JP7163168B2 - Medical image processing device, system and program - Google Patents

Description

The embodiments of the present invention relate to medical image processing apparatuses, systems, and programs.

With the development of machine learning, application of machine learning is progressing also in the medical field. When creating a learned model by machine learning a model using medical data, it is necessary to prepare a large amount of medical data for learning.
It is difficult to use a trained model once created for processing for a certain purpose for processing for a fundamentally different purpose. For example, a trained model for detecting brain tumors specializes in detecting tumors in medical images of the brain. Therefore, even if the trained model is used to detect a tumor from another site such as a liver tumor, the processing capability for medical images of the liver is considered to be low.

Therefore, for example, a user who has generated or purchased a trained model for detecting tumors from medical images of the brain wants to use the trained model for detecting liver tumors. In this case, it is necessary to prepare new liver tumor medical data and generate a trained model of the liver tumor, or separately purchase a trained model of the liver tumor. As a result, previously used medical data and trained models cannot be reused, which imposes a heavy burden on the user in terms of labor and cost.

U.S. Patent Application Publication No. 2010/0121178 U.S. Pat. No. 9,208,405 U.S. Patent Application Publication No. 2018/0060722

A problem to be solved by the present invention is to improve usability.

A medical image processing apparatus according to this embodiment includes an acquisition unit and a learning unit. The acquisition unit acquires a first trained model trained using the input data and the output data. The generating unit re-learns the first trained model using the preprocessed data and the output data, and removes the processing corresponding to the preprocessing from the first trained model. generate a second trained model.

FIG. 1 is a block diagram showing a medical image processing apparatus according to the first embodiment. FIG. 2A is a diagram showing a specific example of subdivision of a trained model by the learning function. FIG. 2B is a diagram showing a specific example of subdivision of a trained model by the learning function. FIG. 3 is a block diagram showing a medical image processing apparatus according to the second embodiment. FIG. 4 is a diagram showing an operation example of the medical image processing apparatus according to the second embodiment. FIG. 5 is a diagram showing an example of determining subdivision candidates by the model selection function. FIG. 6 is a diagram showing an example of a GUI (Graphical User Interface) by the medical image processing apparatus according to the third embodiment. FIG. 7 is a diagram showing a first example of model search processing in the GUI. FIG. 8 is a diagram showing a second example of model search processing in the GUI. FIG. 9 is a diagram showing a second example of model search processing in the GUI. FIG. 10 is a block diagram showing a medical image diagnostic system according to the fourth embodiment. FIG. 11 is a diagram showing an operation example of the medical image processing apparatus in the medical image diagnostic system according to the fourth embodiment.

A medical image processing apparatus, system, and program according to the present embodiment will be described below with reference to the drawings. In the following embodiments, it is assumed that parts denoted by the same reference numerals perform the same operations, and overlapping descriptions will be omitted as appropriate. An embodiment will be described below with reference to the drawings.

The medical image processing apparatus according to the present embodiment described below may be included in the console of each medical image diagnostic apparatus such as an X-ray CT (Computed Tomography) apparatus or MR (Magnetic Resonance), or may be included in a workstation. may be included in a server such as a cloud server.

(First embodiment)
The medical image processing apparatus according to the first embodiment subdivides the processing content to be executed by the trained model and changes it to a trained model specialized for certain processing content.
A medical image processing apparatus according to the first embodiment will be described with reference to the block diagram of FIG.
A medical image processing apparatus 10 according to the first embodiment includes a processing circuit 11 , a data storage section 12 and a communication interface 13 . The processing circuitry 11 includes a model acquisition function 111 , a process division function 113 , a learning function 115 and a display control function 117 .

With the model acquisition function 111, the processing circuit 11 acquires a trained model to be processed. A trained model to be processed is a model generated by learning using a set of input data and output data as learning data. The trained model to be processed may be acquired from the outside, or a trained model stored in advance in the data storage unit 12, which will be described later, may be acquired as the trained model to be processed. Also, the trained model to be processed may be a trained model that is generated by learning a model using the learning function 115, which will be described later.

With the processing division function 113, the processing circuit 11 subdivides the processing content to be executed by the trained model to be processed, and obtains subdivision information indicating how the processing content is realized by combination of processing. Specifically, assume a trained model that receives CT raw data and outputs images in which pulmonary nodules are segmented. The process of outputting an image in which the lung nodule is segmented from the raw data in the trained model includes, for example, a process of CT image reconstruction of the raw data, a process of segmenting the lung field from the CT image, and a process of segmenting the lung field from the CT image. It can be subdivided into the process of segmenting nodules. Therefore, the processing circuit 11 may obtain information about the subdivided processing as the subdivision information by the processing division function 113 .
In addition, as a method for subdividing the processing content, it is assumed that the user inputs to the medical image processing apparatus 10 what kind of processing is to be subdivided. The subdivision method obtains a subdivision example of the processing content from the outside, or obtains a subdivision example of the processing content executed in the medical image processing apparatus 10 in the past, and looks up the subdivision example. It is stored in the data storage unit 12 as a table. The processing circuit 11 may refer to the lookup table by the processing division function 113 .

Based on the subdivision information, the processing circuit 11 by the learning function 115 generates preprocessed data that has been preprocessed differently from the input data used to generate the trained model, and output data used to generate the trained model. is used as learning data, and the learned model is re-learned. As a result of the re-learning, the processing circuit 11 generates a re-learned model by removing the processing corresponding to the pre-processing from the original trained model by the learning function 115 .
By means of the display control function 117, the processing circuit 11 causes the display to display the learned model, subdivision information of the learned model, and the like.

The data storage unit 12 is a storage device such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), or an integrated circuit storage device that stores various information. The data storage unit 12 includes portable storage media such as CDs (Compact Discs), DVDs (Digital Versatile Discs), flash memories, semiconductor memory devices such as RAMs (Random Access Memory), etc., in addition to HDDs and SSDs. It may be a driving device that reads and writes various information between. Moreover, the storage area of the memory may be in the medical image processing apparatus 10 or in an external storage device connected via a network.
The data storage unit 12 stores, for example, input data, correct data (output data) for relearning a trained model, and subdivision information. Examples of input data include raw data, sinogram data, volume data, and image data. Examples of correct data include sinogram data, volume data, image data, determination scores (lesion scores, etc.), and analysis results. The learning data may be a set of input data and correct data, or may be a set of preprocessed data and correct data. The data storage unit 12 may further store a preprocessing model, which is a trained model that preprocesses input data. Preprocessed data can be generated from input data by using a preprocessed model.

The communication interface 13 is an interface for communicating with the outside, and since a general interface may be used, description thereof will be omitted here.

Next, a specific example of segmentation of the trained model by the learning function 115 will be described with reference to FIGS. 2A and 2B.
A trained model is generated by machine learning a model such as a multi-layer network according to a model learning program based on learning data, which is a set of input data and output data. Machine learning algorithms are assumed to be neural networks, deep learning, random forests, etc., but other machine learning algorithms such as support vector machines and kernel regression may also be used. Note that the trained model can also be said to be a parameterized composite function obtained by combining a plurality of functions. A parameterized composite function is defined by a combination of multiple adjustable functions and parameters. A trained model can be any parameterized composite function that meets the above requirements.

Here, as shown in FIG. 2A, (a) as a trained model to be processed, a trained model 20 whose processing content is "segmentation of lung tumor" is assumed. When volume data is input as input data 21 , the trained model 20 outputs image data obtained by segmenting a lung tumor as output data 22 .

With the processing division function 113, the processing circuit 11 subdivides the processing contents of the learned model 20 up to the segmentation of the lung tumor from the volume data based on the subdivision information. The details of processing may be subdivided by a method of appropriately inputting by the user as described above or a method of referring to a lookup table. Furthermore, when the state of the trained model 20 to be processed can be confirmed during processing, it detects what kind of processing is being performed during processing, and subdivides the processing contents according to the detected processing. good too.

Here, it is assumed that the processing is subdivided into a first step of segmenting the lung from the volume data, a second step of dividing the segmented lung into lobes, and a third step of segmenting the tumor. . Here, when the user wishes to generate a trained model that performs the processing of the third step, the first and second steps are called preprocessing.

With the learning function 115, the processing circuit 11 learns, as preprocessing for the input data 21, data obtained by performing the first step on the input data 21 as the input data 23 and data obtained by segmenting the lung tumor as the output data 22. The trained model 20 is relearned (for example, reinforced learning) based on the data for use ((b) in FIG. 2A). As the input data 23, volume data obtained by segmenting the lungs may be prepared in advance in the data storage unit 12, and the volume data obtained by segmenting the lungs may be used. Alternatively, volume data obtained by segmenting lungs obtained from the outside may be used. Alternatively, by performing a general segmentation process on the volume data stored in the data storage unit 12, volume data obtained by segmenting the lungs may be generated and used. Alternatively, the output of a trained model that receives volume data and outputs lung segmentation data may be used.

The relearned model 24 obtained as a result of relearning is changed to a trained model that processes the first preprocessed data because data obtained by segmenting the lungs from the volume data is used as input data. That is, the relearned model 24 is changed to a model obtained by removing the first preprocessing function from the trained model 20 ((c) in FIG. 2A). The retrained model 24 receives data obtained by segmenting lungs from volume data, and outputs data obtained by segmenting lung tumors.

Similarly, an example of retraining a retrained model 24 using preprocessed data and output data 22 for the second step is shown in FIG. 2B.
Here, in the second step, the relearned model 24 is re-trained based on the learning data in which the input data 25 is data obtained by dividing the lung into lobes, and the output data 22 is data obtained by segmenting the lung tumor. Learn ((d) in FIG. 2B). The input data 25 may be acquired in the same manner as the input data 23 . For example, as the input data 25, data preliminarily divided into lung lobes may be used, or data obtained by segmenting the lung is input, and the output of a trained model that outputs data divided into lung lobes is used. may be

The re-learned model 26 obtained as a result of re-learning is changed to a trained model that processes the second preprocessed data because the data divided for each lung lobe is used as input data. That is, the relearned model 26 is changed to a model obtained by removing the second preprocessing function from the relearned model 24 ((e) in FIG. 2A). The retrained model 26 receives data segmented by lung lobe, and outputs data obtained by segmenting lung tumors.

In the re-learning stages shown in (b) and (d) of FIGS. 2A and 2B, the learning function 115 causes the processing circuit 11 to perform processing corresponding to preprocessing in the multilayer network of the trained model 20. If it is possible to detect the layer in which the preprocessing is being performed, the model from which the layer in which the processing corresponding to the preprocessing is being performed is eliminated may be re-learned to generate a re-learned model.

Furthermore, in FIGS. 2A and 2B, the learned model is re-learned for each subdivided process, but this is not restrictive. The re-learning of (d) may be executed using the processed data as input data.

According to the first embodiment described above, by subdividing the processing content of a trained model and re-learning (reinforcement learning) to become a highly versatile trained model, a trained model with high versatility can be obtained. A trained model can be changed into a model. That is, for example, a trained model that outputs an image obtained by segmenting a lung tumor is retrained, and the trained model is changed to a trained model that specializes in the process of segmenting a tumor. When data obtained by segmenting the liver is input to the modified trained model, the modified trained model outputs an image obtained by segmenting the tumor with respect to the data on which some segmentation has been performed. Tumors can be segmented not only from medical data, but also from medical data of other organs such as the liver.
That is, according to the first embodiment, it is possible to generate a more versatile trained model and improve usability.

(Second embodiment)
In the second embodiment, desired processing is realized by combining a plurality of general-purpose trained models instead of realizing desired processing with one trained model.

A medical image processing apparatus 10 according to a second embodiment will be described with reference to the block diagram of FIG.
The medical image processing apparatus 10 according to the second embodiment further includes a model selection function 119 in the processing circuit 11 in addition to the medical image processing apparatus 10 according to the first embodiment. Also, the configurations of the data storage unit 12 and the processing division function 113 of the processing circuit 11 are different from those of the first embodiment.

The data storage unit 12 stores a plurality of trained models. Each of the trained models has different processing contents. For example, a trained model for extracting a tumor, a trained model for lung segmentation, etc. are stored according to each application. Note that a plurality of trained models may be stored according to differences in input data even if the processing content is the same. For example, a trained model for segmenting tumors from CT images and a trained model for segmenting tumors from MR images are stored respectively.

In addition, additional information is attached to the trained model. Supplementary information includes the processing details of the trained model, the correct answer rate of the data output from the trained model, the input data and output data of the trained model, the processing time required to execute the processing details of the trained model, etc. .

Further, the data storage unit 12 may store a processing model designed on a rule basis (hereinafter referred to as a rule model) in addition to the learned model. Supplementary information is also attached to the rule model.

With the processing division function 113, the processing circuit 11 subdivides the processing desired by the user (hereinafter referred to as target processing) into a plurality of processings. In other words, the target process is converted into a combination of multiple processes (steps). When subdividing the target process, the same method as in subdividing the processing content of the trained model according to the first embodiment may be used.
With the model selection function 119, the processing circuit 11 selects a combination of trained models for achieving the target processing based on the incidental information.

Next, a specific operation example of the medical image processing apparatus 10 according to the second embodiment will be described with reference to FIG.
As the target process 40, a case of "segmentation of pulmonary nodules from raw data acquired by an X-ray CT apparatus" is assumed. Raw data acquired by the X-ray CT apparatus is hereinafter referred to as CT raw data.

The processing division function 113 causes the processing circuit 11 to subdivide the target processing 40 into image reconstruction processing from CT raw data, lung field segmentation processing, and lung nodule segmentation processing. As a result, a combination of first trained model 41, second trained model 43 and third trained model 45 is obtained.

The granularity of subdivision of the target process (for example, the number of processes to be combined) may be determined so as to preferentially use the trained model owned by the user (stored in the data storage unit 12). . For example, when the target process can be subdivided into five processes from the first process to the fifth process, if the user already owns a trained model that summarizes the fourth and fifth processes, the learned model In order to use the finished model, the target process should be subdivided into three processes. The granularity of subdivision may be determined as a combination of trained models that minimizes the number of subdivisions, or conversely, it is determined as a combination of trained models that maximizes the number of subdivisions. may

With the model selection function 119, the processing circuit 11 selects from the data storage unit 12 each trained model corresponding to the image reconstruction processing, the lung field segmentation processing, and the nodule segmentation in which the target processing is subdivided. The model selection function 119 allows the processing circuit 11 to automatically select a learning model, and the user may manually select a model.

When subdividing the target process, the target process may not be subdivided into one pattern, and a plurality of patterns may be assumed as subdivision candidates. In this case, model selection function 119 may allow processing circuitry 11 to determine which candidate refinement is most appropriate.

An example of determination of subdivision candidates by the model selection function 119 will be described with reference to FIG.
The example of FIG. 5 shows a case where two subdivision candidates 51 and 53 are generated for the objective process 40 "segmentation of pulmonary nodules from raw data acquired by an X-ray CT apparatus" shown in FIG. The subdivision candidate 51 is a combination of trained models 511 in the order indicated by “image reconstruction processing→lung segmentation→lung lobe segmentation→tumor segmentation”, and the subdivision candidate 53 is a combination of “lung segmentation→TDC acquisition processing”. →TDC-based tumor segmentation” is assumed to be a combination of trained models 531 in the order shown.

Additional information 513 is attached to the trained model 511 , and additional information 533 is attached to the trained model 531 . Further, the subdivision candidate 51 is attached with overall additional information 52 , and the subdivision candidate 53 is also provided with overall additional information 54 . In addition, in the example of FIG. 5, for convenience of explanation, the contents of only the incidental information 513 are illustrated, but the same information is included in other incidental information.

The correct answer rate of the data included in the overall supplementary information 52 and supplementary information 54 may be the average of the correct answer rates of the trained models included in the subdivision candidates, or the subdivision candidate 51 or subdivision candidate 53 It may be the correct answer rate obtained by processing the sample data by . Alternatively, feedback data of correct answer rate when actually used by a third party may be used. As the processing time included in the overall incidental information, the time obtained by accumulating the processing time of each learned model may be used.

With the model selection function 119, the processing circuit 11 selects an optimum subdivision candidate (combination candidate) as a processing model based on the incidental information. In the example of FIG. 5, the subdivision candidate 53 is selected as the processing model if the correct answer rate is prioritized, and the subdivision candidate 51 is selected as the processing model if the processing time is prioritized. selected. It should be noted that the user may manually select the optimal subdivision candidate as the processing model, taking into account the presented correct answer rate or processing time.

According to the second embodiment described above, the target process is subdivided into a plurality of processes, and the model selection function 119 allows the processing circuit 11 to combine trained models for each process, thereby realizing the target process. As a result, for example, even if the user does not own a trained model whose processing content is the target process, the same target process can be realized by combining the trained model in hand, and the user already purchased or created a hand-held It is possible to effectively use the trained model of

(Third embodiment)
The third embodiment shows a medical image processing apparatus that provides a GUI (Graphical User Interface) that supports combination of trained models.

An example of the GUI according to the third embodiment will be described with reference to FIG.
FIG. 6 shows a GUI that assists the user in combining trained models, and the display control function 117 causes the processing circuit 11 to control the display of the GUI on, for example, a console device or workstation display.

The user can input to the GUI or move the cursor 68 by directly touching an external input device such as a keyboard, mouse, or trackball, or by directly touching a touch-panel display. It is possible to carry out a design including a combination work of a finished model.

In FIG. 6, it is assumed that input data and output data are determined for a certain target process. The display control function 117 causes the processing circuit 11 to display a work area 60 for combining a plurality of trained models for target processing, an AI model list 62 for displaying a plurality of trained models 601, and a rule base list for displaying a rule model 640. 64, and a connector list 66 that displays connectors 660, described below. It is assumed that the AI model list 62 includes a plurality of trained models stored in the data storage unit 12 shown in the second embodiment. It may be included in the model list 62.

The user selects a plurality of learned models 601 from the AI model list 62, selects a rule model 640 from the rule base list 64, and combines the learned models 601 and the rule model 640 so as to achieve the desired processing. to run. Further, although not shown, by selecting the learned model 601, additional information attached to the learned model 601 may be displayed. Supplementary information may also be attached to the rule model 640 . Hereinafter, for convenience, the trained model 601 and the rule model 640 are simply referred to collectively as "models."

The AI model list 62, the rule base list 64, and the connector list 66 may be sorted and displayed for each category of model processing content. For example, the user's visibility can be improved by displaying the model separately in the category of "segmentation" and the category of "tumor determination".

The subdivision candidates shown in the second embodiment may be displayed in the work area 60 by default, and the user may decide whether to use the subdivision candidates as they are or to rearrange the subdivision candidates.

A trained model 601 displayed on the GUI includes an explanation block 6011 , an input terminal 6013 and an output terminal 6015 . A description block 6011 is a block describing the processing details of the trained model 601 . The input terminal 6013 indicates the format of the input data by a symbol, and the symbol is connected to one end of the explanation block 6011 via a branch. The output terminal 6015 indicates the format of output data by a symbol, and the symbol is connected to the other end of the explanation block 6011 via a branch.

As the format of each data of the input terminal 6013 and the output terminal 6015, for example, pixel data, voxel data, raw data, intermediate data (for example, sinogram), etc. are assumed. Each data is assigned a different symbol to uniquely identify the format of each input data.

A user can easily determine whether or not the input and output data formats match by looking at the shapes of the input terminal 6013 and the output terminal 6015 . In the example of FIG. 6, assuming that the input data is volume data and the symbol is "○", the input terminal 6013 of the trained model 601 for lung segmentation processing is the symbol "◯", so the user can It can be determined that the model 601 can be connected to the input. Note that the rule model 640 has the same display format as the learned model 601, so the explanation is omitted.

Here, in the work of combining models, there may be a case where the input/output relationship between the input and output of the target process is not aligned. For example, when the output from the previous model is volume data and the input of the trained model 601 to be connected is image data, conversion from volume data to image data is required. Also, even if the input and output are the same image data, if the matrix size of the image is different, it is necessary to convert the matrix size of the output source according to the matrix size of the input destination.

Therefore, a connector list 66 stores a plurality of connectors 660 that indicate data conversion that matches the input/output relationship between models, such as data format conversion and matrix size conversion. The connector 660 may use a mask image for matrix size conversion, and may use a general data conversion formula for data conversion.
By preparing the connector 660, the user can visually and easily match the data input/output relationship between trained models.

Next, a first example of model search processing in the GUI will be described with reference to FIG.
FIG. 7 shows a state in which one trained model 601 is arranged in the work area 60. FIG. The display control function 117 causes the processing circuit 11 to display the candidate area 70 on the GUI. In the candidate area 70, model candidates 701 and necessary connectors 703 are displayed in association with each other.
A model candidate 701 indicates a model that can be combined with the model arranged in the work area 60 . For the model candidate, for example, the processing circuit 11 may extract a model connected to the arranged trained model 601 based on the segmentation information by the processing division function 113 .
A required connector 703 indicates a connector required when connecting the model candidate 701 to the model arranged in the work area 60 . For example, the display control function 117 causes the processing circuit 11 to refer to the supplementary information, and based on the information on the output of the model arranged in the work area 60 and the input of the model that can be combined, the input/output relationship is the same. If they do not match, a conversion that matches the input/output relationship, that is, a connector can be selected and displayed. If a connector is not required, write "none" or leave it blank.

When a model is placed in the work area 60 by the display control function 117, the processing circuit 11 refers to the placed model and selects model candidates from the AI model list 62, the rule base list 64, and the connector list 66. A model or connector that becomes is extracted and displayed in the candidate area 70 .

Next, a second example of model search processing in the GUI will be described with reference to FIGS. 8 and 9. FIG. 8 and 9 assume a so-called incremental search.
FIG. 8 is an example of a user entering characters into a search window 80 for a model that the user wishes to use in combining models. When "ha" is entered, candidates for learned models 601 and rule models 640 beginning with "ha" are displayed in the candidate field 81. FIG. By giving a name to the connector 660, the connector 660 may be set as a search target together with the trained model 601 and the rule model 640. FIG. When the user subsequently enters "yes" in search window 80, the models displayed in candidate column 81 are limited to learned models 601 and rule models 640 beginning with "yes". As such a trained model 601, for example, a lung segmentation model is displayed.

According to the third embodiment described above, by providing a GUI that supports the combination of learned models, the user can freely combine models within a range in which the target process can be achieved. can improve the usability of the design of

(Fourth embodiment)
Next, as a fourth embodiment, an example of purchasing and selling models using the cloud will be described.
A medical image diagnostic system including a medical image diagnostic apparatus according to the fourth embodiment will be described with reference to FIG.

A medical image diagnostic system 1 shown in FIG. 10 includes one or more medical image processing apparatuses 10 and a cloud server 3 . It is assumed that one or more medical image processing apparatuses 10 are arranged for each hospital. Each medical image processing apparatus 10 is connected to the cloud server 3 via the network 5 .

The medical image processing apparatus 10 may have the configuration of the medical image processing apparatus 10 according to the first embodiment or the second embodiment, but here the processing circuit 11 includes the display control function 117 .

The cloud server 3 includes a data storage unit 12 , a processing circuit 11 and a communication interface 13 . Processing circuit 11 includes model acquisition function 111 , learning function 115 and processing division function 113 . In this way, the cloud server 3 includes the learning function 115, which is assumed to have a large processing load, and the data storage unit 12, which has a large memory load, and the other components are included in the medical image processing apparatus 10. may be dispersed. When distributing the configuration, access authority is granted to each medical image processing apparatus 10, and among learning data and various models (learned models, rule models) stored in the cloud server 3, only data with access authority access may be restricted to allow access to

The cloud server 3 may cause the medical image processing apparatus 10 to download data in response to download requests from the medical image processing apparatus 10 for learning data and various models. Alternatively, the cloud server 3 may cause the medical image processing apparatus 10 to upload data in response to an upload request from the medical image processing apparatus 10 for learning data and various models.

By adopting the configuration shown in FIG. 10, on the side of each medical image processing apparatus 10, the throughput can be improved while reducing the load on the processor and memory that implement the processing circuit 11. FIG. Further, by performing learning processing on the cloud server 3 side, the learning time can be shortened, and centralized management of data can be realized.

Next, FIG. 11 shows an operation example of the medical image processing apparatus 10 according to the fourth embodiment.
FIG. 11 is an example of a GUI showing a work area 60 and candidate areas 72, similar to FIG.

The display control function 117 causes the processing circuit 11 to display the candidate area 72 on the GUI. The candidate area 72 displays the candidate model 701 and an ownership flag 705 indicating whether or not the user owns the candidate model 701 in association with each other. Models that the model candidate does not own (hereinafter referred to as unpossessed models 721) are displayed with the ownership flag 705 as "x". The non-possessed model 721 may be displayed with a dashed line, a light color, or a mark as shown in FIG. 10 . In other words, the non-possessed model 721 may be displayed in a manner to be distinguished from the model owned by the user.

Furthermore, the display control function 117 causes the processing circuitry 11 to display a purchase button 723 for purchasing the non-possessed model 721 . By moving the cursor 68 to the purchase button 723 and clicking it, the user displays a model and price confirmation screen (not shown), and further presses the confirmation button to enable the purchase of the non-possessed model 721 . The method of purchasing the non-possessed model 721 is, for example, to store a large number of learned models, rule models, and connectors in the cloud server 3, access the cloud server 3 from the medical image processing apparatus 10 used by the user, and purchase the non-possessed model 721. Model 721 should be made available for purchase.

In addition, it is also possible to allow users to register learned models, rule models and connectors created by users in the cloud server 3, and to sell the registered models to third parties.

According to the fourth embodiment described above, learning data and various models are stored in the cloud server, and the user can purchase and sell various models from the cloud server, thereby improving usability.

The processing circuit 11 according to the present embodiment includes, for example, a processor such as a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), a ROM (Read Only Memory), and a RAM ( Random Access Memory) and other memories. The processing circuit 11 may also be an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other complex programmable logic device (CPLD). , may be implemented by a Simple Programmable Logic Device (SPLD). The processing circuit 11 executes a model acquisition function 111, a processing division function 113, a learning function 115, a display control function 117, and a model selection function 119 by means of a processor that executes programs developed in memory. Each function (model acquisition function 111, process division function 113, learning function 115, display control function 117, and model selection function 119) is not limited to being realized by a single processing circuit. The processing circuit 11 may be configured by combining a plurality of independent processors, and each function may be implemented by executing a program by each processor.

Note that each function according to the above-described embodiment can also be realized by installing a program for executing the processing in a computer such as a workstation and deploying them on the memory. At this time, a program that can cause a computer to execute the method is stored in a storage medium such as a magnetic disk (hard disk, etc.), an optical disk (CD-ROM, DVD, Blu-ray (registered trademark) disk, etc.), a semiconductor memory, etc. It is also possible to distribute

According to the medical image processing apparatuses according to the first to third embodiments and the medical image processing system according to the fourth embodiment, usability can be improved.

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

1 Medical Image Diagnosis System 3 Cloud Server 5 Network 10 Medical Image Processing Apparatus 11 Processing Circuit 12 Data Storage Unit 13 Communication Interface 20 Trained Model 21, 23, 25 Input Data 22 Output Data 24, 26 Re-Trained Model 40 Objective Process 41 , 43, 45, 511, 531, 601 Trained model 51, 53 Subdivision candidate 52, 54, 513, 533 Additional information 60 Work area 62 AI model list 64 Rule base list 66 Connector list 68 Cursor 70, 72 Candidate area 80 Search window 111 Model acquisition function 113 Processing division function 115 Learning function 117 Display control function 119 Model selection function 640 Rule model 660 Connector 701 Model candidate 703 Required connector 705 Possession flag 721 Unpossessed model 723 Purchase button 6011 Explanation block 6013 Input terminal 6015 Output terminal

Claims (10)

an acquisition unit that acquires a first trained model trained using the input data and the output data;
Re-learning the first trained model using preprocessed preprocessed data and the output data, and second learning in which processing corresponding to the preprocessing is removed from the first trained model a learning unit that generates a trained model;
A medical image processing apparatus comprising: further comprising a selection unit,
the second trained model is accompanied by supplementary information regarding the processing content and input/output of the second trained model;
2. The medical image processing apparatus according to claim 1, wherein said selection unit selects a combination of second trained models for achieving target processing based on said supplementary information. A GUI (Graphical User Interface) displaying a first list relating to a plurality of second trained models and a work area for performing work on combining the trained models including the second trained models selected from the first list; 3. The medical image processing apparatus according to claim 2, further comprising a display controller for controlling. 4. The medical image processing apparatus according to claim 3, wherein said first list is displayed by sorting said plurality of second trained models by category of processing content. When a second trained model is selected from the first list and the second trained model is arranged in the work area, the display control unit controls a learned model that can be combined with the arranged second trained model. 5. The medical image processing apparatus according to claim 3, wherein model candidates are displayed. 6. The medical image processing apparatus according to any one of claims 3 to 5, wherein said display control unit displays on said GUI a second list of rule models, which are processing models designed based on rules. the second trained model and the rule model are displayed as icons;
The icon includes a block describing the processing details of the second trained model, a symbol connected to one end of the block representing input data, and a symbol connected to the other end of the block representing output data. The medical image processing apparatus according to claim 6, comprising: The display control unit creates a third list related to connectors indicating data conversion that matches input/output relationships between the second trained models, between the rule models, and between the second trained models and the rule models. 8. The medical image processing apparatus according to claim 6, wherein the GUI is displayed. A medical image processing system including one or more medical image processing apparatuses and a server,
The server is
an acquisition unit that acquires a first trained model trained using the input data and the output data;
Re-learning the first trained model using preprocessed preprocessed data and the output data, and second learning in which processing corresponding to the preprocessing is removed from the first trained model a learning unit that generates a trained model;
a storage unit that stores a plurality of trained models including the input data, the output data, the preprocessed data, and the second trained model;
and
Each of the one or more medical image processing devices
a display control unit for controlling a GUI (Graphical User Interface) that displays a first list relating to the plurality of trained models and a work area for combining trained models selected from the first list; , medical imaging systems. to the computer,
an acquisition function for acquiring a first trained model trained using input data and output data;
Re-learning the first trained model using preprocessed preprocessed data and the output data, and second learning in which processing corresponding to the preprocessing is removed from the first trained model A medical image processing program that realizes a learning function that generates a finished model.

Images