target image. The result is a segmentation of the target image into skeletal sites as depicted in the bone scintigram 300. Lower arms and lower limbs are not considered for analysis.

  The health reference image set used as the source image consists of 10 real cases of healthy patients with typical image quality and normal appearance and anatomy. An algorithm is used to generate an anterior and posterior image of a fictitious normal healthy patient having an average intensity and anatomy calculated from a group of case images. The normalization unit 30 performs this task as described in Non-Patent Document 3. The health reference image set shows the result when the resulting anatomy does indeed appear to have a normal health appearance. This anatomy is the result of averaging the anatomy of several patients, indicating high lateral symmetry.

  Hereinafter, the normalization method of the bone scintigram by the normalization part 30 is demonstrated.

  As described above, the hot spot detection unit 120 uses information from the shape identification unit 110. The hot spot detection unit 120 has two purposes. Its primary purpose is to separate hot spots in the front and back patient images. A hot spot is a high-intensity separated image site and may be an indicator of metastatic disease when in the skeleton. Its second purpose is to adjust the intensity of the image to a predefined reference level. Such intensity adjustment represents normalization of the image. Here, the algorithm will be described. This algorithm simultaneously isolates hot spots and estimates normalization factors. This algorithm is also executed separately for the front and rear images. In the following, the need for proper normalization will be briefly described first, followed by a description of this algorithm.

  Multiple bone scintigrams vary greatly in intensity levels depending on patient, study, and hardware configuration. This difference is presumed to be increasing and zero intensity is presumed to be a common reference level for all images. Thus, normalizing the source image with respect to the target image is to find a scalar factor that makes the intensity of the source image equal to the target image. Here, if the average intensity of the healthy part of the skeleton in the source image is equal to the corresponding part in the target image, the intensity of the two skeleton images is defined as equal. FIG. 9 is a flowchart showing a normalization method performed by the normalization unit 30. The normalization method shown in this flowchart has the following steps.

1. Identification of image elements corresponding to the skeleton (S510).
2. Identification of hot spots included in the image (S520).
3. The hot spot element is removed from the skeleton element (S530).
4). Calculation of the average intensity of the remaining (health) elements (S540).
5. Calculation of an appropriate normalization factor (S550).
6). Adjusting the source image intensity by multiplying by a normalization factor (S560).

  As described above, S510 is executed using the information on the image part belonging to the skeleton provided by the converted anatomical part obtained by the shape identifying unit 110. Polygonal parts are converted into binary image masks that define image elements belonging to the respective parts of the skeleton.

  In S520, the hot spot is divided using one image filtering operation and one threshold operation. Images are filtered using a DoG (difference-of-Gaussians) -BPF (band-pass filter) that emphasizes small areas of high intensity with respect to their surroundings. The filtered image is then thresholded at a certain level and the resulting binary image defines hot spot elements.

  In S530, among the elements calculated in S510, elements that match the hot spot element calculated in S520 are deleted. The remaining elements are assumed to correspond to healthy skeletal sites.

  In S540, the average strength of the health skeleton element is calculated. This average intensity is represented by A.

  In S550, an appropriate normalization factor is determined in relation to a predefined reference intensity level. Here, for example, this level can be set to 1000. The normalization factor B is calculated as (B = 1000 / A).

  In S560, the intensity of the source image is adjusted by multiplying B.

  Hot spot segmentation in S520 depends on the overall intensity level of the image, which is sequentially determined by the normalization factor calculated in S550. Nonetheless, the normalization factor calculated in S550 depends on the hot spot split from S520. Since the results of S520 and S550 are interdependent, S520 to S560 may be repeated, for example, until no further change in normalization factor occurs in S570. Extensive testing has shown that this process usually converges in 3 or 4 iterations.

  FIG. 10 is a diagram illustrating an example of a normalized bone scintigram set. The normalized image 600a shows a front image of the normalized bone scintigram set, and the normalized image 600b shows a rear image of the normalized bone scintigram set. The divided hot spots 620 are shown in the normalized image 600 a and the normalized image 600 b as dark spots that appear in the divided image 610. Therefore, normalizer 30 will classify this patient as having cancer metastasis.

  The normalization unit 30 stores the result of normalization of the corrected image by the above-described normalization method in the normalized image storage unit 35 as a normalized image.

  The term “point” in the above description may be used to represent at least one pixel in the image.

  Hereinafter, a specific example of input to the first determination unit 140 will be described.

The first determination unit 140 is provided with the following 27 features for measuring the size, shape, direction, localization, and intensity distribution of each hot spot.
-Skeletal lesions. A measure of the skeletal volume occupied by the extracted hot spot site based on the 2D hot spot region, the 2D region of the corresponding skeletal region, and a coefficient representing the volume fraction represented by the skeletal site for the entire skeleton. . Calculated as (hot spot area / local area * coefficient).
• Relative area. Hot spot area for the corresponding skeletal site. Scale independent of image resolution and scanner field of view.
-Relative barycentric position (2 functions). Relative barycentric position relative to the bounding box of the corresponding skeletal part. Values range from 0 (upper, left) to 1 (lower, right).
• Relative center of mass (2 functions). Similar to the function of the center of gravity, but takes into account the intensity of the hot spot site when calculating the x and y values.
-Relative height. The height of the hot spot relative to the height of the corresponding skeletal site.
-Relative width. The width of the hot spot relative to the width of the corresponding skeletal site.
・ Minimum strength. The minimum intensity calculated from all hot spot elements on the corresponding normalized image.
・ Maximum strength. Maximum intensity calculated from all hot spot elements on the corresponding normalized image.
・ Total strength. The sum of the intensities calculated from all hot spot elements on the corresponding normalized image.
・ Average strength. The average intensity calculated from all hot spot elements on the corresponding normalized image.
・ Standard deviation of intensity. Standard deviation of the intensity calculated from all hot spot elements on the corresponding normalized image.
・ Boundary length. The length of the hotspot boundary measured in pixels.
・ Consistency. The percentage of the convex hull area of the hot spot represented by the hot spot area.
・ Eccentricity. The elongation of the hot spot represented by 0 (circle) to 1 (line).
・ Total number of hot spot counts. Sum of intensities in all hot spots across the skeleton.
• Number of hot spot counts at the site. Sum of hot spot intensities contained in the skeleton corresponding to the current hot spot.
-Total hot spot range. The area of all hot spots in the entire skeletal site relative to the entire skeletal site in the corresponding image.
-Hot spot range at the site. The area of all hot spots in the skeletal site corresponding to the current hot spot for the area of the skeletal site.
・ Total number of hot spots. Number of hot spots across the skeleton.
-The number of hot spots at the site. The number of hot spots in the skeletal site corresponding to the current hot spot.
Hot spot locality (2 functions). The X coordinate ranges from 0 (innermost) to 1 (most distal) with respect to the midline calculated from the reference anatomical structure converted in the step of the shape identification unit 110. The Y coordinate ranges from 0 (best) to 1 (worst). All measures are relative to the corresponding skeletal site.
• Distance asymmetry. Minimum Euclidean distance between the relative center of mass of the current hot spot and the mirror image of the mass of the hot spot in the contralateral skeletal region. It is calculated only for skeleton regions that naturally have corresponding contralateral skeleton sites.
• Range asymmetry. The minimum difference in range between the current hot spot and the range of hot spots in the contralateral skeletal site.
-Strength asymmetry. The minimum difference in intensity between the current hot spot and the intensity of the hot spot in the contralateral skeletal region.

  The first determination unit 140 can be further provided with inspection data obtained from an inspection other than the inspection using the image of the subject.

  Hereinafter, a specific example of input to the second determination unit 160 will be described.

The second determination unit 160 that determines a patient level diagnosis related to the presence or absence of a metastatic disease uses the 34 features listed below as an input. All the functions used by the second determination unit 160 are calculated from hot spots classified as having a high transition probability by the first determination unit 140.
・ Total lesions. Sum of skeletal lesions across the skeleton.
-Skull lesions. Sum of skeletal lesions within the skull region.
-Cervical spine lesions. Sum of skeletal lesions in the cervical spine column.
-Thoracic column lesions. Sum of skeletal lesions in the thoracic column.
・ Lumbar spine lesions. Sum of skeletal lesions in the lumbar region.
-Upper limb lesions. Sum of skeletal lesions in the upper limbs.
-Lower limb lesions. Sum of skeletal lesions in the lower limbs.
-Chest lesions. Sum of skeletal lesions in the chest region.
・ Pelvic lesions. Sum of skeletal lesions in the pelvic region.
• Total number of “high” hotspots.
• The number of “high” hot spots in the skull area.
• Number of “high” hot spots in the cervical spine region.
• Number of “high” hot spots in the chest column area.
• Number of “high” hot spots in the lumbar column area.
• Number of “high” hot spots in the upper limb region.
• Number of “high” hot spots in the lower limb area.
• Number of “high” hot spots in the chest area.
• Number of “high” hot spots in the pelvic region.
The maximum value calculated by the first determination unit 140 at the skull site.
The maximum value calculated by the first determination unit 140 at the cervical vertebra site.
The maximum value calculated by the first determination unit 140 at the thoracic vertebra site.
The maximum value calculated by the first determination unit 140 at the lumbar region.
The maximum value calculated by the first determination unit 140 at the sacral region.
The maximum value calculated by the first determination unit 140 at the humerus part.
The maximum value calculated by the first determination unit 140 at the clavicle site.
The maximum value calculated by the first determination unit 140 at the scapula site.
The maximum value calculated by the first determination unit 140 in the thigh region.
The maximum value calculated by the first determination unit 140 at the sternum site.
The maximum value calculated by the first determination unit 140 at the rib site.
-The second largest value calculated by the first determination unit 140 in the rib part.
The third largest value calculated by the first determination unit 140 at the rib site.
The maximum value calculated by the first determination unit 140 in the pelvic region.
-The second largest value calculated by the first determination unit 140 in the pelvic region.
The third largest value calculated by the first determination unit 140 in the pelvic region.

  The second determination unit 160 can be further provided with inspection data obtained from an inspection other than the inspection using the image of the subject.

  As described above, the shape identifying unit 110 can identify a skeletal part in a bone scintigram regardless of the patient's body shape by deforming a predefined skeleton model according to the bone scintigram. it can. Further, the hot spot detection unit 120 can detect abnormal accumulation by detecting a hot spot that is a region having a count value higher than a predetermined threshold value from the identified skeleton site. The first determination unit 140 can determine cancer metastasis for each hot spot by calculating the probability of cancer metastasis for each hot spot. The second determination unit 160 can determine whether or not the patient has cancer metastasis, for example, by calculating the probability of cancer metastasis of the patient.

  In the following description, the normalized count value is referred to as a normalized count value. FIG. 11 is a histogram showing the distribution of count values of the bone scintigram before normalization. The distribution of count values of normal bone regions (normal regions) that are healthy skeletal elements differs between the bone scintigram 3010 and the bone scintigram 3020 before normalization. That is, the average count value 3011 of the normal bone region of the bone scintigram 3010 is different from the average count value 3021 of the normal bone region of the bone scintigram 3020. FIG. 12 is a histogram showing the distribution of normalized count values of the normalized bone scintigram. A bone scintigram 3030 with normalized bone scintigram 3010 and a bone scintigram 3040 with normalized bone scintigram 3020 are shown. In the bone scintigram 3030 and the bone scintigram 3040, the distribution of normalized count values of normal bone regions overlap. Further, the average count values 3011 and 3021 of the normal bone region are converted into 1000 in the normalized count value by normalization.

  The hot spot detection unit 120 reads out the imaging condition data corresponding to the normalized image from the modified image storage unit 25 and stores it in the normalized image storage unit 35 in association with the normalized image.

  The hot spot detection unit 120 extracts a healthy skeletal element from which a hot spot is removed from a skeletal part, and normalizes the bone scintigram with an average count value of the healthy skeletal element, thereby not affecting the hot spot. The bone scintigram concentration can be made uniform.

  The normalization unit 30 determines the transfer probability of the hot spot, which is the output of the first determination unit 140, using a preset threshold value, classifies the hot spot into a high risk part and a low risk part, and determines the classification result. It may be included in the feature data. The normalization unit 30 may calculate a BSI (bone scan index) that is a ratio of the area of the high-risk part to the area of all the skeletal parts, and include the calculation result in the feature data. In addition, the normalization unit 30 may calculate the number of high-risk parts and include the calculation result in the feature data.

  The normalizing unit 30 may generate a normalized image using another normalization method.

  Hereinafter, the adjustment unit 40 will be described.

  The size of the region of the patient in the captured image varies depending on the detector (gamma camera) to be imaged and the acquisition conditions. The pixel size that is the actual size of the region represented by each pixel in the captured image is represented by (reference field of view / matrix size / magnification ratio). The reference field is the length of one side of the detector field square. The matrix size is the number of pixels corresponding to one side of the reference visual field in the matrix constituting the captured image. The magnification in bone scintigram is usually 1. The imaging condition data includes a pixel size, a reference visual field, and a matrix size. When imaging multiple bone scintigrams, the size of the patient area differs depending on the detector (gamma camera) and acquisition conditions, allowing the reader to determine the actual size of the region between the multiple bone scintigrams displayed. It becomes difficult to compare.

  The adjustment unit 40 reads the normalized image and the imaging condition data corresponding to the normalized image from the normalized image storage unit 35. Thereafter, the adjustment unit 40 performs, for example, enlargement or reduction of the pixel size and change of the matrix size so that the normalized image becomes a specific reference visual field based on the imaging condition data corresponding to the normalized image. An adjusted image is generated, and the adjusted image is stored in the adjusted image storage unit 45. Thereby, the reference visual field of the adjusted image becomes constant.

  The adjustment unit 40 reads out the imaging condition data corresponding to the adjustment image from the corrected image storage unit 25 and stores it in the adjustment image storage unit 45 in association with the adjustment image.

  The specific reference field may be the latest reference field among the reference fields of a plurality of bone scintigrams. The specific reference visual field may be the largest reference visual field among the reference visual fields of a plurality of bone scintigrams. Further, the specific reference visual field may be a preset reference visual field.

  According to the adjustment unit 40, the reference field of view of the plurality of bone scintigrams becomes constant, so that the image interpreter can compare the actual sizes of the regions in the plurality of displayed bone scintigrams.

  Hereinafter, the conversion unit 50 will be described.

  According to the normalization method described above, by displaying the adjusted image using the normalized count value as the density, the density ranges of the pixels in the normal bone region are made uniform. However, even if the adjustment image is displayed as it is, it is not necessarily a density scale suitable for recognition of a normal bone region and a hot spot by an image interpreter. For example, if the normalized count value corresponding to the maximum density is too high, most of the normal bone region is displayed in white, and the image interpreter may not be able to recognize the normal bone region in the displayed bone scintigram.

  Therefore, when displaying the adjusted image, the conversion unit 50 converts the normalized count value of the adjusted image into an appropriate gray scale density.

  The conversion unit 50 determines the scale range so that the specific saturation reference value is converted to the maximum gray scale density. The saturation reference value is a value obtained by multiplying a specific maximum reference value by a predetermined range coefficient. For example, the maximum reference value is determined based on a normalized count value obtained by normalizing a bone scintigram in the above-described captured image database by the above-described normalization method. The maximum reference value is, for example, the same or larger than the maximum normalized count value in the captured image database. The range coefficient is greater than 0 and less than 1. A more preferable saturation reference value is larger than 1000 and smaller than the maximum normalized count value in the captured image database.

  FIG. 13 is a diagram illustrating the scale range. In this figure, the horizontal axis indicates the normalized count value of the pixel of the adjusted image, and the vertical axis indicates the density of the pixel of the adjusted image. Thereby, the density of the pixel whose normalized count value is zero is displayed as zero density. Pixels whose normalized count value is greater than or equal to the saturation reference value are displayed at the maximum density. Pixels having a normalized count value greater than zero and less than the saturation reference value are displayed with a gradation corresponding to the normalized count value, for example, a curve gradation such as a linear gradation or a log gradation.

  According to the conversion unit 50, by interpreting the normalized count value of the adjusted image into an appropriate gray scale density, the image interpreter can easily recognize the normal bone region and the hot spot in the displayed bone scintigram. Can do. Note that the conversion unit 50 may adjust the range coefficient or the saturation reference value based on the operation of the input device 4 by the interpreter.

  Hereinafter, the generation unit 80 will be described.

  The generation unit 80 identifies a patient to be displayed and a plurality of examination dates based on the operation of the input device 4 by a radiogram interpreter. The adjustment image corresponding to the specified patient and the examination date is selected from the plurality of adjustment images in the adjustment image storage unit 45. The production | generation part 80 produces | generates the display screen for displaying on the display apparatus 5, and arrange | positions the selected adjustment image in a display screen according to a time series.

  FIG. 14 is a screen showing the first display mode. The display screen according to the first display mode includes a front image area 8100 that is an area for arranging a front image, and a rear image area 8200 for arranging a rear image. In this example, it is assumed that three adjustment image sets are selected. The generation unit 80 arranges the plurality of front images 8511, 8521, and 8531 in the selected plurality of adjustment image sets in the horizontal direction in the front image region 8100 in the order of the inspection date and time, and sets the plurality of images in the selected adjustment image set. The rear images 8512, 8522, and 8532 are arranged in the horizontal direction in the rear image area 8200 in the order of the inspection date. That is, the newer front image is placed more to the right in the front image area 8100, that is, the newer rear image is placed to the right in the rear image area 8200. At this time, the generation unit 80 aligns the horizontal center of the plurality of front images 8511, 8521, and 8531 aligned with the horizontal center 8101 in the front image region 8100, and arranges a plurality of rear images 8512, 8522 arranged. , 8532 in the horizontal direction is aligned with the horizontal center 8201 in the rear image area 8200. The front image area 8100 and the rear image area 8200 may be arranged on the left and right, or may be arranged on the top and bottom. The generation unit 80 may attach an examination date to each of the front image and the rear image in the display screen based on the imaging condition data associated with the adjustment image set. Further, the generation unit 80 generates a hot spot transfer determination result, a patient transfer determination result, a BSI, a high risk for each of the front image and the rear image in the display screen based on the feature data associated with the adjustment image set. You may attach | subject the number of parts.

  According to the first display mode, the image reader can easily change the displayed bone scintigram over time by arranging each of the front image and the rear image in which the influence due to the difference in imaging conditions is reduced in time series. Can grasp.

  FIG. 15 is a screen showing the second display mode. The display screen according to the second display mode includes a pair area 8300 that is an area for arranging an adjustment image set that is a pair of a front image and a rear image. The generation unit 80 arranges the selected adjustment image sets 8610, 8620, 8630 in the horizontal direction in the area 8300 in the order of examination date and time. That is, the newer adjusted image set is arranged more right in the pair area 8300. At this time, the generation unit 80 matches the horizontal center of the plurality of adjusted image sets 8610, 8620, and 8630 arranged with the horizontal center 8301 in the pair region 8300. Here, the adjustment image set 8610 includes a front image 8511 and a rear image 8512, the adjustment image set 8620 includes a front image 8521 and a rear image 8522, and the adjustment image set 8630 includes a front image 8531 and a rear image 8532. Have The generation unit 80 may assign an examination date to each of the adjustment image sets in the display screen based on the imaging condition data associated with the adjustment image set. Further, the generation unit 80 generates a hot spot metastasis determination result, a patient metastasis determination result, a BSI, and the number of high risk sites for each of the adjustment image sets in the display screen based on the feature data associated with the adjustment image set. Etc. may be attached.

  The generation unit 80 may be configured to accept a re-determination request for determination of hot spot metastasis and patient metastasis on the display screen of FIG. 14 or FIG. 15. For example, the generation unit 80 may display a predetermined button on the display screen as a redetermination request accepting unit. For example, when the display screen accepts a redetermination request in response to the operation of the input device 4 of the radiogram interpreter, the normalization unit 30 may execute the hot spot transfer determination and / or the patient transfer determination again. .

For example, when the metastasis probability of a hot spot or the metastasis probability of a patient is an intermediate value (for example, in the range of 0.4 to 0.6), it is useful as information that supports diagnosis of whether or not it is metastasis. I can not say. In such a case, a new feature amount for the subject may be accepted as additional data and re-determination may be performed. This is common in both cases where the first determination unit 140 and the second determination unit 160 are configured by a single discriminator C and when a plurality of discriminators C are included. Preferably, inspection data may be used as a new feature amount. This is because examination data other than bone scintigrams may greatly affect the results of metastasis determination in determining metastasis of hot spots and / or metastasis of patients.
When redetermining, the hot spot metastasis probability and / or patient metastasis probability before and after the redetermination may be displayed in the display screen. This is because whether the hot spot metastasis probability or the patient metastasis probability has increased or decreased before and after the re-determination can be important information in support of diagnosis. For example, even when the metastasis probability of a hot spot after re-determination or the metastasis probability of a patient is the same 0.5, when the metastasis probability before re-determination increases to 0.5, it decreases to 0 The meaning is different from the case of .5. Therefore, it can be avoided that the radiogram reader easily makes a judgment by looking only at the transfer probability after the re-determination.

  On the other hand, when the first determination unit 140 includes a plurality of discriminators C, the generation unit 80 further displays the output value of the individual discriminator C associated with the hot spot transfer determination result in the feature data. It may be displayed in the screen. Similarly, when the second determination unit 160 includes a plurality of discriminators C, the generation unit 80 displays the output value of the individual discriminator C associated with the patient metastasis determination result on the display screen. Also good.

  At this time, the output value of the individual discriminator C is considered to indicate the reliability of each determination result. In other words, even if the transfer probability of the hot spot or the transfer probability of the patient is not an intermediate value, for example, when the output value of the individual discriminator C is different even when it is 0.2 or less or 0.8 or more, the probability The reliability of may vary. For example, even when the patient's metastasis probability is the same “0.8”, the outputs of the three discriminators C are “1.0”, “0.7”, “0.4”, and It is considered that the reliability of the transfer probability “0.8” is different from the cases of “0.8”, “0.8”, and “0.8”. For this reason, when it is considered necessary when the image interpreter sees the transfer probability and the output of the discriminator C together, an instruction for a redetermination request can be accepted.

  The reliability calculated from the output value of the individual discriminator C may be displayed. For example, in the above example, the reliability may be determined by variations in the output values of the three discriminators C.

  Furthermore, the generation unit 80 can display the reliability by another method regardless of whether the first determination unit 140 and the second determination unit 160 are single or plural. . The output value of the first determination unit 140 or the output of the second determination unit 160, which is the probability of whether or not the subject is in a predetermined state, is directly related to the reliability. In particular, when the metastasis probability of the hot spot or the metastasis probability of the patient is an intermediate value, the reliability is considered to be low, and a display indicating that the reliability is low may be displayed together with the result of the metastasis determination. For example, the reliability display may be divided into a plurality of classes such as low reliability, medium reliability, and high reliability, and one of the classes may be displayed. The class can be arbitrarily divided by a predetermined threshold according to the output values of 0 to 1 of the discriminator C or the discriminator D, and may be set in advance according to the purpose of inspection.

  The generation unit 80 may also display the reliability based on the hot spot feature set data that is an input to the first determination unit 140 of the subject and the patient feature set data that is an input to the second determination unit 160. it can. For example, in one feature of the hot spot feature set data input to the first determination unit 140, a feature range for outputting a determination result with a predetermined accuracy or higher can be obtained in advance. In addition, in one feature of the patient feature set data input to the second determination unit 160, a feature range for outputting a determination result with a predetermined accuracy or higher can be obtained in advance. That is, if the characteristics of the subject to be input are within a predetermined characteristic range, the determination result calculated based on this characteristic is considered to have a predetermined accuracy or higher, and this can be used as the reliability. For example, the reliability may be displayed according to the number of each feature in the input hot spot feature set data or patient feature set data of the subject to be input within a predetermined feature range. In this example, the greater the number, the higher the reliability and the lower the reliability. Further, for example, the reliability may be displayed depending on whether or not a particular feature is within a predetermined feature range among the features in the input subject's hot spot feature set data or patient feature set data. Good. In this example, if the specific feature is within the predetermined feature range, the reliability is high, and if the specific feature is not within the predetermined feature range, the reliability is low. Here, the specific feature is a variation factor having a large contribution to the determination result, and whether or not the specific feature is within the range of the predetermined feature in the determination result is strongly related.

  According to the second display mode, by arranging a plurality of pairs of the front image and the rear image in which the influence due to the difference in the photographing conditions is reduced according to the time series, the image reader can change the displayed bone scintigram over time. It can be easily grasped.

  The generation unit 80 may select one of the first and second display modes according to the operation of the input device 4 by the image interpreter.

  The generation unit 80 may superimpose the outline of the skeletal part indicated in the feature data on the adjustment image. Further, the generation unit 80 may add different colors to the low risk part and the high risk part in the adjusted image.

  The generation unit 80 attaches feature data such as the possibility of transfer, BSI, and the number of high-risk parts to each adjustment image in the display screen, so that the radiogram interpreter can know the temporal change of the feature data.

  The image processing apparatus 1 may include an analysis unit that analyzes the adjusted image. In this case, the analysis unit may compare specific skeleton parts in the adjusted images obtained by the plurality of examinations using the feature data in the feature data storage unit 36. Further, the analysis unit may use the feature data in the feature data storage unit 36 to associate time series of specific hot spots in the adjusted images obtained by a plurality of examinations. By analyzing the plurality of adjusted images in which the influence due to the difference in the imaging conditions is reduced by the analysis unit, it is possible to improve the accuracy of the analysis as compared with the case of analyzing the plurality of captured images.

  The image processing apparatus 1 may have the above-described captured image database. The image processing apparatus 1 may register a captured image satisfying a predetermined condition among the captured images stored in the captured image storage unit 10 in the captured image database.

  The above-described embodiments of the present invention are examples for explaining the present invention, and are not intended to limit the scope of the present invention only to those embodiments. Those skilled in the art can implement the present invention in various other modes without departing from the gist of the present invention. For example, in the above-described embodiment, the target is a bone scintigram, but other images such as CT, MRI, SPECT, PET, and X-rays (X-rays) may be used.

1: Image processing device 4: Input device 5: Display device 10: Image processing device body 15: Captured image storage unit 17: Inspection data storage unit 20: Correction unit 25: Correction image storage unit 30: Normalization unit 35: Normalization Image storage unit 36: Feature data storage unit 40: Adjustment unit 45: Adjustment image storage unit 50: Conversion unit 60: Display control unit 70: Storage unit 80: Generation unit

Claims (13)

A computer program for a diagnosis support apparatus that accepts subject data that is data relating to a subject and supports diagnosis of the subject,
When the computer program is executed by the diagnosis support device,
A first discriminator including a learned algorithm obtained by learning a first machine learning algorithm with a first data set, wherein the first discriminator receives an input of the subject data and outputs a first discrimination result. Discriminator and
A second discriminator including a learned algorithm obtained by learning the first machine learning algorithm with a second data set different from the first data set; The second discriminator for outputting the discrimination result of 2;
A computer program for realizing, in the diagnosis support apparatus, a determination device that outputs diagnosis support information that is information for supporting diagnosis of the subject according to the first determination result and the second determination result. A computer program for a diagnosis support apparatus that accepts subject data that is data relating to a subject and supports diagnosis of the subject,
When the computer program is executed by the diagnosis support device,
A first discriminator including a learned algorithm obtained by learning a first machine learning algorithm with a first data set, wherein the first discriminator receives an input of the subject data and outputs a first discrimination result. Discriminator and
A first algorithm including a learned algorithm obtained by learning a second machine learning algorithm different from the first machine learning algorithm from the first data set or a second data set different from the first data set; A second discriminator that receives the input of the subject data and outputs a second discrimination result;
A computer program for realizing, in the diagnosis support apparatus, a determination device that outputs diagnosis support information that is information for supporting diagnosis of the subject according to the first determination result and the second determination result. The first and second discrimination results are probabilities as to whether or not the subject is in a predetermined state,
The diagnosis support information is information for supporting diagnosis, which is a probability of whether or not the subject is in the predetermined state or whether or not the subject is in the predetermined state. Computer program. The first and second determination results are whether or not the subject is in a predetermined state,
The computer program according to claim 1, wherein the diagnosis support information is information that supports diagnosis of whether or not the subject is in the predetermined state.   The computer program according to claim 1, further causing the diagnosis support apparatus to realize means for causing the display apparatus to display the diagnosis support information and the first and second determination results.   The computer program according to any one of claims 1 to 4, further causing the diagnosis support apparatus to realize means for displaying the diagnosis support information and a reliability of the diagnosis support information on a display device.   The computer program according to claim 1, wherein the first machine learning algorithm is an artificial neural network.   The computer program according to claim 1, wherein the subject data is data obtained from a bone scintigram. The diagnostic support apparatus further realizes an accepting means for accepting a re-determination request,
When the accepting unit accepts the re-determination request, the first discriminator and the second discriminator are updated based on the subject data and the feature data indicating the subject characteristics. The computer program according to claim 1, wherein the determination result is output, and the determination device outputs new diagnosis support information. A diagnosis support apparatus that accepts subject data that is data relating to a subject and supports diagnosis of the subject,
A first discriminator including a learned algorithm obtained by learning a first machine learning algorithm with a first data set, wherein the first discriminator receives an input of the subject data and outputs a first discrimination result. Discriminator and
A second discriminator including a learned algorithm obtained by learning the first machine learning algorithm with a second data set different from the first data set; The second discriminator for outputting the discrimination result of 2;
A diagnosis support apparatus comprising: a determination device that outputs diagnosis support information that is information for supporting diagnosis of the subject according to the first determination result and the second determination result. A diagnosis support apparatus that accepts subject data that is data relating to a subject and supports diagnosis of the subject,
A first discriminator including a learned algorithm obtained by learning a first machine learning algorithm with a first data set, wherein the first discriminator receives an input of the subject data and outputs a first discrimination result. Discriminator and
A first algorithm including a learned algorithm obtained by learning a second machine learning algorithm different from the first machine learning algorithm from the first data set or a second data set different from the first data set; A second discriminator that receives the input of the subject data and outputs a second discrimination result;
A diagnosis support apparatus comprising: a determination device that outputs diagnosis support information that is information for supporting diagnosis of the subject according to the first determination result and the second determination result. A method performed by a diagnosis support apparatus that accepts subject data that is data related to a subject and supports diagnosis of the subject,
The diagnosis support apparatus includes:
A first discriminator including a learned algorithm obtained by learning a first machine learning algorithm with a first data set;
A second discriminator including a learned algorithm obtained by learning the first machine learning algorithm with a second data set different from the first data set;
A determination device that performs a determination according to the output of the first determination device and the output of the second determination device;
The first discriminator accepts input of the subject data and outputs a first discrimination result;
The second discriminator receives the input of the subject data and outputs a second discrimination result;
A method in which the determination device outputs diagnosis support information, which is information for supporting diagnosis of the subject, according to the first determination result and the second determination result. A method performed by a diagnosis support apparatus that accepts subject data that is data related to a subject and supports diagnosis of the subject,
The diagnosis support apparatus includes:
A first discriminator including a learned algorithm obtained by learning a first machine learning algorithm with a first data set;
A first algorithm including a learned algorithm obtained by learning a second machine learning algorithm different from the first machine learning algorithm from the first data set or a second data set different from the first data set; 2 discriminators,
A determination device that performs a determination according to the output of the first determination device and the output of the second determination device;
The first discriminator accepts input of the subject data and outputs a first discrimination result;
The second discriminator receives the input of the subject data and outputs a second discrimination result;
A method in which the determination device outputs diagnosis support information, which is information for supporting diagnosis of the subject, according to the first determination result and the second determination result.

Images