Medical diagnostic support device, identification device, medical diagnostic support method and program

Abstract

Even with biased training data, stable classification results can be obtained. [Solution] The identification device 100 is an identifier 30 that identifies an object in a medical image by comparing an estimated value of the object to be identified with a cutoff value. The identifier 30 is used in which the cutoff value is adjusted for each category so that the index value evaluated using evaluation data 122 divided into categories approaches a target value common to multiple categories, thereby identifying the object to be identified in the medical image.

Description

Technical Field

[0006] , , , ,

[0001] The present invention relates to a medical diagnosis support device, an identifier generation device, a medical diagnosis support method, and a program.

Background Art

[0002] A technique for identifying an object to be identified in an image using an identifier learned by machine learning is known. For example, Patent Document 1 discloses a medical image classification device that classifies medical images into a plurality of types of case regions using a discriminator learned by a deep learning method.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, depending on the object to be identified, it may be difficult to collect training data for machine learning in a balanced manner. In such a case, since the training data is biased, there is a risk of bias in the identification result.

[0005] The present invention is for solving the above problems, and an object thereof is to obtain a stable identification result even when there is bias in the training data.

Means for Solving the Problems

[0006] To achieve the above objective, the medical diagnostic support device according to the present invention is characterized by comprising: an identifier that identifies an object to be identified in a medical image by comparing an estimated value of the object to be identified with a cutoff value, and an identification processing unit that identifies the object to be identified in the medical image using the identifier, the identifier having been adjusted for each category such that the index value evaluated using evaluation data divided into categories approaches a target value common to multiple categories. [Effects of the Invention]

[0007] According to the present invention, stable classification results can be obtained even if the training data is biased. [Brief explanation of the drawing]

[0008] [Figure 1] This is a block diagram showing the configuration of the identification device according to Embodiment 1. [Figure 2] This is a diagram showing the configuration of the identifier according to Embodiment 1. [Figure 3] This figure shows the relationship between the sensitivity and specificity of the classifier according to Embodiment 1 and the cutoff value. [Figure 4] This figure shows the sensitivity and specificity of the classifier by age group when the cutoff value is fixed in Embodiment 1. [Figure 5] This figure shows the sensitivity and specificity of the classifier for each age group when the cutoff value is set according to each age group in Embodiment 1. [Figure 6] This flowchart shows the flow of the classifier generation process performed by the identification device according to Embodiment 1. [Figure 7] This flowchart shows the flow of the cutoff value adjustment process performed by the identification device according to Embodiment 1. [Figure 8] This flowchart shows the flow of the identification process performed by the identification device according to Embodiment 1. [Figure 9] Figures (a) to (c) show the results of evaluating the classifier according to Embodiment 1. [Figure 10]This diagram shows the configuration of the identifier according to Embodiment 2. [Figure 11] This is a block diagram showing the configuration of the identifier generation device according to Embodiment 3. [Figure 12] This is a block diagram showing the configuration of the identification device according to Embodiment 3. [Modes for carrying out the invention]

[0009] Embodiments of the present invention will be described below with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals. The identification device 100 according to Embodiment 1 is a device that uses a classifier 30 generated by machine learning to identify whether an object to be identified in an input image is benign or malignant. In particular, the identification device 100 according to Embodiment 1 functions as a medical diagnostic support device that identifies whether a lesion in a medical image in which a lesion has been photographed is benign or malignant. Here, a medical image is an image taken for medical diagnostic purposes, in which a part of the living body (area of ​​interest) suspected of having a disease has been photographed. An example of a medical image is a dermatological image, which is an image of a lesion on the skin. Alternatively, the medical image is not limited to this, and may be any other image capable of photographing a lesion, such as an endoscopic image, X-ray image, CT (Computed Tomography) image, or ultrasound image.

[0010] As shown in Figure 1, the identification device 100 comprises a control unit 11, a storage unit 12, an operation unit 13, a display unit 14, and a communication unit 15. The control unit 11 includes a CPU (Central Processing Unit), ROM (Read Only Memory), and RAM (Random Access Memory). The CPU is a central processing unit that includes a microprocessor and performs various processes and calculations. In the control unit 11, the CPU reads the control program stored in the ROM and controls the operation of the entire identification device 100 while using the RAM as work memory. The processing of the control unit 11 may be executed by a single CPU or by multiple CPUs.

[0011] The memory unit 12 is a non-volatile memory such as flash memory or a hard disk. The memory unit 12 stores programs and data executed by the control unit 11, and data generated by the control unit 51. Specifically, the memory unit 12 stores training data 121, evaluation data 122, and the classifier 30. Details of these will be described later.

[0012] The operation unit 13 is equipped with input devices such as a keyboard, mouse, and touch panel, and accepts user input. The display unit 14 is equipped with a display device such as a liquid crystal display or an organic EL (Electro-Luminescence) display, and displays various images under the control of the control unit 11. The communication unit 15 is equipped with a communication interface for communicating with external devices of the identification device 100. For example, the communication unit 15 communicates with external devices in accordance with well-known communication standards such as LAN (Local Area Network) and USB (Universal Serial Bus).

[0013] The control unit 11 performs processing in two phases: a first phase, which is the learning phase of the classifier 30, and a second phase, which is the inference phase by the classifier 30. The control unit 11 includes a learning unit 111, a target value setting unit 112, and a cutoff adjustment unit 113 as functions in the learning phase. The control unit 11 also includes an image receiving unit 114, a classification processing unit 115, and a result output unit 116 as functions in the inference phase. In the control unit 11, the CPU functions by reading a program stored in ROM into RAM and executing that program to control each of these units. Note that in the control unit 11, one CPU may function as the learning phase and the inference phase units, or multiple CPUs may jointly function as the learning phase and the inference phase units.

[0014] First, the learning phase will be described. The learning phase is a phase in which an identifier 30 that can accurately identify the object to be identified in the input image is generated using machine learning techniques. Here, the identifier 30 is a computer program for identifying the quality of the object to be identified in the input image, and is a learned model learned by machine learning using training data 121.

[0015] Specifically, as shown in FIG. 2, the identifier 30 receives an input of a medical image in which a lesion to be identified is photographed, and age information of a patient having the lesion photographed in the medical image. Then, as output information for the input of the medical image, the identifier 30 outputs information indicating an identification result as to whether the lesion photographed in the medical image is benign or malignant.

[0016] More specifically, the identifier 30 includes an NN (Neural Network) unit 31 and a benign / malignant discrimination unit 32. The NN unit 31 is an arithmetic unit that executes the main arithmetic in the identifier 30. Specifically described, the NN unit 31 outputs a malignancy degree M for the input medical image by a method such as logistic regression or DNN (Deep Neural Network). Here, the malignancy degree M is a value indicating the probability that the lesion is malignant. The malignancy degree M takes a value between 0 and 1, meaning that the closer the malignancy degree M is to 0, the higher the probability that the lesion is benign, and the closer the malignancy degree M is to 1, the higher the probability that the lesion is malignant.

[0017] As an example, the NN unit 31 is composed of a neural network with a multi-layer structure, and has a plurality of layers including an input layer to which input data is input, an intermediate layer (hidden layer) that performs operations such as convolution and pooling on the input data, and an output layer (fully connected layer) that outputs the operation result. The NN unit 31 calculates the malignancy degree M of the lesion photographed in the medical image input to the input layer in the intermediate layer, and outputs the calculated malignancy degree M from the output layer.

[0018] The benign / malignant discrimination unit 32 outputs output information indicating whether the lesion captured in the input medical image is benign or malignant, based on the malignancy grade M output from the NN unit 31. Specifically, the benign / malignant discrimination unit 32 compares the malignancy grade M output from the NN unit 31 with a preset cutoff value. The benign / malignant discrimination unit 32 then determines that the lesion is malignant if the malignancy grade M is greater than the cutoff value, and determines that the lesion is benign if the malignancy grade M is less than the cutoff value. The cutoff value is preset to an appropriate value between 0 and 1 so that the benign / malignant discrimination unit 32 can appropriately distinguish between benign and malignant lesions. The benign / malignant discrimination unit 32 dynamically changes the cutoff value according to the age information input to the classifier 30, as will be described in more detail later.

[0019] The classifier 30 outputs the discrimination result from the benign / malignant discrimination unit 32 as output information. For example, if the classifier 30 determines that the lesion is malignant, it outputs the value "1" as output information, and if it determines that the lesion is benign, it outputs the value "0" as output information. In this way, the classifier 30 determines whether the object to be identified in the input medical image is benign or malignant by comparing the predicted value of benign or malignant with the cutoff value.

[0020] Returning to Figure 1, in the learning phase, the learning unit 111 performs machine learning using the training data 121. Here, the training data 121 is a dataset (training dataset) used by the learning unit 111 to perform machine learning. The training data 121 includes multiple training input images (hereinafter referred to as "training images") used as training data. Each of the multiple training images is an image of a lesion, and the correct answer of whether the captured lesion is benign or malignant is known in advance. In the training data 121, each training image is pre-assigned a correct malignancy score LM, which is the correct answer for the captured lesion, as a training label (correct label). The correct malignancy score LM is represented as a binary value of 1 or 0, where 1 is when the lesion is malignant and 0 is when the lesion is benign. For example, the results of a pathological diagnosis (biopsy) can be used as the correct malignancy score LM.

[0021] The learning unit 111 performs machine learning using multiple training images included in the training data 121 as training data. Specifically, the learning unit 111 inputs each of the multiple training images included in the training data 121 into the classifier 30. In the classifier 30, the NN unit 31 calculates and outputs an estimated value of the malignancy degree M of the lesion captured in the input training image. The learning unit 111 adjusts the calculation parameters of the NN unit 31 using methods such as backpropagation so that the malignancy degree M output from the NN unit 31 approaches the correct malignancy degree LM assigned to the input training image. The calculation parameters of the NN unit 31 are, for example, the weights of the connections in each layer of the neural network in the NN unit 31, that is, weights indicating the connection strength between multiple neurons (nodes). By adjusting the calculation parameters, the estimated value of malignancy degree M output by the NN unit 31 for the input training image changes. The learning unit 111 adjusts the computation parameters by varying them in various ways so that the malignancy score M output from the NN unit 31 approaches the correct malignancy score LM as closely as possible. The learning unit 111 performs this computation parameter adjustment process for each of the multiple training images included in the training data 121, thereby optimizing the computation parameters of the NN unit 31 and constructing the neural network in the NN unit 31.

[0022] Returning to Figure 1, the target value setting unit 112 sets target values ​​for the discrimination index, which is an index value related to the discrimination of the classifier 30. Here, the discrimination index is an index value for evaluating the performance of the classifier 30 in distinguishing between benign and malignant lesions. Specifically, the target value setting unit 112 sets the target value of the sensitivity of the classifier 30 and the target value of the specificity of the classifier 30 as target values ​​for the discrimination index. Sensitivity is the proportion of malignant cases that were correctly identified as malignant. In contrast, specificity is the proportion of benign cases that were correctly identified as benign. Sensitivity and specificity have a trade-off relationship; if sensitivity is made too high, specificity will decrease, and conversely, if specificity is made too high, sensitivity will decrease.

[0023] More specifically, as shown in Figure 3, the sensitivity and specificity of the classifier 30 vary depending on the cutoff value used in the benign / malignant determination unit 32. For example, if the cutoff value is set too high, the identification results will be biased towards benign, resulting in high specificity and low sensitivity. Conversely, if the cutoff value is set too low, the identification results will be biased towards malignant, resulting in high sensitivity and low specificity. Therefore, by changing the cutoff value used in the benign / malignant determination unit 32, the combination of sensitivity and specificity, i.e., the balance between sensitivity and specificity, can be adjusted. Note that Figure 3 is a simplified representation of the relationship between the cutoff value and sensitivity and specificity, and is not necessarily accurate.

[0024] Incidentally, discrimination indicators such as sensitivity and specificity also vary depending on the category of the object to be identified, including age. For example, the sensitivity or specificity of identifying lesions in younger patients is not necessarily the same as the sensitivity or specificity of identifying lesions in older patients. In particular, if there is a bias between categories in the multiple training images included in the training data 121, it is possible that the sensitivity or specificity of a particular category will differ significantly from that of other categories. To avoid such a situation and generate a classifier 30 that can reliably identify any category, it is desirable to stabilize the balance between sensitivity and specificity as much as possible, regardless of the category. The target value setting unit 112 sets target values ​​for sensitivity and specificity for this purpose. In the following, we will explain using age groups as an example of a category.

[0025] First, the target value setting unit 112 sets target values ​​common to multiple categories (multiple age groups) as target values ​​for sensitivity and specificity. To do this, the target value setting unit 112 uses evaluation data 122. Here, evaluation data 122 is a dataset used to evaluate the discrimination index of the classifier 30. Evaluation data 122 includes multiple input images for evaluation (hereinafter referred to as "evaluation images"). Each of the multiple evaluation images is pre-labeled with the correct malignancy grade LM of the captured lesion as a training label, similar to the training images. In addition, each of the multiple evaluation images is labeled with information on the age LA of the patient with the captured lesion. Note that all or part of the multiple evaluation images in evaluation data 122 may be the same as the training images. Alternatively, machine learning in the NN unit 31 and adjustment of the cutoff value in the benign / malignant discrimination unit 32 may be performed using only the training data 121 by a cross-validation method.

[0026] The target value setting unit 112 inputs each of the multiple evaluation images included in the evaluation data 122 into the classifier 30 without separating them by age group (category). In other words, although each evaluation image included in the evaluation data 122 is accompanied by information on the age LA of the patient with the captured lesion, the target value setting unit 112 does not use this age LA information and inputs evaluation images of lesions from all age groups into the classifier 30.

[0027] When each evaluation image is input to the classifier 30, the target value setting unit 112 compares the output information indicating benign or malignant, which is output from the classifier 30 for each input evaluation image, with the correct malignancy score LM for each evaluation image. The target value setting unit 112 determines whether the output information and the correct malignancy score LM match for each evaluation image and calculates the discrimination indices, sensitivity and specificity. Specifically, the target value setting unit 112 obtains the "sensitivity" value by calculating the proportion of the output information of the classifier 30 that is malignant (1) for a given number of input evaluation images where the correct malignancy score LM is malignant (1). The target value setting unit 112 also obtains the "specificity" value by calculating the proportion of the output information of the classifier 30 that is benign (0) for a given number of input evaluation images where the correct malignancy score LM is benign (0).

[0028] The target value setting unit 112 calculates sensitivity and specificity in this way by varying the cutoff value in multiple ways, thereby obtaining multiple combinations of sensitivity and specificity. Then, the target value setting unit 112 sets the sensitivity and specificity of the combination that satisfies a specific condition from among the obtained multiple combinations as the target values, sensitivity A and specificity A. Here, the specific condition is one that does not excessively reduce both sensitivity and specificity, and that allows both to achieve a reasonably high value. For example, one specific condition can be set as one in which both sensitivity and specificity are above a predetermined lower limit. The lower limit can be set to an appropriate value such as 0.7 or 0.75.

[0029] More specifically, in addition to the condition that both sensitivity and specificity are above a predetermined lower limit, a condition is set that sensitivity should be higher than specificity in order to minimize the missed detection of malignant cases. In other words, the target value setting unit 112 sets the target value of sensitivity to a value higher than the target value of specificity. In the example in Figure 3, when the cutoff value is set to C0, the sensitivity becomes 0.85 and the specificity becomes 0.8. Such sensitivity (=0.85) and specificity (=0.8) are higher than the lower limit and satisfy the condition that sensitivity is higher than specificity. Therefore, the target value setting unit 112 can set the sensitivity (=0.85) and specificity (=0.8) at the cutoff value C0 as the target value of sensitivity and the target value of specificity, respectively.

[0030] It is generally assumed that there are typically multiple combinations of sensitivity and specificity that satisfy specific conditions. The target value setting unit 112 may automatically set the target sensitivity and specificity combination from among these multiple combinations. Alternatively, the user may manually select the desired target combination from among the multiple combinations. In the following explanation, we will use the example shown in Figure 3, where the target value setting unit 112 sets sensitivity (=0.85) and specificity (=0.8) as target values ​​for sensitivity and specificity common to multiple age groups (hereinafter referred to as "sensitivity A" and "specificity A," respectively). In this case, the target value setting unit 112 sets a cutoff value C0 that can achieve sensitivity A and specificity A as a cutoff value common to multiple age groups.

[0031] Returning to Figure 1, the cutoff adjustment unit 113 sets cutoff values ​​for each age group. In other words, while the target value setting unit 112 sets a cutoff value C0 common to multiple age groups, the cutoff adjustment unit 113 sets individual cutoff values ​​for each of the multiple age groups. To explain in more detail, in identification using the cutoff value C0 set by the target value setting unit 112, sensitivity A and specificity A can be achieved for the entire evaluation data 122. On the other hand, when multiple evaluation images included in the evaluation data 122 are separated by age group, depending on the age group, sensitivity A and specificity A may not be achieved with the cutoff value C0.

[0032] For example, Figure 4 shows the sensitivity and specificity of the classifier 30 when the cutoff value used in the benign / malignant discrimination unit 32 is fixed at C0, for all age groups, young age groups, middle age groups, and elderly age groups. Here, the young age groups, middle age groups, and elderly age groups are, for example, the ranges of 0-39 years, 40-59 years, and 60 years and older, respectively. To explain in more detail, first, the multiple evaluation images included in the evaluation data 122 are divided into three age groups: young age group, middle age group, and elderly age group, according to the age of the patient in the evaluation image. Then, for each of the three age groups, the sensitivity and specificity of the classifier 30 are evaluated using the evaluation images divided into each age group, and the results are shown as the data for the young age group, middle age group, and elderly age group in Figure 4. In contrast, the data for the all age group in Figure 4 is the result of evaluating the sensitivity and specificity of the classifier 30 using all of the multiple evaluation images included in the evaluation data 122 without dividing them by age group.

[0033] In Figure 4, the sensitivity and specificity for all age groups are the target values ​​set by the target value setting unit 112: sensitivity A (=0.85) and specificity A (=0.8). Furthermore, the sensitivity and specificity for the middle-aged and older age groups do not differ significantly from those for all age groups, being approximately 0.85 and 0.8, respectively. In contrast, the sensitivity and specificity for the younger age group differ significantly from those for the other age groups, with sensitivity being lower than specificity. This situation can occur when the probability of malignant cases is low in the younger age group and the number of training images for the younger age group in the training data 121 is small. To avoid an imbalance between sensitivity and specificity in any age group and to ensure stable identification in all age groups, the cutoff adjustment unit 113 sets cutoff values ​​individually for each age group.

[0034] Specifically, the cutoff adjustment unit 113 first divides the multiple evaluation images included in the evaluation data 122 by age group. Specifically, as in Figure 4, the cutoff adjustment unit 113 divides the multiple evaluation images into three age groups according to the age of the patient captured in the evaluation image: young age group (0-39 years old), middle age group (40-59 years old), and old age group (60 years and older). Note that the division of age groups is not limited to these. Also, the number of age group divisions is not limited to three; it may be two, four or more.

[0035] Next, the cutoff adjustment unit 113 adjusts the cutoff value for each age group so that the discrimination index of the classifier 30, when evaluated using evaluation data 122 divided by age group, approaches the target value set by the target value setting unit 112. Specifically, for each of the evaluation data 122 divided by age group, the cutoff adjustment unit 113 calculates the loss function E represented by equation (1) below, while varying the cutoff value used in the benign / malignant discrimination unit 32. The loss function E is a function for evaluating how close the sensitivity and specificity of the classifier 30 have come to their respective target values, sensitivity A and specificity A. E = α × (Sensitivity A - Sensitivity)^2 + (1 - α) × (Specificity A - Specificity)^2 …(1)

[0036] In equation (1) above, α is a parameter that indicates the weights of the sensitivity term and the specificity term in the loss function E. Parameter α is an external variable that can be freely set by the user within the range of 0 to 1. For example, if sensitivity and specificity are treated equally, parameter α is set to 0.5. Alternatively, if sensitivity is prioritized over specificity when adjusting the cutoff value, parameter α may be set to a value greater than 0.5.

[0037] The cutoff adjustment unit 113 inputs each evaluation image from the evaluation data 122, which is divided by age group, into the classifier 30, and determines whether the output information indicating benign or malignant, output from the classifier 30 for each input evaluation image, matches the correct malignancy score LM for each evaluation image. Then, the cutoff adjustment unit 113 evaluates the sensitivity and specificity based on the result of the determination, and calculates the loss function E from the evaluated sensitivity and specificity. The cutoff adjustment unit 113 performs this process of calculating the loss function E by changing the cutoff value in multiple ways, and searches for the cutoff value that makes the loss function E smallest, that is, the closest to 0. As a result of the search, the cutoff adjustment unit 113 determines the cutoff value that makes the loss function E smallest as the cutoff value for each age group.

[0038] The cutoff adjustment unit 113 determines the cutoff value individually for each of the following age groups by performing a process to search for such cutoff values ​​for each of the evaluation data 122, which is divided into young age groups, middle age groups, and old age groups. In this way, the cutoff adjustment unit 113 independently determines the cutoff value C1 for the young age group, the cutoff value C2 for the middle age group, and the cutoff value C3 for the old age group. Here, the cutoff values ​​C1 to C3 for each age group represent the cutoff values ​​used by the benign / malignant discrimination unit 32 when medical images of lesions of patients belonging to the corresponding age group are input to the classifier 30. Note that the cutoff values ​​C1 to C3 for each age group may, as a result of the adjustment, be the same as the cutoff value C0 common to all age groups.

[0039] Figure 5 shows examples of the sensitivity and specificity of the classifier 30 for evaluation images of all age groups, young age groups, middle age groups, and elderly age groups, when using cutoff values ​​individually adjusted for each age group by the cutoff adjustment unit 113. In Figure 5, it can be seen that the sensitivity and specificity for the young age group are closer to the target values ​​of sensitivity A (=0.85) and specificity A (=0.8) set by the target value setting unit 112, compared to the case in Figure 4. The sensitivity and specificity for the middle age group and elderly age group were close to the sensitivity and specificity for all age groups even in Figure 4, so they have not changed significantly in Figure 5. In this way, because the cutoff adjustment unit 113 adjusts the cutoff value individually for each age group, it is possible to avoid an imbalance between the sensitivity and specificity of the classifier 30 in a particular age group. Therefore, it becomes possible to generate a classifier 30 that can stably identify any age group.

[0040] Next, with reference to Figure 6, the flow of the classifier generation process performed by the classifier device 100 during the learning phase will be explained. The classifier generation process shown in Figure 6 starts when the operation unit 13 receives a start command from the user. The classifier generation process shown in Figure 6 is an example of a classifier generation method.

[0041] When the classifier generation process is started, the control unit 11 prepares training data 121 and evaluation data 122 (step S101). Next, the control unit 11 functions as a learning unit 111 and performs machine learning using the training data 121 (step S102). Specifically, the control unit 11 inputs each training image included in the training data 121 into the classifier 30. Then, the control unit 11 adjusts the calculation parameters of the NN unit 31 so that the malignancy score M output from the classifier 30 approaches the correct malignancy score LM assigned to the input training image. After performing machine learning, the control unit 11 adjusts the cutoff value used for determining benign or malignant in the benign / malignant discrimination unit 32 (step S103). Details of the cutoff value adjustment process in step S103 will be explained with reference to Figure 7.

[0042] When the cutoff value adjustment process shown in Figure 7 is started, the control unit 11 functions as a target value setting unit 112 and sets target values ​​for sensitivity and specificity, which are the identification indices of the classifier 30, that are common to multiple age groups (step S201). Specifically, the control unit 11 evaluates the sensitivity and specificity of the classifier 30 using evaluation data 122 that has not been divided by age group. The control unit 11 performs this evaluation of sensitivity and specificity by changing the cutoff value in multiple ways, and sets the combination of sensitivity and specificity that satisfies specific conditions as the target value from among the multiple combinations of sensitivity and specificity obtained. For example, as shown in Figure 3, the control unit 11 sets sensitivity A and specificity A at cutoff value C0.

[0043] Once a target value is set, the control unit 11 divides the evaluation data 122 by age group (step S202). Specifically, the control unit 11 divides each evaluation image into multiple age groups, such as young age group, middle age group, and old age group, based on the age LA information attached to each evaluation image included in the evaluation data 122. After dividing the evaluation data 122, the control unit 11 functions as a cutoff adjustment unit 113 and uses the evaluation data 122 divided into multiple age groups to set a cutoff value for one of the multiple age groups (step S203). Specifically, the control unit 11 selects one age group from the multiple age groups. Then, the control unit 11 evaluates the sensitivity and specificity of the classifier 30 using the evaluation data 122 of the selected age group and calculates the loss function E expressed by equation (1) above. The control unit 11 performs the calculation of such loss function E for each of the multiple cutoff values ​​and sets the cutoff value that minimizes the loss function E as the cutoff value for the selected age group.

[0044] When the cutoff value for one age group is adjusted, it is determined whether or not there are any unprocessed age groups among the multiple age groups (step S204). If there are unprocessed age groups (step S204; YES), the control unit 11 returns to step S203. The control unit 11 then selects one new unprocessed age group and sets the cutoff value for the newly selected age group. In this way, the control unit 11 sets the cutoff value individually for each of the multiple age groups. Finally, when there are no more unprocessed age groups (step S204; NO), the control unit 11 terminates the cutoff value adjustment process. This completes the classifier generation process shown in Figures 6 and 7.

[0045] Returning to Figure 1, the second phase will be described. The inference phase is a phase in which the classifier 30 generated in the learning phase is used to identify whether an unknown lesion in an unknown medical image is benign or malignant. In the inference phase, the image receiving unit 114 receives input of an unknown medical image to be identified. Here, an unknown medical image is an image in which a lesion of unknown benign or malignant nature has been captured. The image receiving unit 114 accepts the designation of an unknown medical image to be identified from among a plurality of medical images pre-stored in the storage unit 12, according to the user's instructions from the operation unit 13. Alternatively, the image receiving unit 114 may receive the unknown medical image to be identified from an external source via the communication unit 15.

[0046] In addition to receiving medical images, the image reception unit 114 also accepts age information. Here, age information is information indicating the age of the person to be identified in the input medical image, and specifically, it is information indicating which age group—young, middle-aged, or elderly—the patient with the lesion belongs to. In Embodiment 1, the age information is known in advance from information such as medical records and is pre-associated with each of the multiple medical images that may be the target of identification.

[0047] The identification processing unit 115 uses a classifier 30 to which cutoff values ​​adjusted by category (age group) by the cutoff adjustment unit 113 are applied to identify whether the lesion in the unknown medical image received by the image receiving unit 114 is benign or malignant. First, as shown in Figure 2, the identification processing unit 115 selects a cutoff value to be used in the benign / malignant determination unit 32 according to the age information received by the image receiving unit 114. Specifically, the identification processing unit 115 selects the cutoff value for the age group corresponding to the input age information from among the cutoff values ​​individually adjusted for each of the multiple age groups by the cutoff adjustment unit 113, and uses this cutoff value to be used in the benign / malignant determination unit 32.

[0048] Next, the identification processing unit 115 inputs the unknown medical image received by the image receiving unit 114 to the classifier 30. In the classifier 30, the NN unit 31 calculates an estimated value of the malignancy degree M of the lesion captured in the input unknown medical image using computation parameters for which machine learning has been performed by the learning unit 111. The benign / malignant discrimination unit 32 determines whether the lesion is benign or malignant by comparing the estimated value of malignancy degree M calculated by the NN unit 31 with a cutoff value selected according to age information. Specifically, the benign / malignant discrimination unit 32 determines that the lesion is malignant if the estimated value of malignancy degree M exceeds the cutoff value, and determines that the lesion is benign if the estimated value of malignancy degree M falls below the cutoff value. The classifier 30 outputs output information indicating whether it is benign or malignant as determined by the benign / malignant discrimination unit 32. The identification processing unit 115 identifies whether a lesion is benign or malignant based on the output information output from the classifier 30 for an unknown medical image.

[0049] The result output unit 116 outputs the identification result from the identification processing unit 115. Specifically, the result output unit 116 displays on the display unit 14 output information indicating whether the lesion in the unknown medical image is benign or malignant, which is the result identified by the identification processing unit 115. Alternatively, the result output unit 116 may output the output information as audio, or output it to an external device via the communication unit 15. This allows the user to confirm the identification result from the identification device 100.

[0050] Next, with reference to Figure 8, the flow of the identification process performed by the identification device 100 in the inference phase will be explained. The identification process shown in Figure 8 starts when the operation unit 13 receives a start instruction from the user, while the identification device 30 generated by the identification device generation process shown in Figure 6 is stored in the storage unit 12.

[0051] When the identification process is started, the control unit 11 functions as an image receiving unit 114 and receives input of an unknown medical image in which the unknown lesion to be identified is captured, and age information (step S301). The control unit 11 functions as an identification processing unit 115 and selects a cutoff value to be used in the benign / malignant discrimination unit 32 according to the input age information (step S302). Specifically, the control unit 11 selects the cutoff value for the age group corresponding to the input age information from among cutoff values ​​that have been individually adjusted for each of several age groups, and uses this as the cutoff value to be used in the benign / malignant discrimination unit 32.

[0052] Next, the control unit 11 functions as an identification processing unit 115 and inputs an unknown medical image to the classifier 30. In the classifier 30, the NN unit 31 estimates the malignancy degree M of the lesion captured in the input unknown medical image (step S303). Then, the benign / malignant discrimination unit 32 compares the estimated malignancy degree M calculated by the NN unit 31 with the cutoff value selected in step S302 according to the age information to determine whether the lesion is benign or malignant (step S304). The control unit 11 obtains the benign or malignant identification result from the classifier 30. Next, the control unit 11 functions as a result output unit 116 and outputs output information indicating the acquired identification result (step S305). With this, the identification process shown in Figure 8 is completed. Because the identification process described above uses a classifier 30 whose cutoff value is adjusted for each age group in the classifier generation process, it can accurately distinguish between benign and malignant lesions even if there is a bias in the training data 121 with respect to age groups.

[0053] As described above, the identification device 100 according to Embodiment 1 sets target values ​​common to multiple age groups as target values ​​for sensitivity and specificity, which are the identification indices of the classifier 30, and adjusts the cutoff value for each age group so that the sensitivity and specificity evaluated using evaluation data 122 separated by age group approaches the target value. Then, the identification device 100 according to Embodiment 1 uses the classifier 30 with the cutoff value adjusted for each age group to identify whether an unknown lesion captured in an unknown medical image is benign or malignant. In this way, because the identification device 100 according to Embodiment 1 adjusts the cutoff value of the classifier 30 for each age group, even if the training images are not collected in a balanced manner for each of the multiple age groups and there is a bias in the training data 121 with respect to the patient's age group, it is possible to set an appropriate balance of sensitivity and specificity between age groups. As a result, stable identification results can be obtained.

[0054] In particular, in clinical practice, many cases in young patients are benign diseases. Therefore, it is difficult to collect enough training images of malignant cases in young patients. When there is a bias in the disease distribution within the training data 121, the identification results are more likely to be biased. For example, if the number of malignant cases is smaller than the number of benign cases, the identification results are more likely to be biased towards the benign side. In contrast, the identification device 100 according to Embodiment 1 sets cutoff values ​​individually for each of the multiple age groups and dynamically sets the cutoff values ​​according to the age group of the lesion to be identified. Therefore, even if there is a large bias in the case distribution of the training data 121, stable identification results can be obtained.

[0055] Figures 9(a) to (c) show the results of experimental evaluation of the classifier 30 according to Embodiment 1 described above. In Figures 9(a) to (c), the horizontal axis represents the age groups of patients captured in medical images, divided into three age groups: young (~40 years old), middle-aged (40 to under 70 years old), and elderly (70 years old and over). In Figures 9(a) and (b), the vertical axis represents the sensitivity and specificity of the classifier 30 evaluated without limiting the type of lesion, respectively. In Figure 9(c), the vertical axis represents the sensitivity of the classifier 30 evaluated only when the lesion is melanoma. As for the evaluation method, using multiple evaluation images divided by age group, the sensitivity or specificity of the classifier 30, which uses cutoff values ​​adjusted for each age group according to the method described in Embodiment 1, and a conventional classifier that uses a common cutoff value for multiple age groups were evaluated for each age group. In Figures 9(a) to 9(c), the solid lines show the results obtained by the classifier 30 in Embodiment 1, while the dashed lines show the results obtained by the conventional (original) classifier.

[0056] As a result, as shown in Figures 9(a) to (c), there was no significant difference in the sensitivity and specificity of the classifier 30 of Embodiment 1 and the conventional classifier between the middle-aged and older age groups, but a difference was observed in the younger age group. More specifically, the sensitivity and specificity in the younger age group differed significantly from that of other age groups in the conventional (original) classifier, but the classifier 30 of Embodiment 1 approached the sensitivity and specificity of other age groups. In particular, for melanoma shown in Figure 9(c), the sensitivity in the younger age group was greatly improved in the classifier 30 of Embodiment 1 compared to the conventional classifier. Since melanoma is a dangerous malignant tumor, it is especially important to identify it with high accuracy. On the other hand, melanoma is rare, and there are few cases regardless of age, but there are especially few cases of melanoma in young children. Therefore, it is difficult to collect melanoma in young children as training data 121. In contrast, it was confirmed that the classifier 30 of Embodiment 1 can identify melanoma in younger age groups, where there are fewer cases, with a sensitivity similar to that of other age groups. Based on these experimental results, it can be seen that the classifier 30, which uses cutoff values ​​adjusted for each age group, is effective in obtaining stable identification results even when there is a bias in the case distribution in the training data 121.

[0057] Next, Embodiment 2 will be described. The same configuration and functions as in Embodiment 1 will not be described. In Embodiment 1, the image receiving unit 114 received input of an unknown medical image as well as input of age information of the subject to be identified in that medical image. In contrast, in Embodiment 2, the image receiving unit 114 does not receive input of age information, and the classifier 30 estimates the age of the subject to be identified from the medical image.

[0058] Specifically, as shown in Figure 10, the classifier 30 includes an NN unit 33 in addition to the NN unit 31 and the benign / malignant discrimination unit 32 described in Embodiment 1. The NN unit 33 is an estimation unit that estimates the age of the object to be identified in a medical image, given that the input is a medical image. In other words, the NN unit 33 functions as a category estimation unit that estimates the category to which the object to be identified in the input medical image belongs from among multiple categories. Like the NN unit 31, the NN unit 33 is configured, for example, by a multi-layered neural network. While the NN unit 31 calculates and outputs an estimated value of the malignancy degree M of the object to be identified in a medical image, given that the input is a medical image, the NN unit 33 calculates and outputs an estimated value of the age of the object to be identified in a medical image, given that the input is a medical image. Like the NN unit 31, the NN unit 33 is trained by machine learning during the learning phase. Specifically, the learning unit 111 inputs multiple training images, each labeled with the correct age LA, into the NN unit 33, and learns the computational parameters of the NN unit 33 so that the estimated age output from the NN unit 33 approaches the correct age LA attached to the input training images.

[0059] In the inference phase, when an unknown medical image is input by the image receiving unit 114, the identification processing unit 115 inputs the unknown medical image to the classifier 30. In the classifier 30, the NN unit 31 calculates an estimated value of the malignancy degree M of the lesion captured in the input unknown medical image. The benign / malignant discrimination unit 32 selects a cutoff value corresponding to the age estimated by the NN unit 33 from among the cutoff values ​​adjusted by the cutoff adjustment unit 113 for each age group. Then, the benign / malignant discrimination unit 32 determines whether the lesion is benign or malignant by comparing the estimated value of malignancy degree M calculated by the NN unit 31 with the selected cutoff value. The classifier 30 outputs output information indicating benign or malignant as determined by the benign / malignant discrimination unit 32. Based on the output information output from the classifier 30 for the unknown medical image, the identification processing unit 115 identifies whether the lesion is benign or malignant. Thus, the classifier 30 according to Embodiment 2 can estimate the age of the subject to be identified from a medical image. Therefore, even for medical images where the age of the subject to be identified is unknown, it can distinguish whether the subject is benign or malignant by using cutoff values ​​adjusted for each age group.

[0060] Next, Embodiment 3 will be described. The same configuration and functions as in Embodiments 1 and 2 will not be described. The identification device 100 according to Embodiments 1 and 2 had both a learning phase function and an inference phase function, as shown in Figure 1. In contrast, the identification device 100 according to Embodiment 3 does not have a learning phase function, and a separate device, the classifier generation device 200, has the learning phase function.

[0061] Specifically, as shown in Figure 11, the classifier generation device 200 according to Embodiment 3 comprises a control unit 21, a storage unit 22, an operation unit 23, a display unit 24, and a communication unit 25. The hardware configuration of each of these units is the same as that of the control unit 11, storage unit 12, operation unit 13, display unit 14, and communication unit 15 in the classifier device 100, so a description is omitted. The control unit 21 includes a learning unit 111, a target value setting unit 112, and a cutoff adjustment unit 113 as functions for the learning phase. In the control unit 21, the CPU functions by reading the program stored in ROM into RAM and executing the program to control these units. The storage unit 22 stores training data 121 and evaluation data 122 as data used in the learning phase.

[0062] Each function of the control unit 21 is the same as the learning unit 111, target value setting unit 112, and cutoff adjustment unit 113 provided in the identification device 100 in Embodiment 1. Specifically, the control unit 21 performs the classifier generation process shown in Figure 6 using the functions of the learning unit 111, target value setting unit 112, and cutoff adjustment unit 113. As a result, the control unit 21 generates a classifier 30 that identifies whether the lesion is benign or malignant using cutoff values ​​adjusted for each category (age group).

[0063] On the other hand, the identification device 100 according to Embodiment 3 is configured as shown in Figure 12. In the identification device 100 according to Embodiment 3, the control unit 11 does not have a learning phase function, but as an inference phase function, it includes an image receiving unit 114, an identification processing unit 115, and a result output unit 116. The functions of each of these units are the same as in Embodiment 1. The identification device 100 acquires the classifier 30 generated by the classifier generation device 200 from the classifier generation device 200, for example, through communication via the communication unit 15, and stores it in the storage unit 12. The control unit 11, using the functions of the image receiving unit 114, the identification processing unit 115, and the result output unit 116, uses the classifier 30 acquired from the classifier generation device 200 to execute the identification process shown in Figure 8. As a result, the control unit 11 identifies whether the lesion captured in the medical image is benign or malignant. Thus, in Embodiment 3, since the functions of the learning phase and the inference phase are separated into different devices, more flexible operation becomes possible, such as utilizing the classifier 30 generated by the classifier generation device 200 in multiple classifier devices 100.

[0064] Although embodiments of the present invention have been described above, these embodiments are merely examples, and the scope of application of the present invention is not limited thereto. In other words, the embodiments of the present invention can be applied in various ways, and all embodiments fall within the scope of the present invention.

[0065] For example, in the above embodiment, the discrimination indices, which are index values ​​related to the discrimination of the classifier 30, were sensitivity and specificity. However, the discrimination indices may be index values ​​other than sensitivity or specificity, such as accuracy rate or precision rate. Furthermore, the discrimination indices are not limited to using multiple discrimination indices such as sensitivity and specificity; for example, only one of sensitivity or specificity may be used, and the cutoff value may be adjusted for each category. Also, in the above embodiment, the target value setting unit 112 varied the cutoff value in multiple ways to search for a combination of sensitivity and specificity that satisfies specific conditions, and set sensitivity A and specificity A that satisfies specific conditions as a target value common to multiple categories. However, the target value setting unit 112 is not limited to searching by cutoff value in this way; for example, it may set a typical value of sensitivity or specificity used in clinical practice as a fixed target value.

[0066] In the above embodiment, the target value setting unit 112 sets target values ​​common to multiple categories as target values ​​for sensitivity and specificity, and the cutoff adjustment unit 113 adjusts the cutoff value for each category so that the sensitivity and specificity for each age group approach the target values ​​common to multiple categories, namely sensitivity A and specificity A. However, the cutoff adjustment unit 113 is not limited to adjusting the cutoff value using one same target value for multiple categories, but may also change the target value for each category. In other words, the cutoff adjustment unit 113 may adjust the cutoff value using the target value common to multiple categories itself, as long as it is based on the target value common to multiple categories set by the target value setting unit 112, or it may adjust the cutoff value using a value determined based on the target value common to multiple categories. This makes it possible to change the balance between sensitivity and specificity in a particular category compared to other categories.

[0067] For example, considering the significant social impact of developing cancer at a young age, the target sensitivity for a specific age group, namely the younger age group, may be set higher than the target sensitivity for other age groups. In this case, the cutoff adjustment unit 113 sets the target sensitivity and specificity for age groups other than the younger age group to sensitivity A and specificity A, which are common target values ​​for multiple age groups set by the target value setting unit 112, while setting the target sensitivity for the younger age group to sensitivity B, which is higher than sensitivity A, and specificity B, which is lower than specificity A. The cutoff adjustment unit 113 then adjusts the cutoff value for the younger age group so that the sensitivity and specificity when the classifier 30 identifies the quality of the object to be identified in the younger age group using the age-separated evaluation data 122 approaches sensitivity B and specificity B, respectively. Additionally, if necessary, the weight of the sensitivity term in the loss function E may be increased by making the value of α in equation (1) greater than 0.5. In this way, different sensitivity or specificity target values ​​can be set for specific categories such as young age groups. Therefore, even if the risks and social impacts differ between categories, the sensitivity and specificity of the classifier 30 can be adjusted more flexibly.

[0068] In the above embodiment, age group was used as an example of a category when adjusting the cutoff value by category. However, the category is not limited to age group; it may also be information indicating, for example, the size of the object to be identified, body part, race, etc. Here, the size of the object to be identified refers to the size (typically the longest diameter) of the lesion to be identified. Generally, just as there are few malignant cases at a young age, there are few malignant cases of small size. Therefore, by dividing the evaluation data 122 according to the size of the lesion and adjusting the cutoff value according to the size, stable identification results can be obtained even if there is a bias in the training data 121 with respect to age, as in the above embodiment. Furthermore, the body part to be identified refers to the body part where the lesion to be identified is located (e.g., face, palms and soles, mucous membranes, etc.). The race to be identified refers to the race of the patient who has the lesion to be identified (e.g., Caucasian, Black, Asian, etc.). Because it can be difficult to collect training images in a balanced manner depending on body part or race, by dividing the evaluation data 122 by body part or race and adjusting the cutoff value for each body part or race, stable classification results can be obtained even if there is a bias in the training data 121. In this way, using information that is prone to bias in the training data 121 as a category has the effect of obtaining stable classification results.

[0069] In the above embodiment, the identification device 100 identified whether the lesion was benign or malignant. However, the identification device 100 is not limited to this and may also be a device that identifies recommendations or suggestions for the identified subject. Here, a recommendation for a referral means recommending that a patient with the identified lesion be referred to another hospital. For example, it is conceivable that a small clinic would refer a patient to a larger hospital capable of more specialized examinations. In the above embodiment, each training image was assigned a correct malignancy score LM indicating malignancy (1) or benignity (0) as a teacher label. However, when the identification device 100 identifies a recommendation for a referral, instead of the correct malignancy score LM, information indicating whether a recommendation for a referral is necessary (1) or unnecessary (0) is assigned to each training image as a teacher label. The learning unit 111 performs machine learning using such training data 121. As a result, the learning unit 111 generates a classifier 30 that, in response to an input medical image, outputs output information indicating whether a recommendation for a referral is necessary or not (presence or absence of a recommendation for a referral) for the lesion captured in that medical image. Alternatively, the training label may remain the correct malignancy grade LM, but the cutoff value used in the benign / malignant discrimination unit 32 may be set higher (i.e., more sensitive) than when identifying whether the lesion site is benign or malignant, thereby allowing the discrimination device 100 to be used as a referral recommendation discrimination device.

[0070] In the above embodiment, the identification device 100 was a medical diagnostic support device that identified whether a lesion captured in a medical image was benign or malignant. However, the identification device 100 is not limited to being a medical diagnostic support device. For example, the identification device 100 may be an inspection device that receives input of inspection images of structures such as buildings, roads, and bridges, and identifies whether or not there is an abnormality in the structure based on cracks, surface shape, etc., of the structure captured in the inspection image. In this case, the object of identification corresponds to the structure captured in the inspection image, not the lesion captured in the medical image. Furthermore, the goodness or badness of the object of identification corresponds to the presence or absence of an abnormality in the structure, not whether the lesion is benign or malignant.

[0071] Furthermore, the identification device 100 is not limited to outputting binary information as an identification result, such as whether the lesion is benign or malignant, whether a referral is necessary, or whether the building is in good condition (whether there is an abnormality). It may also output information that is more than binary as an identification result. For example, the identification device 100 may use an identifier 30 capable of identifying eight diseases to output an identification result indicating which of the eight diseases the lesion belongs to. Here, the eight diseases are, for example, eight major diseases: melanoma, basal cell carcinoma, other malignant diseases, pigmented nevus, seborrheic keratosis, dermatofibroma, hemangioma, and other benign diseases. In this case, the NN unit 31 outputs a probability value indicating the probability that the lesion captured in the input medical image corresponds to each of the eight diseases, instead of the malignancy grade M in the above embodiment. For example, if the input medical image is a melanoma image, the learning unit 111 trains the NN unit 31 so that the probability value for melanoma approaches 1, and the probability values ​​for the other seven diseases approach 0. The benign / malignant discrimination unit 32 then determines that the disease with the largest probability value among the eight probability values ​​output from the NN unit 31 is the disease corresponding to the lesion captured in the input medical image. For example, if the probability value for melanoma is the largest, the benign / malignant discrimination unit 32 determines that the lesion is melanoma. The classifier 30 outputs the discrimination result from the benign / malignant discrimination unit 32 as output information. In the identification device 100, the identification processing unit 115 identifies which of the eight diseases the lesion captured in the input medical image corresponds to based on the output information output from the classifier 30, and the result output unit 116 outputs the discrimination result.

[0072] In the above embodiment, the control units 11 and 21 functioned as the respective parts shown in Figures 1, 11, or 12 by having the CPU execute a program stored in the ROM or storage unit 12. However, the control units 11 and 21 may be dedicated hardware. Dedicated hardware includes, for example, single circuits, complex circuits, programmed processors, ASICs (Application Specific Integrated Circuits), FPGAs (Field-Programmable Gate Arrays), or combinations thereof. If the control units 11 and 21 are dedicated hardware, each function of each part may be realized by separate hardware, or the functions of each part may be realized together by a single piece of hardware. Furthermore, some of the functions of each part may be realized by dedicated hardware, and other parts by software or firmware. In this way, the control unit 110 can realize the above-mentioned functions by hardware, software, firmware, or a combination thereof.

[0073] It is also possible to make an existing computer, such as a personal computer or cloud server, function as the identification device 100 or identification device 200 by applying the program that defines the operation of the identification device 100 or identification device 200 described above to the computer. Furthermore, the distribution method of such a program is arbitrary; for example, it may be distributed by storing it on a computer-readable recording medium such as a CD-ROM (Compact Disk ROM), DVD (Digital Versatile Disk), MO (Magneto Optical Disk), or memory card, or it may be distributed via a communication network such as the Internet.

[0074] Although preferred embodiments of the present invention have been described above, the present invention is not limited to the embodiments described above, and various modifications and substitutions can be made to the embodiments described above without departing from the scope of the claims. [Explanation of symbols]

[0075] 30... Identifier, 112... Target value setting unit, 113... Cutoff adjustment unit, 115... Identification processing unit, 100... Identification device

Claims

1. A classifier that identifies an object in a medical image by comparing an estimated value of the object with a cutoff value, wherein the classifier is configured such that the cutoff value is adjusted for each category so that the index value evaluated using evaluation data divided by category approaches a target value common to multiple categories, thereby identifying the object in the medical image. A medical diagnostic support device characterized by the following features.

2. The aforementioned index value is at least one of the sensitivity of the classifier and the specificity of the classifier. The cutoff value is adjusted on a category-by-category basis so that at least one of the sensitivity and specificity when the classifier identifies the target object by category using the evaluation data approaches the target value. The medical diagnostic support device according to feature 1.

3. The target values ​​set are the target value for the sensitivity of the classifier and the target value for the specificity of the classifier. The target value for the sensitivity is set to a value higher than the target value for the specificity. The medical diagnostic support device according to feature 2.

4. The cutoff value for a specific category among the aforementioned multiple categories is adjusted so that when the classifier identifies the object to be identified in the specific category using the evaluation data, the sensitivity and specificity approach values ​​that are higher than the target value for sensitivity and lower than the target value for specificity, respectively. The medical diagnostic support device according to claim 2 or 3, characterized by the above.

5. When a medical image is input, the classifier identifies the target to be identified by comparing the estimated value with the cutoff value corresponding to the category to which the target to be identified in the input medical image belongs, from among the cutoff values ​​adjusted for each category. A medical diagnostic support device according to any one of claims 1 to 3.

6. The aforementioned discriminator is When the aforementioned medical image is input, the category to which the identified object in the input medical image belongs is estimated from among the multiple categories. The object to be identified is identified by comparing the estimated cutoff value corresponding to the estimated category among the cutoff values ​​adjusted for each category, with the estimated value. The medical diagnostic support device according to feature 5.

7. The aforementioned category is information indicating at least one of the age, size, body part, and race of the identified subject. A medical diagnostic support device according to any one of claims 1 to 3.

8. The classifier identifies the quality of the object to be identified by comparing an estimated value of the quality of the object to be identified in the medical image with the cutoff value. A medical diagnostic support device according to any one of claims 1 to 3.

9. A device for generating a classifier that identifies an object in a medical image by comparing an estimated value with a cutoff value, A target value setting unit that sets target values ​​common to multiple categories, A cutoff adjustment unit adjusts the cutoff value for each category so that the index value evaluated using evaluation data divided by category approaches the target value set by the target value setting unit, A discriminator generating device characterized by comprising the following features.

10. A classifier that identifies an object in a medical image by comparing an estimated value of the object with a cutoff value, wherein the classifier is configured such that the cutoff value is adjusted for each category so that the index value evaluated using evaluation data divided by category approaches a target value common to multiple categories, thereby identifying the object in the medical image. A medical diagnostic support method characterized by the following.

11. Computers, A classifier that identifies an object in a medical image by comparing an estimated value of the object with a cutoff value, wherein the classifier is configured such that the cutoff value is adjusted for each category so that the index value evaluated using evaluation data divided by category approaches a target value common to multiple categories, and means for identifying the object in the medical image. A program designed to function as such.

Images