JP2020048783A - Information collection system and information collection device - Google Patents

Abstract

To provide an information collection system and an information collection device for collecting useful data for learning.SOLUTION: An information collection system includes a first storage unit, an acquisition unit, a determination unit, and a second storage unit. The first storage unit stores information on an arithmetic result output by an intermediate layer in a first learned model including one or more intermediate layers that execute arithmetic processing to information that is input, and an output layer for outputting a diagnostic result on the basis of the arithmetic result of the intermediate layer. The acquisition unit acquires the arithmetic result output by the intermediate layer upon input of medical information into an input layer of the first learned model. The determination unit determines whether or not the arithmetic result acquired by the acquisition unit satisfies a condition with the information on the arithmetic result stored by the first storage unit. The second storage unit stores information on the medical information corresponding to the arithmetic result that is determined to satisfy the condition.SELECTED DRAWING: Figure 5

Description

  An embodiment of the present invention relates to an information collection system and an information collection device.

  An image processing technology using machine learning is known. In machine learning, for example, a learned model configured by a multilayer neural network is used. In this technique, a large amount of learning data including various images having different characteristics is required in order to increase the accuracy of the learned model.

  At the time of learning a learned model, if there is a large amount of monotonous learning data having a high degree of similarity, the accuracy of determination may be reduced when a rare case image is input.

US Patent Publication No. 2016/0361025 U.S. Pat. No. 9,208,672 US Patent Publication No. 2017/0177798

  The problem to be solved by the present invention is to collect useful learning data.

  The information collection system according to the embodiment includes a first storage unit, an acquisition unit, a determination unit, and a second storage unit. The first storage unit is a first learned model including at least one intermediate layer that performs an arithmetic process on input information and an output layer that outputs a diagnosis result based on the arithmetic result of the intermediate layer. The information about the operation result output from the intermediate layer is stored. The acquisition unit acquires an operation result output by the intermediate layer in response to input of medical information to an input layer of the first learned model. The determining unit determines whether or not the calculation result obtained by the obtaining unit satisfies a condition between the calculation result and the information on the calculation result stored in the first storage unit. The second storage unit stores information on the medical information corresponding to a calculation result determined to satisfy the condition.

FIG. 1 is a diagram illustrating an example of the configuration of the information collection system according to the first embodiment. FIG. 2 is a diagram for explaining the flow of the learning process according to the first embodiment. FIG. 3 is a diagram illustrating an example of a distribution of calculation results according to the first embodiment. FIG. 4 is a diagram illustrating an example of the activation function according to the first embodiment. FIG. 5 is a diagram illustrating an example of functional blocks of the information collection system according to the first embodiment. FIG. 6 is a diagram illustrating an example of a flow of a learning process according to the first embodiment. FIG. 7 is a diagram illustrating an example of the flow of the determination process according to the first embodiment. FIG. 8 is a diagram illustrating an example of the calculation result information according to the first embodiment. FIG. 9 is a diagram illustrating an example of the statistical data according to the first embodiment. FIG. 10 is a diagram illustrating an example of useful data according to the first embodiment. FIG. 11 is a diagram illustrating an example of the model data according to the first embodiment. FIG. 12 is a diagram illustrating an example of a flow of an information collection process according to the first embodiment. FIG. 13 is a diagram illustrating an example of a reception screen according to the first embodiment. FIG. 14 is a flowchart illustrating an example of the acquisition process according to the first embodiment. FIG. 15 is a flowchart illustrating an example of the determination process according to the first embodiment. FIG. 16 is a flowchart illustrating an example of a reception process according to the first embodiment. FIG. 17 is a diagram illustrating an example of a functional block of the information collection system according to the second embodiment. FIG. 18 is a diagram illustrating an example of useful data according to the second embodiment. FIG. 19 is a diagram illustrating an example of functional blocks of the information collection system according to the third embodiment. FIG. 20 is a diagram illustrating an example of a distribution of calculation results according to the fourth embodiment.

  Hereinafter, embodiments of an information collection system and an information collection device will be described in detail with reference to the drawings. The information collection system according to the present embodiment is useful for improving the accuracy of a learned model, such as information corresponding to a rare case, out of medical information input to a learned model in, for example, CAD (Computer Aided Diagnosis). Collect important information. The learned model in the present embodiment can be configured by, for example, a multilayer neural network. In the following, an intermediate operation result of the learned model may be referred to as “intermediate operation result” or simply “operation result”. In the following, “information collection” may be simply referred to as “collection”.

  For example, when a rare case is input to a trained model, a peculiar value is often output in the calculation result in the hidden layer, and as a result, the discrimination result may become invalid. For example, for input information to be determined as “lung cancer”, a determination result may be output such that the accuracy of “pulmonary nodule” is greater than the accuracy of “lung cancer”. In this case, the output determination result may not be useful as support information for a doctor to diagnose whether it is “lung cancer” or “pulmonary nodule”. In the following, the “rare case” is not limited to a rare case, and includes medical information including a parameter that is not a rare case but is considered to be out of the learned range in the learned model.

  For example, when learning a trained model, if there is a large amount of monotonous training data having a high degree of similarity, the robustness to data or the like corresponding to a rare case is impaired. The discrimination accuracy for the data may be reduced. Therefore, in order to improve the accuracy of the learned model, it is useful to prepare learning data in which image data corresponding to a rare case is associated with correct answer data.

  Note that, for example, in a multilayer neural network, it is often not possible to clearly specify which parameter of the learning data has been reflected in the learned model, that is, the extent of the learned range. Therefore, it may be difficult to determine whether the medical information input to the learned model corresponds to a rare case in the learned model. Therefore, in the present embodiment, a configuration will be described in which medical image data corresponding to a rare case is collected using a calculation result output from an intermediate layer in a trained model. In the following, the determination that the medical image data is “corresponding to a rare case” is an example of that the medical information “satisfies the condition”.

(First embodiment)
First, a first embodiment will be described. In the first embodiment, an information collection system 1 including an information collection device 10 will be described as an example.

  FIG. 1 is a diagram illustrating an example of the configuration of the information collection system according to the first embodiment. As shown in FIG. 1, the information collection system 1 according to the present embodiment includes an information collection device 10, a group of medical image diagnostic devices 20, an image storage device 21, a learning device 30, and a terminal 40.

  As shown in FIG. 1, the information collection device 10, the medical image diagnostic device group 20, the learning device 30, and the terminal 40 are communicably connected to each other via a network NW. Although FIG. 1 illustrates the information collection system 1 including one information collection device 10, one image storage device 21, one learning device 30, and one terminal 40, the number of each device is arbitrary. Further, the information collection system 1 directly or indirectly communicates with each server device (not shown) included in other systems in the hospital such as a hospital accounting system (not shown) and an electronic medical record system (not shown). , May be communicably connected to each other. Further, an apparatus described below as a single computer device may be realized by a server group (cloud) connected via a network NW.

  The information collection device 10 collects useful information for improving the accuracy of the learned model, such as information corresponding to a rare case, from the medical information input to the learned model. The information collection device 10 determines whether the medical information is useful information, for example, by using a calculation result such as a feature amount output by an intermediate layer of the trained model at the time of the discriminating process by the learning device 30. . The information collection device 10 collects information on medical information determined to be useful information as needed. The information collection device 10 outputs information on the collected medical information at an arbitrary timing, and receives, for example, input of correct answer data from a CAD administrator (not shown). For example, the information collection device 10 is realized by a computer device such as a workstation.

  The medical image diagnostic apparatus group 20 is a plurality of apparatuses (modalities) that collect medical information from a subject. The medical information processed as image data is also described as medical image data. As shown in FIG. 1, the information collection system 1 according to the present embodiment includes a plurality of modalities XXX, XXY, and YYY as a medical image diagnostic apparatus group 20. Hereinafter, when a plurality of modalities XXX, XXY, and YYY are expressed without distinction, they may be simply expressed as “modality”.

  For example, the medical image diagnostic apparatus group 20 collects medical image data from a subject and transmits the collected medical image data to the learning apparatus 30. An example of the medical image diagnostic apparatus group 20 is an X-ray diagnostic apparatus, but an X-ray CT (Computed Tomography) apparatus, a SPECT (Single Photon Emission Computed Tomography) apparatus, and a SPECT apparatus and an X-ray CT apparatus are integrated. Other medical image diagnostic apparatuses such as a SPECT-CT apparatus and an ultrasonic diagnostic apparatus may be used. In FIG. 1, the information collection system 1 has three modalities as the medical image diagnostic apparatus group 20, but the number of modalities of the information collection system 1 according to the present embodiment is not limited to this.

  The image storage device 21 is a database (DB) or the like that stores medical image data collected by the medical image diagnostic device group 20 via the network NW. For example, in FIG. 1, the image storage device 21 acquires medical image data from the medical image diagnostic device group 20, and stores the acquired medical image data in a memory provided inside or outside the device. Further, the image storage device 21 may further store correct answer data including information such as a case corresponding to the medical image data. For example, the image storage device 21 is realized by a computer device such as a server device.

  The learning device 30 acquires medical image data from the medical image diagnostic device group 20 and the like via the network NW, and executes various processes using the acquired medical image data. For example, the learning device 30 acquires medical image data from the medical image diagnostic device group 20 or the image storage device 21 via the network NW. The learning device 30 inputs the acquired medical image data to the trained model, performs a discrimination process, and outputs a corresponding case, a reliability indicating a certainty corresponding to the case, and the like. The learning device 30 performs learning on the learned model using medical image data that is input information to the learned model and learning data including, for example, correct data such as a case as a data set. For example, the learning device 30 is realized by a computer device such as a workstation.

  In the present embodiment, the learning device 30 includes, for example, three learned models 341, 342, and 343, as illustrated in FIG. 1, but the number of learned models according to the present embodiment is not limited thereto. Further, the learned models 341, 342, and 343 may be implemented in different learning devices. The learned models 341, 342, and 343 are learned models used in CAD used in a hospital.

  The terminal 40 is installed, for example, in an examination room of a hospital, and outputs a determination result in the learning device 30 to a user such as a doctor. The terminal 40 is, for example, a desktop computer as shown in FIG. 1, but is not limited thereto, and may be a portable computer such as a smartphone or a tablet.

  Note that the location where the information collection device 10, the group of medical image diagnostic devices 20, the image storage device 21, the learning device 30, and the terminal 40 are installed is arbitrary as long as connection is possible via the network NW. For example, the learning device 30 may be installed in a hospital different from the hospital where the information collecting device 10 is installed. That is, the network NW may be configured by a local network closed in the hospital, or may be a network via the Internet.

  Next, information collection processing in the information collection system 1 will be described with reference to FIG. FIG. 2 is a diagram for explaining the flow of the learning process according to the first embodiment. As shown in FIG. 2, the learning device 30 receives, for example, an input of medical image data 91 such as X-ray image data from the medical image diagnostic device group 20. The learning device 30 inputs the medical image data 91 to a learned model formed by a multilayer neural network. As shown in FIG. 2, the multilayer neural network according to the present embodiment has a total (L-2) layer of an input layer (l = 1) and an intermediate layer (l = 2, 3,..., L-1). ) And an output layer (l = L).

  The learning device 30 uses, for example, pixel values of pixels constituting the medical image data 91 as each component xp (p = 1, 2,..., N) of the input signal x. In the intermediate layers (l = 2, 3,..., L-1 layers) subsequent to the input layer, the arithmetic processing is sequentially performed on the input signals, whereby each of the (L-2) intermediate layers is processed. The output z can be calculated. Each intermediate layer has a plurality of filters. The output layer L can calculate the output yn (n = 1, 2,..., M) by executing the arithmetic processing on the output z of each layer. The output yn indicates, for example, the reliability for each case. That is, in the example shown in FIG. 2, the output yn outputs the reliability for each of the M cases. The arithmetic processing in each layer can use, for example, a known neural network technique, and thus a detailed description is omitted.

  The learning device 30 outputs, to the terminal 40, information on the determination result based on the output yn in the output layer L. The terminal 40 outputs information on the determination result output from the learning device 30 as shown in FIG. 2, for example. In the example shown in FIG. 2, a determination result indicating that the probability of a lung nodule is 60% and the probability of lung cancer is 30% is output. Note that FIG. 2 illustrates an example in which the terminal 40 displays the reliability of each case as the information on the determination result. However, the present invention is not limited to this, and the terminal 40 simply displays only the case with the highest reliability. May be displayed. For example, the terminal 40 may display "Possible lung nodules". Further, for example, the terminal 40 may display, in the medical image data 91, an area in which a lung nodule is suspected with 60% accuracy.

  In the present embodiment, the information collection device 10 acquires a calculation result or the like in the intermediate layer of the learned model of the learning device 30 and collects medical information recognized as falling under the rare case. In the present embodiment, the information collecting apparatus 10 acquires the operation result in, for example, the pooling layer among the intermediate layers of the learned model configured by the multilayer neural network, but is not limited to this, and the convolution layer, the ReLU (Rectified) A calculation result in each layer such as a (Liner Unit) layer may be obtained.

  In the present embodiment, the information collection device 10 sequentially acquires and stores the calculation results output in the intermediate layer, for example, at the time of learning by the learning device 30 and at the time of determination processing. In addition, the information collection device 10 determines whether the newly obtained calculation result is based on the rare case, for example, by comparing the newly obtained calculation result with the stored calculation result.

  FIG. 3 is a diagram illustrating an example of a distribution of calculation results according to the first embodiment, with regard to a criterion for determining whether an input image is a rare case. In FIG. 3, the horizontal axis indicates the value of the calculation result, and the vertical axis indicates the frequency of appearance of the calculation result corresponding to the value. For example, as shown in FIG. 3, when the frequency of appearance of the calculation result has a normal distribution, the calculation result within the range of ± 3σ from the average value μ (hereinafter, the standard deviation may be referred to as “σ”). 97a is determined not to be based on the rare case. On the other hand, the calculation results 97b and 97c deviating from the average value μ by ± 3σ are determined to be based on the rare case. The criterion for determining whether or not the calculation result is based on the rare case is an example, and may be determined based on another criterion. Further, when the frequency of appearance of the calculation result does not take a normal distribution, whether or not the calculation result is a calculation result based on a rare case may be similarly determined based on the distance from the average value μ.

  In many cases, the feature of the calculation result based on the rare case is not propagated to the output layer by the activation function. FIG. 4 is a diagram illustrating an example of the activation function according to the first embodiment. FIG. 4 shows an example of a graph of the ReLU function often used in an intermediate layer of a trained model formed by a multilayer neural network. For example, even if the calculation result based on the rare case deviates by more than −3σ from the average value μ of the calculation result, the output becomes 0 when input to the ReLU function as shown in FIG. Therefore, since the characteristics when the rare case is input to the input layer are not often reflected in the determination result in the output layer, it is not possible to determine whether the input information is a rare case based on the determination result. difficult. Therefore, in the present embodiment, a configuration will be described in which it is determined whether or not input information is a rare case using a calculation result in the intermediate layer.

  In such a configuration, the information collection device 10 in the present embodiment acquires the calculation result output from the intermediate layer (l = 2, 3,..., L-1 layer) of the learned model of the learning device 30. I do. Then, the information collection device 10 determines whether the medical image data is useful as learning data based on the calculation result, and collects information on the useful medical image data.

  FIG. 5 is a diagram illustrating an example of functional blocks of the information collection system according to the first embodiment. As shown in FIG. 5, the information collection system 1 includes an information collection device 10, a group of medical image diagnostic devices 20, an image storage device 21, a learning device 30, and a terminal 40. As shown in FIG. 5, the learning device 30 includes a communication I / F (Interface) 31, an input I / F 32, a display 33, a storage circuit 34, and a processing circuit 35. Note that the functions of the processing circuit 35 of the learning device 30 may be implemented on a cloud, and the content stored by the storage circuit 34 may be stored on the cloud.

  The communication I / F 31 is connected to the processing circuit 35, and transmits and communicates various data between the information collection device 10, the medical image diagnostic device group 20, the image storage device 21, and the like connected via the network NW. Control. For example, the communication I / F 31 is realized by a network card, a network adapter, a NIC (Network Interface Controller), and the like.

  The input I / F 32 is connected to the processing circuit 35 and converts an input operation received from, for example, a CAD administrator into an electric signal and outputs the electric signal to the processing circuit 35. For example, the input I / F 32 is a switch button, a mouse, a keyboard, a touch panel, or the like.

  The display 33 is connected to the processing circuit 35 and displays various information and various image data output from the processing circuit 35. For example, the display 33 is realized by a liquid crystal monitor, a CRT (Cathode Ray Tube) monitor, a touch panel, or the like.

  The storage circuit 34 is connected to the processing circuit 35 and stores various data. For example, the storage circuit 34 is realized by a semiconductor memory element such as a random access memory (RAM) or a flash memory, a hard disk, an optical disk, or the like. In the present embodiment, the storage circuit 34 stores a program for causing a circuit included in the learning device 30 to realize its function. Further, for example, the storage circuit 34 stores a plurality of learned models 341, 342, and 343, as shown in FIG. The storage circuit 34 may be realized by a server group (cloud) connected to the learning device 30 via the network NW.

  The learned models 341, 342, and 343 are learned models incorporated in the discrimination processing performed by the processing circuit 35 to discriminate cases based on input medical image data. The learned models 341, 342, and 343 include, for example, as shown in FIG. 2, an input layer (l = 1), an intermediate layer (l = 2, 3,..., L−1), and an output layer (l = L). ) Is constituted by a multilayer neural network composed of L layers. The learned models 341, 342, and 343 are generated by a model generation function 352 described later.

  Returning to FIG. 5, the processing circuit 35 executes a control function 351, a model generation function 352, a determination function 353, and an output function 354. Here, for example, each processing function of a control function 351, a model generation function 352, a determination function 353, and an output function 354, which are components of the processing circuit 35 shown in FIG. 5, is a storage circuit in the form of a program executable by a computer. 34. The processing circuit 35 is a processor that reads out each program from the storage circuit 34 and executes the read out program to realize a function corresponding to each program. In other words, the processing circuit 35 in a state where each program is read has each function shown in the processing circuit 35 of FIG.

  Note that all the processing functions of the control function 351, the model generation function 352, the determination function 353, and the output function 354 may be recorded in the storage circuit 34 in the form of one program executable by a computer. For example, such a program is also called a learning program. In this case, the processing circuit 35 reads the learning program from the storage circuit 34 and executes the read learning program to realize a control function 351, a model generation function 352, a determination function 353, and an output function 354 corresponding to the learning program. .

  The control function 351 in the processing circuit 35 controls various operations related to operations on the learning device 30, communication with other devices, and data acquisition from other devices. For example, the control function 351 acquires learning data including a set of medical image data and correct answer data from, for example, the image storage device 21 in response to a trigger for starting the model generation processing, and outputs the learning data to the model generation function 352. . Further, for example, the control function 351 acquires medical image data from, for example, the image storage device 21 in response to a trigger for starting the determination processing, and outputs the medical image data to the determination function 353. Note that the trigger for starting the learning process and the determination process may be, for example, an operation received from a CAD administrator via the input I / F 32 or may be received from the terminal 40 or the like via the communication I / F 31. .

  The model generation function 352 executes a process related to generation of the learned models 341, 342, and 343. For example, when the model generation function 352 first receives an input of learning data from the control function 351, the model generation function 352 determines correct data such as a case from medical image data included in the learning data based on the learning data. Let the model learn. FIG. 6 is a diagram illustrating an example of a flow of a learning process according to the first embodiment.

  As shown in FIG. 6, for example, the model generation function 352 adjusts model parameters by using medical image data included in the learning data as input data and correct data included in the learning data as output data. Then, a learned model that approximates a function for determining a case or the like is generated. When the learned model to be generated is a learned model based on a neural network, the parameter to be adjusted is, for example, at least one of a weight and a bias. Then, the model generation function 352 stores the generated learned models 341, 342, and 343 in the storage circuit 34. The learned models 341, 342, and 343 output the calculation results in the hidden layer. Note that when new learning data is acquired, the model generation function 352 may perform a re-learning process on the learned models 341, 342, and 343 in order to increase the discrimination accuracy.

  Note that the model generation function 352 may use medical image data included in the learning data as input-side data, or may use data based on medical image data as input-side data. For example, as shown in FIG. 6, the model generation function 352 performs pre-processing (for example, image processing using a spatial filter or a convolution filter) on medical image data, and performs input processing as image data after pre-processing. Data may be used as input data.

  The types of the image processing filters include, for example, a moving average (smoothing) filter, a Gaussian filter, a median filter, a recursive filter, and a bandpass filter. The parameters of the image processing filter include, for example, the strength of each filter, the vertical and horizontal sizes of the moving average (smoothing) filter, the Gaussian filter and the median filter, the number of frames used in the recursive filter, and the Weight, setting of a frequency in a band-pass filter, and the like.

  The determination function 353 performs a determination process on the input medical image data based on the learned model 341, 342, or 343. Specifically, the discriminating function 353 executes a discriminating process of discriminating a case, for example, by inputting pre-processed medical image data to the trained model 341, 342, or 343. Then, the determination function 353 outputs the determination result output from the learned model 341, 342, or 343 to the output function 354. FIG. 7 is a diagram illustrating an example of the flow of the determination process according to the first embodiment.

  For example, the discriminating function 353 first reads the learned model 341, 342, or 343 from the storage circuit 34. Next, as shown in FIG. 7, the discriminating function 353 inputs the medical image data to the trained models 341, 342, or 343, and executes a discriminating process for discriminating a case. The learned model 341, 342, or 343 outputs a determination result based on the input medical image data. Further, the learned model 341, 342, or 343 outputs a calculation result in the hidden layer. Note that, as shown in FIG. 7, the discrimination function 353 may perform pre-processing on medical image data as in the learning process, and use the pre-processed image data as input-side data.

  The output function 354 outputs the result of the determination by the determination function 353. Specifically, the output function 354 outputs the determination results output from the learned models 341, 342, and 343 to the terminal 40 illustrated in FIG.

  Next, the information collection device 10 includes a communication I / F 11, an input I / F 12, a display 13, a storage circuit 14, and a processing circuit 15, as shown in FIG. Note that the functions of the processing circuit 15 of the information collection device 10 may be implemented on a cloud, and the contents stored by the storage circuit 14 may be stored on the cloud.

  The communication I / F 11 is connected to the processing circuit 15 and controls transmission and communication of various data performed with the learning device 30 and the like connected via the network NW. For example, the communication I / F 11 is realized by a network card, a network adapter, or the like.

  The input I / F 12 is connected to the processing circuit 15, converts an input operation received from a CAD administrator into an electric signal, and outputs the electric signal to the processing circuit 15. The input I / F 12 receives an input operation such as mode selection by a CAD administrator, for example. For example, the input I / F 12 is a switch button, a mouse, a keyboard, a touch panel, or the like.

  The display 13 is connected to the processing circuit 15 and displays various information output from the processing circuit 15. The display 13 is realized by, for example, a liquid crystal monitor, a CRT monitor, a touch panel, or the like.

  The storage circuit 14 is connected to the processing circuit 15 and stores various data. For example, the storage circuit 14 is realized by a semiconductor memory element such as a RAM and a flash memory, a hard disk, an optical disk, and the like. In the present embodiment, the storage circuit 14 stores a program for a circuit included in the information collection device 10 to realize its function. In addition, for example, as illustrated in FIG. 5, the storage circuit 14 stores operation result information 141, statistical data 142, useful data 143, and model data 144. The storage circuit 14 may be realized by a server group (cloud) connected to the information collection device 10 via a network NW. Note that the storage circuit 14 is an example of a first storage unit.

  The calculation result information 141 stores the calculation result output from the intermediate layer of the learned model 341 in the learning process or the determination process in the learning device 30. For example, the calculation result information 141 stores the calculation result output in the learning process illustrated in FIG. 6 and the calculation result output in the determination process illustrated in FIG. The calculation result information 141 is stored by, for example, a control function 151 described later. The calculation result information 141 is an example of information stored in the first storage unit.

  FIG. 8 is a diagram illustrating an example of the calculation result information according to the first embodiment. For example, as shown in FIG. 8, the calculation result information 141 holds one table for each filter included in each intermediate layer of each learned model. The calculation result information 141 illustrated in FIG. 8 stores, for example, a calculation result output from the “2” -th filter of the “1” -th intermediate layer of the learned model with the model ID “M001”.

  For example, the calculation result information 141 is information in which “acquisition date and time”, “image ID”, “processing category”, and “calculation result” are associated with each other, as shown in FIG. “Acquisition date and time” indicates the date and time when the operation result was acquired. “Image ID” indicates an identifier (IDentifier) that uniquely identifies medical image data input when the operation result is output. The “processing category” indicates whether the calculation result is output during “learning” or output during “discrimination”. The “calculation result” indicates a calculation result output from the filter of the intermediate layer. The calculation result accumulated in the calculation result information 141 is used, for example, for generating statistical data described later. In FIG. 8, for example, at “at the time of learning” using “image ID: XXX1”, the output from the “2” -th filter of the “1” -th intermediate layer at “15:20 on June 2, 2018” It is stored that the calculated result is “0.604”. In FIG. 8, for example, in “determination time” for “image ID: XXY1”, “1” -th intermediate layer outputs “2” -th filter at “10:00” on June 30, 2018. It is stored that the calculated result is “0.317”.

  Note that the items shown in FIG. 8 are examples, and the calculation result information 141 may further store other information such as a “determination result”, and some of the items shown in FIG. 8 are not stored. Such a configuration may be adopted. The calculation result information 141 may hold one table for each calculation result obtained in all filters of one layer, or may hold all calculation results in one table. The same applies to each table described below.

  The statistical data 142 is statistical data such as an average value μ and a standard deviation σ of the calculation results calculated by performing statistical processing on the calculation results stored in the calculation result information 141. The statistical data 142 is stored by the control function 151, for example. Note that the statistical data 142 is an example of information stored in the first storage unit.

  FIG. 9 is a diagram illustrating an example of the statistical data according to the first embodiment. For example, as shown in FIG. 9, the statistical data 142 holds one table for each learned model. Statistical data 142 shown in FIG. 9 stores, for example, data calculated by performing statistical processing on calculation results output from each filter of each intermediate layer of the learned model with the model ID “M001”.

  For example, as shown in FIG. 9, the statistical data 142 includes “intermediate layer No.” and “filter No.”, “record number”, “average value”, “−3σ” and “+ 3σ”. This is the associated information. Here, “intermediate layer No.” and “filter No.” in FIG. 9 indicate information for uniquely identifying the layer and the filter to which the operation result is output. The “record number” indicates the number of calculation results output from the filter of the corresponding layer and stored in the calculation result information 141. “Average value” indicates the average value μ of the calculation result. "-3 [sigma]" and "+3 [sigma]" indicate values that are ± 3 [sigma] apart from the average value μ of the calculation result.

  In FIG. 9, “average value”, “−3σ”, and “+ 3σ” are used, for example, as threshold values for determining whether or not the obtained calculation result is based on a rare case. For example, when the calculation result deviates from the average μ by “−3σ” or “+ 3σ”, the determination result 152 described later determines that the calculation result is based on the rare case. In FIG. 9, for example, when the calculation result output from the “1” th filter of the “1” th intermediate layer is smaller than “0.431” or larger than “0.674”, the calculation result depends on the rare case. Is determined. Similarly, in FIG. 9, for example, when the operation result output from the “2” th filter of the “1” th intermediate layer is smaller than “−0.932” or larger than “0.450”, Is determined to be due to the rare case.

  The useful data 143 stores information about medical image data determined to be useful. The useful data 143 is stored by the determination function 152, for example. Note that the useful data 143 is an example of information stored in the second storage unit.

  FIG. 10 is a diagram illustrating an example of useful data according to the first embodiment. For example, as shown in FIG. 10, the useful data 143 includes “image ID”, “modality ID”, “patient ID”, “model ID”, “correct answer data”, “doctor ID”, This is information that is associated with “relevant reason”. Here, “image ID” in FIG. 10 indicates an identifier that uniquely identifies medical image data input when the calculation result is output. “Modality ID” indicates information for uniquely identifying the modality that collected the medical image data. “Patient ID” indicates information for uniquely identifying a patient or the like from which medical image data has been collected. “Model ID” indicates information for uniquely identifying a learned model used in the discrimination processing of medical image data. "Correct data" and "doctor ID" indicate correct data for medical image data and information for uniquely identifying a doctor who has input the correct data, respectively. “Relevant reason” indicates the reason that the medical image data is determined to correspond to useful information.

  For example, the table shown in FIG. 10 indicates that the medical image data of the image ID “XXX1” collected from the patient of the patient ID “0001” by the modality of the modality ID “XXX” is the learned model of the model ID “M001”. Indicates that it is determined that there is a “suspect of a rare case” in the determination process by Further, the table shown in FIG. 10 indicates that the value of the n-th filter of the m-th layer of the trained model “M001” satisfies the condition in the determination processing using the model ID “M001” for the medical image data of the image ID “XXX1”. Is shown. Note that the table shown in FIG. 10 includes only the records related to the calculation result determined to have “rare case suspicion”, but is not limited thereto, and further includes the records related to the calculation result determined not to be the rare case. May be included. Note that the learned models of the model IDs “M001”, “M002”, and “M003” correspond to, for example, the learned models 341, 342, and 343 shown in FIG. 1, respectively. The reason “applicable to the target range” described in the record of the model ID “M003” will be described later in detail.

  The model data 144 stores, for example, information about a learned model used for preprocessing for generating a learning model from medical image data. FIG. 11 is a diagram illustrating an example of the model data according to the first embodiment. For example, as shown in FIG. 11, the model data 144 is information in which "model ID", "input format", "filter type", and "corresponding modality" are associated. In FIG. 11, “model ID” indicates information for uniquely identifying a learned model used in the medical image data determination process. The “input format” indicates information on a format such as the size of medical image data input to the model. “Filter type” indicates information about a filter used for pre-processing of medical image data when medical image data is input to the learned model.

  For example, the first record of the model data 144 shown in FIG. 11 indicates that the model with the model ID “M001” applies “format A” and “filter A” and corresponds to all modalities. In the third record of the model data 144 shown in FIG. 11, the model with the model ID “M003” applies “format C” and “filter C”, and corresponds to the modalities of “CT” and “MR”. To do so.

  Returning to FIG. 5, the processing circuit 15 executes a control function 151, a determination function 152, an output function 153, and a reception function 154. Here, the control function 151 is an example of an acquisition unit. The determination function 152 is an example of a determination unit. The output function 153 is an example of an output unit. The reception function 154 is an example of a reception unit. Here, for example, each processing function of the control function 151, the determination function 152, the output function 153, and the reception function 154, which are components of the processing circuit 15 shown in FIG. It is recorded in. The processing circuit 15 is a processor that reads out each program from the storage circuit 14 and executes the read out program to realize a function corresponding to each program. In other words, the processing circuit 15 in a state where each program is read has each function shown in the processing circuit 15 of FIG.

  Note that all the processing functions of the control function 151, the determination function 152, the output function 153, and the reception function 154 may be recorded in the storage circuit 14 in the form of one program executable by a computer. For example, such a program is also called an information collection program. In this case, the processing circuit 15 reads out the information collection program from the storage circuit 14, and executes the read information collection program to execute the control function 151, the determination function 152, the output function 153, and the reception function 154 corresponding to the information collection program. Realize.

  A control function 151 in the processing circuit 15 controls various operations related to operations on the information collection device 10, communication with other devices, and data acquisition from other devices. For example, the control function 151 acquires a calculation result corresponding to the medical image data input to the learning device from the learning device 30 as needed, and outputs the result to the determination function 152, for example. Further, the control function 151 performs statistical processing on the obtained calculation result and stores the result in the statistical data 142.

  Further, the control function 151 acquires, for example, a patient ID of a patient or the like from which medical image data input to the learning device is collected. For example, when the information collection system 1 has a plurality of modalities XXX, XXY, and YYY as shown in FIG. 1, the control function 151 further acquires each modality ID. Further, for example, when the learning device 30 has a plurality of learned models 341, 342, and 343 as illustrated in FIG. 1, the control function 151 further acquires the model ID of each learned model.

  In the present embodiment, the information collection process is executed as needed, for example, when the control function 151 detects the start of the learning process or the start of the determination process in the learning device 30.

  The determination function 152 determines whether the medical image data accompanying the calculation result is useful. For example, when the determination function 152 first receives the output of the calculation result from the control function 151, the determination function 152 refers to the statistical data 142 and determines that the calculation result ranges from “−3σ” to “+ 3σ” as shown in FIG. It is determined whether or not a deviation occurs, that is, whether or not the calculation result is based on a rare case. FIG. 12 is a diagram illustrating an example of a flow of an information collection process according to the first embodiment. For example, as illustrated in FIG. 12, the determination function 152 compares the result of the comparison between the intermediate calculation result received from the control function 151 and the statistical data 142 into useful data 143 as illustrated in FIG. Store.

  The determination function 152 specifies, as useful data, for example, input data in which the calculation result output from the intermediate layer of the learned model deviates from the range of “−3σ to + 3σ”, that is, input data corresponding to a rare case. However, the embodiment is not limited to this. For example, even if the calculation result is within the range of “−3σ to + 3σ”, the determination function 152 is considered to be useful in generating learning data. A configuration may be adopted in which input data whose calculation result falls within the range is determined to be useful data. For example, when the number of samples of the learning data in which the calculation result falls within the range of “from −σ to the average value μ” is small, the determination function 152 determines that the data whose calculation result falls within the range is useful data. The configuration may be as follows. For example, as illustrated in FIG. 10, the determination function 152 determines that the calculation result of the n-th filter in the m-th layer is “applicable to the target range” in the useful data 143 for the medical image data of the patient ID “0005” It may be stored.

  The output function 153 outputs information on medical image data collected as useful information. The output function 153 outputs, for example, a screen for accepting the input of correct answer data to the medical image data to the display 13 of the information collection device 10. In the following, a case where correct data is input from a CAD administrator who refers to the display 13 will be described, but the output destination of the screen is not limited to this. For example, when the CAD administrator operates another computer such as the terminal 40, the output function 153 may output a screen to the computer via the communication I / F 11.

  FIG. 13 is a diagram illustrating an example of a reception screen according to the first embodiment. As shown in FIG. 13, for example, a reception screen 80 displayed on the display 13 includes a display area 81 on which information on medical image data is displayed, an input area 82 for receiving input of correct answer data, a decision button 83, And image data 91. The display area 81 has information for specifying medical image data, such as an image ID, a modality ID, and a patient ID. The medical image data 91 is medical image data of the image ID displayed in the display area 81. The input area 82 is an area for receiving an input of correct answer data of a case or the like by a CAD administrator, for example. Note that the correct answer data is an example of the determination result.

  The reception function 154 receives information on the correct answer data input on the reception screen 80. When the reception function 154 detects that the determination button 83 has been pressed on the reception screen 80, the reception function 154 acquires information on the case input to the input area 82. The reception function 154 stores the acquired information on the case as correct data in the useful data 143 in association with the image ID, modality ID, patient ID, model ID, and the like displayed on the reception screen 80. Further, the reception function 154 specifies the doctor who has input the information regarding the case, and stores the doctor ID in the useful data 143.

  For example, as shown in FIG. 10, the reception function 154 converts the correct answer data “lung cancer” and the doctor ID “A001” of the doctor who input the correct answer data into the image ID “XXX1”, the modality ID “XXX”, and the patient ID “0001”. And stored in the useful data 143.

  Next, a procedure of a process performed by the information collection system 1 according to the first embodiment will be described with reference to FIGS. FIG. 14 is a flowchart illustrating an example of the acquisition process according to the first embodiment. Here, each process from step S101 to S120 in FIG. 14 is realized, for example, by the processing circuit 15 reading out a program corresponding to the control function 151 from the storage circuit 14 and executing it. Note that the acquisition process illustrated in FIG. 14 is executed as needed when, for example, the learning device 30 detects that the learning process or the determination process has been performed.

  For example, as illustrated in FIG. 14, in the information collection system 1 according to the present embodiment, first, the processing circuit 15 obtains the operation result of the intermediate layer from the learning device 30 via the network NW (step S101). Next, the processing circuit 15 stores the obtained calculation result in the calculation result information 141 (Step S102).

  Subsequently, the processing circuit 15 performs statistical processing on the operation result stored in the operation result information 141 (S103), and stores the result in the statistical data 142 (S104).

  Next, the processing circuit 15 determines whether or not processing for all filters has been completed (step S110). If the processing circuit 15 determines that the processing for all the filters has not been completed (No at Step S110), the processing circuit 15 proceeds to Step S101 and repeats the processing.

  On the other hand, when it is determined that the processing for all the filters has been completed (Yes at Step S110), the processing circuit 15 determines whether the processing for all the layers has been completed (Step S120). When the processing circuit 15 determines that the processing for all the layers has not been completed (No at Step S120), the processing circuit 15 proceeds to Step S101 and repeats the processing. On the other hand, when the processing circuit 15 determines that the processing for all the layers has been completed (Yes at Step S120), the processing ends.

  Next, a determination process based on the newly obtained calculation result will be described. FIG. 15 is a flowchart illustrating an example of the determination process according to the first embodiment. Here, steps S201, S220, and S230 in FIG. 15 are realized, for example, by the processing circuit 15 reading out a program corresponding to the control function 151 from the storage circuit 14 and executing it. Steps S202, S210, and S211 are realized, for example, by the processing circuit 15 reading a program corresponding to the determination function 152 from the storage circuit 14 and executing the program. The determination process illustrated in FIG. 15 is executed as needed when, for example, the information collection process illustrated in FIG. 14 is performed.

  For example, as shown in FIG. 14, in the information collection system 1 according to the present embodiment, first, the processing circuit 15 obtains the operation result of the intermediate layer from the learning device 30 via the network NW (step S201).

  Subsequently, the processing circuit 25 compares the obtained calculation result with the statistical data 142 (S202), and determines whether a condition for determining a rare case is satisfied, for example, whether the calculation result is equal to or greater than a threshold. (Step S210). When the processing circuit 15 determines that the determination result is not equal to or larger than the threshold (No at Step S210), the process proceeds to Step S220.

  On the other hand, when determining that the determination result is equal to or greater than the threshold (Yes at Step S210), the processing circuit 25 stores, for example, medical information corresponding to the calculation result in the useful data 143 as the determination result (Step S211). Then, the process proceeds to step S220.

  Next, the processing circuit 15 determines whether or not processing for all filters has been completed (step S220). When the processing circuit 15 determines that the processing for all the filters has not been completed (No at Step S220), the processing circuit 15 proceeds to Step S201 and repeats the processing.

  On the other hand, when it is determined that the processing for all the filters has been completed (Yes at Step S220), the processing circuit 15 determines whether the processing for all the layers has been completed (Step S230). When the processing circuit 15 determines that the processing for all the layers has not been completed (No at Step S230), the processing circuit 15 proceeds to Step S201 and repeats the processing. On the other hand, when the processing circuit 15 determines that the processing for all the layers has been completed (Yes at Step S230), the processing ends.

  Next, a process for accepting input of correct answer data for useful data will be described. FIG. 16 is a flowchart illustrating an example of a reception process according to the first embodiment. Here, steps S301, S302, and S320 in FIG. 16 are realized, for example, by the processing circuit 15 reading a program corresponding to the output function 153 from the storage circuit 14 and executing the program. Steps S310 and S311 are realized, for example, by the processing circuit 15 reading out a program corresponding to the reception function 154 from the storage circuit 14 and executing it. The reception process shown in FIG. 16 is executed, for example, in response to a trigger for starting the reception process. The trigger for starting the reception process may be, for example, an operation received from a CAD administrator via the input I / F 12 or may be received from the terminal 40 or the like via the communication I / F 11.

  For example, as shown in FIG. 16, in the information collection system 1 according to the present embodiment, first, the processing circuit 15 acquires a record stored in the useful data 143 (Step S301). Next, the processing circuit 15 generates a reception screen based on the obtained useful data, and outputs it to the display 13 or the like of the information collection device 10 (S302). Then, for example, the processing circuit 15 determines whether or not the input of the correct answer data has been received from the input I / F 12 (Step S310). When determining that the input of the correct data has not been received (No at Step S310), the processing circuit 15 waits until the input of the correct data is received.

  When determining that the input of the correct answer data has been received (Yes at Step S310), the processing circuit 15 stores the correct answer data in the useful data 143 (Step S311). After that, it moves to step S320.

  Next, the processing circuit 15 determines whether the processing for all useful data has been completed (step S320). When the processing circuit 15 determines that the processing for all the useful data has not been completed (No at Step S320), the processing circuit 15 proceeds to Step S301 and repeats the processing. On the other hand, when the processing circuit 15 determines that the processing for all the useful data has been completed (Yes at Step S320), the processing ends.

  As described above, the information collection system 1 according to the first embodiment can collect useful learning data. For example, the information collection system 1 according to the first embodiment can collect medical image data corresponding to a rare case and data determined to be useful for learning by a CAD administrator.

(Second embodiment)
In the first embodiment, the configuration has been described in which it is determined whether the medical information satisfies the condition based on the calculation result obtained from the intermediate layer of the learned model. The configuration for collecting unnecessary data is not limited to this. For example, when other medical image data is collected from a patient or the like from which data corresponding to a rare case is collected, even if the data itself does not correspond to a rare case, the correct answer of the data corresponding to the rare case is obtained. May be useful in identifying data. In the present embodiment, “patient or the like” is an example of a subject.

  Thus, in the present embodiment, a configuration will be described in which useful information is collected by specifying a patient or the like from which medical image data has been collected. FIG. 17 is a diagram illustrating an example of a functional block of the information collection system according to the second embodiment. In the following embodiments, the same parts as those shown in the drawings described above are denoted by the same reference numerals, and overlapping description will be omitted.

  As shown in FIG. 17, the information collection system 1a includes an information collection device 10a, a group of medical image diagnostic devices 20, an image storage device 21, a learning device 30, and a terminal 40. The information collection device 10a includes a communication I / F 11, an input I / F 12, a display 13, a storage circuit 14a, and a processing circuit 15a.

  In the present embodiment, the storage circuit 14a stores operation result information 141, statistical data 142, useful data 143a, and model data 144, as shown in FIG. The useful data 143a according to the present embodiment includes, in addition to information on medical image data having a “rare case suspicion” as shown in FIG. And information on medical image data having the same “patient ID”.

  FIG. 18 is a diagram illustrating an example of useful data according to the second embodiment. As illustrated in FIG. 18, the useful data 143a stores a fourth record having the same patient ID “0001” as the first record. In this case, the useful data 143a stores that the corresponding reason of the fourth record is “patient agreement”.

  Returning to FIG. 17, the processing circuit 15 executes a control function 151, a determination function 152, an output function 153, a reception function 154, and a collection function 155. Here, the collection function 155 is an example of a collection unit.

  The collection function 155 in the processing circuit 15 collects medical image data. For example, the collection function 155 acquires medical image data from the medical image diagnostic device group 20 or the image storage device 21. Further, the collection function 155 collects medical image data input from the learning device 30 to the input layer of the learned model 341, 342, or 343. At that time, the collection function 155 acquires, for example, the image ID of the medical image data, the patient ID of the patient from which the medical image data was collected, and the modality ID of the modality that collected the medical image data. Further, the collection function 155 may further acquire the model ID of the learned model to which the medical image data has been input.

  In addition, the collection function 155 determines, for example, whether a patient ID that matches the patient ID of the collected medical image data is stored in the useful data 143a. If the collection function 155 determines that a patient ID that matches the patient ID corresponding to the collected medical image data is stored in the useful data 143a, the information regarding the medical image data corresponding to the patient ID is used as useful data. 143a and store it.

  For example, when collecting the medical image data of the patient ID “0001”, the modality ID “YYY”, and the model ID “M003”, the collection function 155 determines that the matching patient ID “0001” is the first data of the useful data 143a shown in FIG. It is determined that the data is stored in one record. In this case, as shown in FIG. 18, the collection function 155 stores information about the collected medical image data in the fourth record of the useful data 143a.

  In the present embodiment, the medical image collection processing by the collection function 155 is executed as needed, for example, when the control function 151 executes the information collection processing.

  As described above, the information collection system 1 according to the second embodiment includes, in addition to useful data such as medical image data corresponding to a rare case and data determined to be useful for learning by a CAD manager, Medical image data collected from the same subject as the data can be collected. Thus, it is possible to support the identification of a case in medical image data corresponding to a rare case, data determined by a CAD administrator to be useful for learning, and the like.

  Note that the collection function 155 in the present embodiment may anonymize the data corresponding to the patient ID when storing the information on the medical image data with the matching patient ID in the useful data 143a. For example, when registering information about medical image data in the useful data 143a, the collection function 155 may replace the patient ID with a random character string so that the patient cannot be identified. In addition, the collection function 155 replaces, for example, a similar character string with “k” patient IDs corresponding to other medical image data with matching cases and the like in addition to the patient ID corresponding to the medical image data. Thus, “k-anonymization” may be realized.

(Third embodiment)
In the above-described first embodiment, the configuration in which the determination function 152 determines whether to collect the calculation result using the statistical data 142 has been described. On the other hand, in the third embodiment, a case will be described in which the determination function determines a calculation result using a learned model.

  FIG. 19 is a diagram illustrating an example of functional blocks of the information collection system according to the third embodiment. As shown in FIG. 19, the information collection system 1b includes an information collection device 10b, a group of medical image diagnostic devices 20, an image storage device 21, a learning device 30, and a terminal 40. The information collection device 10b includes a communication I / F 11, an input I / F 12, a display 13, a storage circuit 14b, and a processing circuit 15b.

  For example, as illustrated in FIG. 19, the storage circuit 14b further stores a learned model 145 in addition to the calculation result information 141, the statistical data 142, the useful data 143, and the model data 144. The learned model 145 is a learned model that is incorporated in a determination process that is performed by the processing circuit 15b and that outputs, based on an input operation result, whether medical image data corresponding to the operation result is useful or not. is there. The learned model 145 is configured by a multilayer neural network including an input layer, a plurality of intermediate layers, and an output layer, for example, like the learned models 341, 342, and 343 illustrated in FIG. The learned model 341 is generated by, for example, a model generation function 156 described later.

  Further, the processing circuit 15b executes a control function 151, a determination function 152b, an output function 153, a reception function 154, and a model generation function 156. The determination function 152b performs a determination process on the input calculation result based on the learned model 145. More specifically, the determination function 152b performs a determination process of determining whether medical image data corresponding to the calculation result is useful by inputting the calculation result to the learned model 145. Then, the determination function 152b outputs the determination result output from the learned model 145 to the output function 153.

  The model generation function 156 executes a process related to generation of the learned model 145. For example, the model generation function 156 first receives, from the control function 151, input of learning data including a set of an operation result as input data and information indicating whether or not the answer data is useful. The model generation function 156 trains the model based on the learning data so as to determine correct data such as information indicating whether or not it is useful from the calculation results included in the learning data.

  The example of the configuration of the information collection device 10b according to the third embodiment has been described above. With such a configuration, the information collection device 10b can determine whether the medical image data is useful data based on the learned model based on the calculation result.

(Fourth embodiment)
In the above-described first embodiment, the configuration in which the average value μ and the standard deviation σ of the calculation result are used as the statistical data has been described. On the other hand, in the fourth embodiment, a case will be described in which the relationship between the first calculation result and the second calculation result is used as the statistical data. FIG. 20 is a diagram illustrating an example of a distribution of calculation results according to the fourth embodiment.

  FIG. 20 shows a correlation distribution between the operation result 1 as the first value and the operation result 2 as the second value. As shown in FIG. 20, most of the sets of the calculation results 1 and 2 are distributed along the range 98. In this case, the set 99a of the calculation result 1 and the calculation result 2 included in the range 98 is not determined to be the calculation result based on the rare case. On the other hand, the sets 99b and 99c of the calculation result 1 and the calculation result 2 that deviate from the range 98 are determined to be calculation results based on the rare case.

  As described above, in the configuration according to the fourth embodiment, the information collection device can determine the validity of the determination result using the learned model based on the relationship between the plurality of calculation results. In FIG. 20, the configuration for determining whether or not a case is a rare case based on the relationship between two calculation results has been described. Such a configuration may be adopted.

  According to at least one embodiment described above, useful learning data can be collected.

  Although several embodiments of the present invention have been described, these embodiments are provided by way of example and are not intended to limit the scope of the invention. These embodiments can be implemented in other various forms, and various omissions, replacements, and changes can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are also included in the invention described in the claims and equivalents thereof.

  For example, the configuration in which medical image data is collected as medical information has been described. However, the configuration is not limited to this, and a configuration in which medical information other than images, such as electrocardiograms and audio data, may be collected. Further, the configuration in which the learning device 30 has the plurality of learned models 341, 342, and 343 has been described, but the learning device 30 may have a configuration in which only the single learned model 341 is provided.

  Further, the receiving function 154 may further receive from the user an input of a condition for determining whether or not such data is useful data, for example, a condition such as “from −σ to the average μ”. In this case, the determination function 152 may determine the calculation result using the input condition.

  In addition, the reception function 154 may further store, for example, a doctor ID identifying the doctor who registered the correct answer data in the useful data 143. For example, when a plurality of doctors input different cases as correct data for the same medical image data, the reception function 154 stores the case with the largest number of input doctors in the useful data 143 as correct data. You may. In addition, the reception function 154 may determine the evaluation of the doctor in advance and store the case input by the doctor with the highest evaluation in the useful data 143 as correct answer data.

  The term “processor” used in the description of each of the above-described embodiments may be, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an application specific integrated circuit (ASIC), Means a circuit such as a logic device (for example, a Simple Programmable Logic Device (SPLD), a complex programmable logic device (CPLD), and a field programmable gate array (FPGA)). I do. Here, instead of storing the program in the memory, the program may be directly incorporated in the circuit of the processor. In this case, the processor realizes a function by reading and executing a program incorporated in the circuit. Further, each processor of the present embodiment is not limited to the case where each processor is configured as a single circuit, but may be configured as one processor by combining a plurality of independent circuits to realize its function. Good.

  Here, the program executed by the processor is provided by being incorporated in a ROM (Read Only Memory) or a storage unit in advance. This program is a file in a format that can be installed or executed in these devices, such as a CD (Compact Disk) -ROM, an FD (Flexible Disk), a CD-R (Recordable), and a DVD (Digital Versatile Disk). May be provided by being recorded on a computer-readable storage medium. The program may be stored on a computer connected to a network such as the Internet, and may be provided or distributed by being downloaded via the network. For example, this program is configured by modules including each functional unit. As actual hardware, the CPU reads a program from a storage medium such as a ROM and executes the program, whereby each module is loaded on the main storage device and generated on the main storage device.

DESCRIPTION OF SYMBOLS 1 Information collection system 10 Information collection device 20 Medical image diagnostic device group 21 Image storage device 30 Learning device 40 Terminal

Claims (12)

The intermediate layer outputs in a first learned model including one or more intermediate layers for performing an arithmetic process on input information and an output layer for outputting a diagnosis result based on the operation result of the intermediate layer. A first storage unit that stores information about a calculation result;
An acquisition unit configured to acquire an operation result output by the intermediate layer in response to medical information being input to an input layer of the first learned model;
A determining unit that determines whether or not the calculation result obtained by the obtaining unit satisfies a condition between the calculation result and the information on the calculation result stored in the first storage unit;
A second storage unit for storing information on the medical information corresponding to the calculation result determined to satisfy the condition.   The information collection system according to claim 1, wherein the acquisition unit stores information including the acquired operation result in the first storage unit as information regarding the operation result. An output unit that outputs information on medical information stored in the second storage unit;
3. A receiving unit that receives, from a user, an input of a determination result for the medical information output by the output unit, and stores the received determination result in the second storage unit in association with the medical information. Information collection system described in.   The information collection system according to claim 3, wherein the determination unit stores the medical information and the determination result in the second storage unit in association with subject information regarding a subject from which the medical information has been collected.   The obtaining unit further obtains subject information on the subject of the medical information input to the input layer, and subject information that matches the obtained subject information is stored in the second storage unit. 5. The information collection system according to claim 4, wherein in the case, information on the medical information input to the input layer in association with the matching subject information is stored in the second storage unit.   The information collection system according to claim 4, wherein the reception unit stores anonymously information on the subject of the medical information, the medical information, and the determination result.   The information collection system according to claim 3, wherein the receiving unit stores the medical information and the determination result in the second storage unit in association with the user. The acquisition unit acquires, from a plurality of the first learned models, an operation result output by the intermediate layer,
The second storage unit stores the calculation result obtained by the obtaining unit in association with information on the plurality of first learned models that output the calculation result. Information collection system described in. Further comprising a collecting unit for collecting the medical information from a plurality of medical image diagnostic apparatuses,
The reception unit according to any one of claims 3 to 8, wherein the reception unit stores the medical information and the determination result in association with information regarding the medical image diagnostic apparatus from which the collection unit has collected the medical information. Information collection system.   The information collection system according to any one of claims 2 to 9, wherein the acquisition unit stores, in the first storage unit, statistical data obtained by statistically processing the acquired operation results as information regarding the operation results.   The determination unit receives an input of the calculation result obtained by the obtaining unit, and inputs the calculation result to a second learned model that outputs information indicating whether the condition is satisfied. The information collection system according to any one of claims 2 to 10, wherein it is determined whether a condition is satisfied. The intermediate layer outputs in a first learned model including one or more intermediate layers for performing an arithmetic process on input information and an output layer for outputting a diagnosis result based on the operation result of the intermediate layer. A first storage unit that stores information about a calculation result;
An acquisition unit configured to acquire an operation result output by the intermediate layer in response to medical information being input to an input layer of the first learned model;
A determining unit that determines whether or not the calculation result obtained by the obtaining unit satisfies a condition between the calculation result and the information on the calculation result stored in the first storage unit;
A second storage unit for storing information related to the medical information corresponding to a calculation result determined to satisfy the condition.

Images