CN106709917A - Neural network model training method, device and system - Google Patents
Neural network model training method, device and system Download PDFInfo
- Publication number
- CN106709917A CN106709917A CN201710002230.5A CN201710002230A CN106709917A CN 106709917 A CN106709917 A CN 106709917A CN 201710002230 A CN201710002230 A CN 201710002230A CN 106709917 A CN106709917 A CN 106709917A
- Authority
- CN
- China
- Prior art keywords
- network model
- sample
- neural network
- image
- client
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
Landscapes
- Apparatus For Radiation Diagnosis (AREA)
- Medical Treatment And Welfare Office Work (AREA)
Abstract
The invention discloses a neural network model training method, device and system and belongs to the image processing field. The method comprises the following steps that: medical samples sent by a plurality of clients are received, wherein the medical sample sent by the first client comprises a plurality of CT images and first label images corresponding to each CT image, wherein the first label images are used for identifying specified organs contained in the CT images, the first client adopts a local neural network model to segment the plurality of CT images so as to form the first label images, and the first client is any one client in the plurality of clients; and a first neural network model in a server is trained according to the medical samples sent by the plurality of clients, wherein the first neural network model is a neural network model of the latest version in the server. With the neural network model training method, device and system of the invention adopted, the training time of the neural network model is shortened, and the accuracy of the training of the neural network model is effectively improved. The neural network model training method, device and system of the present invention are used for training neural network models.
Description
Technical field
The present invention relates to image processing field, more particularly to a kind of neural network model training method, apparatus and system.
Background technology
Image segmentation is to divide the image into several regions specific, with unique properties and extract interesting target
Technology.It is by the committed step of image procossing to graphical analysis.
At present, in medical domain, it is possible to use neural network model is split to medical image, the neural network model
Initial neural network model can be trained using predetermined training sample and obtain (example, steepest can be used
Descent method is trained).Specifically, server can online under collect a large amount of training samples in advance, each training sample includes
The segmentation result of original image and original image, is trained to original neural network model using the plurality of training sample and obtained
Neural network model after training, after by neural metwork training success, can issue new neural network model version and supply
Client downloads.
But, at present in medical domain, it is necessary to first set up an original neutral net mould during training neural network model
Training sample is collected under type, doubling, the training time of neural network model is more long, and the accuracy of training is relatively low.
The content of the invention
In order to the training time for solving prior art neural network model is more long, the relatively low problem of the accuracy of training, this
Inventive embodiments provide a kind of neural network model training method, apparatus and system.The technical scheme is as follows：
A kind of first aspect, there is provided neural network model training method, is applied to the service of medical image segmentation system
Device, methods described includes：
The medical science sample that multiple client sends is received, wherein, the medical science sample that the first client sends includes：Multiple CT
Image and corresponding first label image of every CT image, first label image are used to identify the finger that the CT images are included
Determine organ, first label image is that first client is entered using local neural network model to described multiple CT images
Row segmentation is obtained, and first client is any client in the multiple client；
According to the medical science sample that the multiple client sends, the first nerves network model in the server is trained,
The first nerves network model is the neural network model of latest edition in the server.
Alternatively, the medical science sample sent according to the multiple client, trains the first god in the server
Through network model, including：
Inaccurate medical science sample in the medical science sample that the multiple client sends is deleted, training sample is obtained；
The first nerves network model in the server is trained using the training sample.
Alternatively, medical science sample inaccurate in the medical science sample for deleting the multiple client transmission, is instructed
Practice sample, including：
Multiple mask images of first sample are shown successively, and each mask image is by a label in the first sample
Imaging importing is formed on corresponding CT images, and the first sample is appointing in the medical science sample that the multiple client sends
One sample；
Receive the artificial interface where any mask image or interface triggering where the first sample to described
The deletion action of first sample；
Deleted the first sample as inaccurate sample according to the deletion action.
Alternatively, medical science sample inaccurate in the medical science sample for deleting the multiple client transmission, is instructed
Practice sample, including：
Every CT image in the first sample is entered using default standard neural network model in the server
Row segmentation, obtains the second label image corresponding with every CT images, and second label image is used to identify the CT
The designated organ that image is included, the first sample is any sample in the medical science sample that the multiple client sends；
Whether the segmentation image difference of the CT images is judged more than preset difference value threshold value, and the segmentation image difference is right
Should be in the image difference of first label image of same CT images and second label image；
When segmentation image difference is more than accounting of the CT images of preset difference value threshold value in the first sample more than default
During ratio, the first sample is stored to manual confirmation database, the sample in the manual confirmation database is used for by people
Work is confirmed whether to delete；
When segmentation image difference be not more than more than the accounting of the CT images in the first sample of preset difference value threshold value it is pre-
If during ratio, the first sample is stored to training sample database, the sample in the training sample database is used to instruct
Practice neural network model；
By the sample in the sample for confirming not delete in the manual confirmation database and/or the training sample database
It is defined as the training sample.
Alternatively, first in the medical science sample sent according to the multiple client, the training server
After neural network model, methods described also includes：
Nervus opticus network model is sent to multiple test clients, the nervus opticus network model is according to described
The neutral net mould obtained by first nerves network model in server described in the medical science sample training that multiple client sends
Type；
Receive marking of the multiple test client to the nervus opticus network model；
Marking according to the multiple test client to the nervus opticus network model, determines test result；
Judge the test result whether more than default passing score；
When the test result is more than default passing score, the nervus opticus network model is defined as latest edition
Neural network model；
When the test result is not more than default passing score, the first nerves network model is defined as latest edition
This neural network model.
A kind of second aspect, there is provided neural network model training method, is applied to the first of medical image segmentation system
Client, methods described includes：
Acquisition includes the sample to be split of multiple computer tomography CT images, and the CT images include designated organ
Image；
Every CT image is split using the local neural network model in first client, obtain with it is described
Corresponding first label image of every CT image, first label image is used to identify the specified device that the CT images are included
Official；
Medical science sample is sent to the server, is taken according to the medical science sample training in order to the server
First nerves network model in business device, the first nerves network model is the neutral net of latest edition in the server
Model, the medical science sample includes：Every CT images and corresponding first label image of every CT images.
A kind of third aspect, there is provided neural network model trainer, is applied to the service of medical image segmentation system
Device, described device includes：
First receiver module, the medical science sample for receiving multiple client transmission, wherein, the doctor that the first client sends
Imitate and originally include：Multiple CT images and corresponding first label image of every CT image, first label image are used to identify
The designated organ that the CT images are included, first label image is that first client uses local neural network model
Described multiple CT images are carried out splitting what is obtained, first client is any client in the multiple client；
Training module, for the medical science sample sent according to the multiple client, in the training server first
Neural network model, the first nerves network model is the neural network model of latest edition in the server.
Alternatively, the training module, including：
Submodule is deleted, for medical science sample inaccurate in the medical science sample for deleting the multiple client transmission, is obtained
To training sample；
Training submodule, for training the first nerves network mould in the server using the training sample
Type.
A kind of fourth aspect, there is provided neural network model trainer, is applied to the first of medical image segmentation system
Client, described device includes：
Acquisition module, the sample to be split of multiple computer tomography CT images, the CT images are included for obtaining
Image comprising designated organ；
Segmentation module, for being divided every CT image using the local neural network model in first client
Cut, obtain the first label image corresponding with every CT images, first label image is used to identify the CT images
Comprising designated organ；
Sending module, for medical science sample to be sent to the server, in order to the server according to the medical science
First nerves network model in server described in sample training, the first nerves network model is newest in the server
The neural network model of version, the medical science sample includes：Every CT images and every CT images are corresponding described
First label image.
5th aspect, there is provided a kind of medical image segmentation system, the medical image segmentation system includes：At least one
Server and at least one first clients being connected with the server；
The server includes the neural network model trainer described in the third aspect；
Each described first client includes the neural network model trainer described in fourth aspect.
The beneficial effect that technical scheme provided in an embodiment of the present invention is brought is：
Neural network model training method provided in an embodiment of the present invention, apparatus and system, by server in multiple visitors
After each client in the end of family carries out CT image segmentations using newest neural network model, receive what multiple client sent
Medical science sample comprising multiple CT images and label image, is trained to newest neural network model, realizes neutral net
The on-line training of model, shortens the training time of neural network model, is effectively improved the standard of neural network model training
True property.
Brief description of the drawings
Technical scheme in order to illustrate more clearly the embodiments of the present invention, below will be to that will make needed for embodiment description
Accompanying drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the present invention, for
For those of ordinary skill in the art, on the premise of not paying creative work, other can also be obtained according to these accompanying drawings
Accompanying drawing.
Fig. 1 is the schematic diagram of the implementation environment involved by neural network model training method provided in an embodiment of the present invention；
Fig. 2 is a kind of flow chart of neural network model training method provided in an embodiment of the present invention；
Fig. 3 is the flow chart of another neural network model training method provided in an embodiment of the present invention；
Fig. 4 is the flow chart of another neural network model training method provided in an embodiment of the present invention；
Fig. 5 be it is provided in an embodiment of the present invention it is a kind of according to multiple client send medical science sample training server in
The flow chart of first nerves network model；
Fig. 6-1 is a kind of schematic diagram of certain mask image of display provided in an embodiment of the present invention；
Fig. 6-2 is the schematic diagram of certain mask image of another display provided in an embodiment of the present invention；
Fig. 7-1 is a kind of model marking interface schematic diagram provided in an embodiment of the present invention；
It is another model marking interface schematic diagram provided in an embodiment of the present invention that Fig. 7-2 is；
Fig. 8 is a kind of structured flowchart of neural network model trainer provided in an embodiment of the present invention；
Fig. 9 is a kind of structured flowchart of training module provided in an embodiment of the present invention；
Figure 10 is the structured flowchart of another neural network model trainer provided in an embodiment of the present invention；
Figure 11 is the structured flowchart of another neural network model trainer provided in an embodiment of the present invention；
Figure 12 is the structured flowchart of another neural network model trainer provided in an embodiment of the present invention.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, below in conjunction with accompanying drawing to embodiment party of the present invention
Formula is described in further detail.
Fig. 1 is referred to, it illustrates involved by the neural network model training method provided in the embodiment of the present invention
Plant the schematic diagram of the implementation environment of medical image segmentation system.The implementation environment can include：Server 110 and multiple have aobvious
The client 120 of display screen.
Server 110 can be a server, or by some server groups into server cluster, or one
Individual cloud computing service center.Client 120 is the Medical Devices with display screen.
Connection, client 120 can be set up by cable network or wireless network between server 110 and client 120
Training sample can be provided for server 110, server 110 can be trained using the sample to neural network model, visitor
Family end 120 can download the neural network model for training from server 110 and carry out medical image segmentation.
Fig. 2 is a kind of flow chart of neural network model training method provided in an embodiment of the present invention, and the method is applied to
The server of medical image segmentation system, as shown in Fig. 2 the method can include：
Step 101, server receive the medical science sample that multiple client sends, wherein, the medical science that the first client sends
Sample includes：Multiple computed tomography (English：Computed Tomography；Referred to as：CT) image and every CT image
Corresponding first label image, first label image is used to identify the designated organ that CT images are included, and the first label image is
First client carries out splitting what is obtained using local neural network model to multiple CT images, and the first client is multiple clients
Any client in end.
The medical science sample that step 102, server send according to multiple client, the first nerves network in training server
Model, the first nerves network model is the neural network model of latest edition in server.
In sum, neural network model training method provided in an embodiment of the present invention, by server in multiple clients
After each client in end carries out CT image segmentations using newest neural network model, the bag that multiple client sends is received
Medical science sample containing multiple CT images and label image, the neural network model to latest edition is trained, and realizes nerve net
The on-line training of network model, shortens the training time of neural network model, is effectively improved the accurate of neural network model
Property.
Fig. 3 is a kind of flow chart of neural network model training method provided in an embodiment of the present invention, and the method is applied to
First client of medical image segmentation system, the first client is any client in multiple client, as shown in figure 3,
The method can include：
Step 201, acquisition include the sample to be split of multiple computer tomography CT images, and the CT images are comprising specified
The image of organ.
Step 202, every CT image is split using the local neural network model in the first client, obtain with
Corresponding first label image of every CT image, first label image is used to identify the designated organ that CT images are included.
Step 203, medical science sample is sent to server, in order to server according in medical science sample training server
First nerves network model, the first nerves network model is the neural network model of latest edition in server, medical science sample
Including：Every CT image and corresponding first label image of every CT image.
In sum, neural network model training method provided in an embodiment of the present invention, each in multiple client
After client carries out CT image segmentations using newest neural network model, by sending comprising multiple CT images to server and
The medical science sample of label image, enables the server to be trained newest neural network model according to medical science sample, realizes
The on-line training of neural network model, shortens the training time of neural network model, is effectively improved neural network model
Accuracy.
Fig. 4 is the flow chart of another neural network model training method provided in an embodiment of the present invention, and server can be with
Newest neural network model in by collecting the sample of each client upload come training server, the neural network model
Training method is applied in implementation environment as described in Figure 1, and the embodiment of the present invention is with the first client in multiple client
Example is illustrated, and in embodiments of the present invention, server, can also be by testing client after new neural network model is obtained
End is tested the new neural network model, and determining the validity of the new neural network model, alternatively, this
One client can also be default test client.As shown in figure 4, the method can include：
Step 301, the first client are obtained includes the sample to be split of multiple computer tomography CT images, CT figures
As the image comprising designated organ.
Alternatively, each sample to be split can be the case of patient.One case of patient is generally by hundreds of
CT images are constituted, and every CT image is the image comprising patient's designated organ.Illustratively, a case of certain sufferer includes
300 CT images, wherein, every CT image is the image comprising patient's designated organ, example, and this every CT image can be
The CT images of patient chest, the designated organ comprising the patient chest, then 300 CT images can constitute a sample to be split
This.
Step 302, the first client are carried out using the local neural network model in the first client to every CT image
Segmentation, obtains the first label image corresponding with every CT image.
Optionally, the local neural network model can download the latest edition for obtaining for the first client from server
Neural network model, first label image be used for identify the designated organ that corresponding CT images are included.
May be stored with the neural network model of multiple versions in practical application, in server, the nerve net of each version
Network model can correspond to a version number for the unique mark neural network model, and the version number is generally divided by server
Match somebody with somebody, such as version number can be 1.1 or 1.2 etc..First client periodically can download nerve net to server request
Network model, server can provide neural network model after receiving the request to the first client, and usual server is each
The neural network model that client is provided is the neural network model of latest edition；Or, server can be to each client
End periodically pushes the neural network model of latest edition, therefore, what the first client was obtained is the neutral net mould of latest edition
Type.
It is common, when the split order of CT images during the first client receives instruction for splitting sample to be split,
Can first judge the local neural network model in first client whether be latest edition neural network model, local god
Through network model for latest edition neural network model when, CT images point are carried out using the neural network model of the latest edition
Cut, when local neural network model is not for the neural network model of latest edition, the first client can be downloaded most from server
The neural network model of redaction, and neural network model using the latest edition carries out CT image segmentations.In practical application,
First client can obtain the version number of the neural network model of latest edition in server, by by the version number and first
The version number of the local neural network model in client is compared, it may be determined that the local neutral net in the first client
Model whether be latest edition neural network model, when the version number of the neural network model of latest edition in server is more than
During the version number of the local neural network model in the first client, the local neural network model in the first client is not most
The neural network model of redaction；When the version number of the neural network model of latest edition in server is equal in the first client
Local neural network model version number when, the local neural network model in the first client is the nerve net of latest edition
Network model.Illustratively, it is assumed that the version number of the neural network model of latest edition is in the server that the first client is obtained
1.5, the version number of the local neural network model in the first client is 1.2, it is known that 1.5>1.2 can determine the first client
Local neural network model in end is not the neural network model of latest edition.
When first client is split using local neural network model to every CT image, using designated organ as point
Cut target, you can obtain the first label image corresponding with every CT image, first label image is used to identify corresponding CT
The designated organ that image is included.Illustratively, the designated organ can be liver organ.Alternatively, the first label for obtaining is split
Image can be bianry image, then the pixel value of the corresponding pixel of designated organ can be 1, non-designated device in the first label image
The pixel value of the corresponding pixel in official position can be 0, or, the first label image that segmentation is obtained can also be gray level image,
Then the pixel value of the corresponding pixel of designated organ can be 255 in the first label image, the corresponding pixel of non-designated organ sites
Pixel value can be 0.Illustratively, when splitting to every CT image using local neural network model, liver may be selected
Organ can obtain the first label image including liver organ corresponding with every CT image as designated organ, then, and this first
The pixel value of the corresponding pixel of liver is 1 in label image, and the pixel value of the corresponding pixel of non-liver region is 0.
Alternatively, the present embodiments relate to neural network model can be convolutional neural networks.
It should be noted that before being split to every CT image using local neural network model, can be to every
Opening CT images carries out image preprocessing.The preprocessing process includes：
Step a, reading CT images, and record the window width W of CT images, window an information M.
Usually, CT images are digital imaging and communications in medicine (English：Digital Imaging and
Communications in Medicine；Referred to as：DICOM) the image of form.
Window width refers to the CT value scopes of CT images.The width of window width influences the contrast of image, the CT value models that narrow window width shows
Enclose small, the CT value amplitudes of every grade of GTG representative are small, then picture contrast is strong.Window position refers to the average of window width upper and lower limit CT values.Window
The height of position influences the brightness of image, and the low brightness of image in window position is high white, and window position hi-vision brightness is low in black.
Step b, the window width according to CT images and window position information determine the intensity profile scope of CT images according to the first formula.
First formula is：α=M-W/2, β=M+W/2；Wherein, W is the window width of CT images, and M is the window position of CT images, α
It is the minimum value of the corresponding pixel of designated organ in DICOM images, β is the maximum of the corresponding pixel of designated organ in CT images
Value.
Step c, greyscale transformation is carried out to CT images according to the intensity profile scope of CT images.
It is alternatively possible to carry out greyscale transformation to CT images according to the second formula, second formula is：
Wherein, C (i, j) is the gray value of the pixel at CT image (i, the j) position before conversion, after D (i, j) is for conversion
CT image (i, j) position at pixel gray value.
Step d, morphological dilation is done to the CT images after greyscale transformation, and obtain the CT images after expansive working
Connected region.
Step e, the target connected region for determining CT images, and CT images are cut according to the target connected region,
CT images after being cut.
It is alternatively possible to the size of occupied area determines target connected region in CT images according to designated organ.Example
Ground, when designated organ is liver, due to liver, occupied area is maximum in every CT image, therefore, it can connect the target
Logical region is defined as largest connected region.
Alternatively, when being cut to CT images, with the center of target connected region, the edge of target connected region is made
Edge with the CT images after cutting is tangent.Due to including multiple organs in every CT image, if directly to including multiple
The CT images of organ are split, and the accuracy of its splitting speed and segmentation result will necessarily be subject to a certain degree of influence, because
This, before being split to every CT image using local neural network model, target that can be according to where designated organ
The size of connected region is cut to improve the speed of segmentation and the accuracy of segmentation result to CT images.
Step f, by the CT image scalings after cutting to target size, and the image after scaling is saved as into target image lattice
Formula.
When being trained to neural network model, the size of every image for training is identical.But, designated organ
Size in every CT image may be different, so after the size of the target connected region according to where designated organ cuts
CT images may have different sizes, accordingly, it would be desirable to by the CT image scalings after cutting to target size.
Illustratively, the target size can be 301*400.In practical application, the target size can be carried out as needed
Adjustment, the embodiment of the present invention is not limited herein.
Alternatively, target image form can be the forms such as BMP forms, jpeg format and PNG format.
Alternatively, after being split to every CT image using local neural network model, can also be using artificial
Manual mode is finely adjusted to segmentation result, and fine setting can show as in segmentation result cut portion image or in segmentation
Increase parts of images in result, further to improve the accuracy of segmentation result.
In practical application, the split order of CT images in the first client receives instruction for splitting sample to be split
Afterwards, if local neural network model is not the neural network model of latest edition, the first client can be presented to user and pointed out
Information, for pointing out user to carry out the renewal of neural network model, if user indicates not carrying out the renewal of neural network model, the
One client can carry out CT image segmentations using local neural network model, because the sample for finally obtaining is not using newest
What the neural network model treatment of version was obtained, the first client can not upload the sample, it is also possible to when the sample is uploaded
The version number of the neural network model corresponding to the sample is marked, is determined whether to use the sample according to preset rules by server
Carry out the training of neural network model.For example, server can be instructed using the sample to its corresponding neural network model
Practice, it is also possible to the neural network model of latest edition is trained using the sample.
Step 303, the first client send to server medical science sample, and the medical science sample includes：Every CT image and
Corresponding first label image of every CT image.
It should be noted that the first client can the client be in idle condition (namely do not carry out other number
According to treatment or transmission state) when, medical science sample is sent to server, with avoid influence client other data place
Reason efficiency.
Optionally, client periodically can also be sent to server newly-increased medical science sample.
The medical science sample that step 304, server send according to multiple client, the first nerves network in training server
Model, the first nerves network model is the neural network model of latest edition in server.
Existing neural network training method, the sample of all collections can be used for the training of neutral net.But, work as receipts
When some of sample of collection sample has error and larger error, if continuing to use the sample training neutral net, not only will not
Increase training precision, can also bring opposite result.Therefore, server receive multiple client send medical science sample it
Afterwards, the first sample for receiving can be screened, then using the sample by screening as training sample, to server
In first nerves network model be trained, to improve the convergence rate and training precision of neutral net.Illustratively, the screening
Process can be medical science sample inaccurate in the medical science sample for deleting multiple client transmission.
Alternatively, as shown in figure 5, the medical science sample sent according to multiple client, the first nerves in training server
The process of network model can include：
Inaccurate medical science sample, is instructed in the medical science sample that step 3041, server deletion multiple client send
Practice sample.
Alternatively, the process of inaccurate medical science sample in the medical science sample that server deletion multiple client sends, can
To there is various achievable modes, the embodiment of the present invention is illustrated in following two achievable modes as an example.
The first can realize mode, and the deletion of inaccurate medical science sample can be realized by artificial screening, specifically may be used
To include：
Step a1, server show multiple mask images of first sample successively, and each mask image is by first sample
A label image be superimposed upon on corresponding CT images formed, the first sample be multiple client send medical science sample in
Any sample.
Alternatively, server can be provided with input/output interface, and it is defeated to connect outside input by the input/output interface
Go out equipment, for example, the input-output equipment includes display screen, server can be on the display screen of the input-output equipment successively
Multiple mask images of first sample are shown, the display screen can be the display screen of the maintained equipment for server configuration.First
Multiple mask images of sample refer to that in first sample a label image is superimposed upon on corresponding CT images and forms image,
In mask image, CT images are shown in the way of artwork, label image is shown in the way of certain transparency
Show, and the corresponding pixel of designated organ has certain pixel value in label image.By such display mode, can be conveniently
Ground carries out control and checks to the designated organ in label image and CT images, in order to judge the accuracy of segmentation result.Example
Ground, certain mask image that display screen shows can be as in Figure 6-1, it is assumed that in the mask image, with an image portion for filling
It is divided into the corresponding image section of designated organ in label image, schemes by CT of the image section of oblique line filling in the mask image
The corresponding image section of designated organ as in, checks, it can be seen that designated organ pair in CT images by the mask image
The corresponding image of designated organ has the part that overlaps each other and not lap in the image and label image answered, and this overlaps each other
Part be reflected as splitting accurate part in segmentation result, nonoverlapping part is reflected as splitting in segmentation result inaccurate
Part, by the contrast, it can be seen that the degree of accuracy of segmentation result, when the poor accuracy to a certain extent when, you can think
The segmentation result is inaccurate.Dotted line can be used for the center of sign picture picture in figure, can judge to specify by the center
The position of organ.
Contrasted for the ease of user, server simultaneously can also include in mask figure the corresponding 3-D view of first sample
In the display interface of picture, the 3-D view is multiple the CT images generation in first sample.Illustratively, it is same in Fig. 6-2
When show the mask image (upper left) and 3-D view (bottom right) of first sample, meanwhile, checked for the ease of user, Fig. 6-2
In also show the elevational perspective view (upper right) and left view perspective view (lower-left) of abdominal cavity of patients, dotted line can be used for sign picture picture in figure
Center, the position of designated organ can be judged by the center.Analysis to Fig. 6-2, can corresponding reference pair
The analysis of Fig. 6-1, here is omitted.
It is right that step b1, interface or interface where first sample where the artificial mask image in office of server reception are triggered
The deletion action of first sample.
Alternatively, when attendant checks in server to multiple mask images of first sample, when it is thought
Multiple mask images are inaccurate when causing the first sample inaccurate, and attendant can be manually deleted first sample,
Attendant can manually deletion of the interface triggering to first sample at interface where any mask image or where first sample
Operation, illustratively, the deletion action can be that attendant is produced by mouse in corresponding interface operation.For example, each
Interface where mask image or interface where each sample can be provided with the corresponding deletion button of first sample, when artificial point
When hitting the deletion button, generation is corresponding to delete instruction.
Step c1, server are deleted according to deletion action using first sample as inaccurate sample.
Alternatively, after server have received the deletion action to first sample of artificial triggering, can be deleted according to this
Division operation is deleted corresponding first sample as inaccurate sample.
Illustratively, the deletion action of this pair of first sample is triggered at the corresponding interface of mask image in attendant,
For example the interface can be the interface shown in Fig. 6-1 or Fig. 6-2, that is to say, server have received artificial in any mask
Triggered on interface where image after deletion action, set according to system, attendant can given tacit consent to and think the mask image pair
The first sample answered is inaccurate sample, then server can be deleted in the interface of current mask image to first sample
Remove, or, attendant accordingly deletes first sample in the corresponding interface triggering deletion action of first sample, server.
It is capable of achieving mode second, inaccurate medical science sample is realized by server and the artificial screening mode for combining
Delete, can specifically include：
Step a2, server are by every CT image in first sample using default standard neural network mould in server
Type is split, and obtains the second label image corresponding with every CT image, and second label image is used to identify CT image bags
The designated organ for containing, the first sample is any sample in the medical science sample that multiple client sends.
Alternatively, in server default standard neural network model be typically that attendant pre-sets it is most reliable
Neural network model, can be the neural network model of the latest edition of storage in server illustratively.
Step b2, server judge that whether the segmentation image difference of CT images, more than preset difference value threshold value, splits image difference
Value is the image difference of the first label image corresponding to same CT images and the second label image.
Alternatively, the image difference of the first label image and the second label image can be by designated organ in correspondence image
Area difference represent.The area difference can show as the pixel number purpose difference that designated organ in image is included.
In practical application, preset difference value threshold value can be configured according to actual conditions, and the embodiment of the present invention is not done to it
It is specific to limit.
Step c2, when CT image accounting in first sample of the segmentation image difference more than preset difference value threshold value is more than pre-
If during ratio, server stores to manual confirmation database first sample, and the sample in manual confirmation database is used for by people
Work is confirmed whether to delete, and performs step e2.
Illustratively, above-mentioned default ratio can be 5%.
Alternatively, segmentation image difference is more than accounting of the CT images of preset difference value threshold value in first sample, Ke Yitong
Over-segmentation image difference is total more than the CT images included in the sum and first sample of the CT images of preset difference value threshold value
Ratio is showed.Illustratively, it is assumed that default ratio is 5%, the sum of the CT images included in first sample is 300, segmentation figure
Aberration value is 21 more than the sum of the CT images of preset difference value threshold value, then split image difference and scheme more than the CT of preset difference value threshold value
As the accounting in first sample is 7%, it is known that 7%>5%, then server the first sample is stored to manual confirmation data
Storehouse, by manual confirmation, whether the first sample deletes.
Step d2, it is not more than more than the accounting of the CT images in first sample of preset difference value threshold value when segmentation image difference
During default ratio, server stores to training sample database first sample, and the sample in training sample database is used to instruct
Practice neural network model, perform step e2.
Step e2, server are by the sample for confirming not delete in manual confirmation database and/or training sample database
Sample be defined as training sample.
It is alternatively possible to will confirm that the sample do not deleted is defined as training sample in manual confirmation database, it is also possible to will
Sample in training sample database is defined as training sample, it is also possible to while will confirm what is do not deleted in manual confirmation database
Sample standard deviation in sample and training sample database is defined as training sample, and the embodiment of the present invention is not specifically limited to it.Show
Example ground, it is assumed that confirm that the sample do not deleted is sample 1 in manual confirmation database, the sample in training sample database is sample
2 and sample 3, then server sample 1 can be defined as training sample, it is also possible to sample 2 and sample 3 are defined as to train sample
This, or, it is also possible to sample 1, sample 2 and sample 3 are defined as training sample.
The first realized side of inaccurate medical science sample in the medical science sample that server deletion multiple client sends
Formula needs artificial participation, and the mode for manually participating in Screening Samples makes it possible to collect more abnormal sample, so that
Split accurate sample range more extensive；Being capable of achieving mode second can reduce artificial workload, can realize nerve
While network model automation updates, it is ensured that the accuracy of segmentation result.
Step 3042, server are using the first nerves network model in training sample training server.
Alternatively, server can pack all training samples, then be used to the training sample after packing train the
One neural network model, the first nerves network model is the neural network model of latest edition in server.In the instruction of packing
Practice in sample, every CT image is corresponded with corresponding first label image.
It is alternatively possible to be trained to first nerves network model using feedforward back-propagation algorithm, it would however also be possible to employ
Other algorithms are trained to first nerves network model, and the training process may be referred to prior art, the embodiment of the present invention pair
It is repeated no more.
Step 305, server send to the first client nervus opticus network model.
The nervus opticus network model is the medical science sample sent according to multiple client, the first god in training server
Through the neural network model obtained by network model.
Alternatively, because the performance of the nervus opticus network model after training is not necessarily better than the first nerves before training
Network model (for example, train selected error sample excessive, cause to train the neural network model precision for obtaining relatively low), because
This, server after the training to nervus opticus network model is completed, send out by the nervus opticus network model that will can be trained
Multiple test clients are delivered to, the image segmentation for testing nervus opticus network model.In embodiments of the present invention,
One client can also be default test client, therefore, server can be sent to first nervus opticus network model
Client.
Alternatively, can be the part client in multiple client for the multiple test clients tested, it is also possible to
It is whole clients.In actual applications, specifically which client serving as test client by can be carried out according to actual needs
Setting, the embodiment of the present invention is not specifically limited to it.
Step 306, the first client using after nervus opticus network model, display model marking interface.
Alternatively, the first client can be default test client.Sent to server by medical science sample,
Serve as the nervus opticus network model that the first client of test client can be sent with the reception server, and use the second god
After network model, display model marking interface.User (usually medical personnel) can be in model marking interface to the
Two neural network models are given a mark, and the marking can characterize nervus opticus network model to the accurate of medical image segmentation result
Property, give a mark higher, then it represents that user thinks that segmentation result is more accurate.Illustratively, the marking can give a mark for hundred-mark system.Marking
Specific standards can be set according to practical application, and the embodiment of the present invention is not specifically limited to it.
Illustratively, Fig. 7-1 is a kind of model marking interface schematic diagram provided in an embodiment of the present invention, model marking interface
Marking region 01 is only show, Fig. 7-2 is another model marking interface schematic diagram provided in an embodiment of the present invention, and the model is beaten
Interface shows marking region 01, segmentation result viewing area 02 and image display area to be split 03.Fig. 7-1 and Fig. 7-
User can give a mark to the judgment criterion according to oneself to nervus opticus network model in the marking region 01 of 2 displays, with
Characterize accuracy of the nervus opticus network model to medical image segmentation result.In practical application, model marking interface can be with
There are other forms, the embodiment of the present invention is not specifically limited to it.
Step 307, the first client receive marking of the user to nervus opticus network model by model marking interface.
Illustratively, in the marking region 01 that Fig. 7-1 and Fig. 7-2 shows, user can click on marking region by mouse
In numeral and " " nervus opticus network model is given a mark, can by mouse click on marking region in "×" fight each other
Divide and modify.
Step 308, the first client will be given a mark and sent to server.
The marking of step 309, server according to the first client to nervus opticus network model, determines test result.
Alternatively, after server receives multiple test clients to the marking of nervus opticus network model, can basis
The marking determines test result, and the test result can be the average mark of multiple test client marking, or multiple is surveyed
Minimum point of examination client marking, or, it is also possible to it is that test result effect is calculated according to multiple test clients
The fraction for going out, illustratively, multiple test clients can be indicated to test result effect by weights.It is actual to answer
In, determine that the mode of test result can be according to actually being set according to multiple marking, the embodiment of the present invention is not done to it
It is specific to limit.
Illustratively, it is assumed that have three clients in multiple client for test client, respectively client 1, client 2
The marking of three clients received with client 3, server is respectively 80,85 and 90, and three clients are to test result shadow
The weights of loud degree are respectively 0.2,0.3 and 0.5.When the average mark that test result is multiple test client marking, can
To determine test result as (80+85+90)/3=85；When the minimum timesharing that test result is multiple test client marking, can
To determine test result as 80；When test result is calculated according to three test clients to test result effect
During fraction, it may be determined that test result is 80*0.2+85*0.3+90*0.5=86.5.
Whether step 3010, server judge test result more than default passing score.
Alternatively, default passing score can be the accurate of the segmentation result of the nervus opticus network model for representing current
Property higher than training before first nerves network segmentation result accuracy lowest fractional.The default passing score can be
A fractional value being manually set, for example, it is 80 that can be manually set the default passing score；Can also be to first nerves
Test result when network model is tested, for example, test result when testing first nerves network model is 82,
Then the default passing score could be arranged to 82；Or, the numerical value of the default passing score can according to circumstances be set
Fixed, the embodiment of the present invention is not specifically limited to it.
Test result is more than default passing score, that is to say, the segmentation knot of currently used nervus opticus network model
Accuracy of the accuracy of fruit higher than the first nerves network segmentation result before training；Test result is not more than default qualifying and divides
Number, that is to say, the accuracy of the segmentation result of current nervus opticus network model is not higher than the first nerves network before training
Segmentation result accuracy.
Step 3011, when test result is more than default passing score, be defined as nervus opticus network model by server
The neural network model of latest edition.
Illustratively, it is assumed that default passing score is 80, the test result of nervus opticus network model is 82, it is known that 82>
80, that is to say that test result more than default passing score, then nervus opticus network model can be defined as the god of latest edition
Through network model.
Step 3012, when test result is not more than default passing score, server determines first nerves network model
It is the neural network model of latest edition.
Illustratively, it is assumed that default passing score is 80, the test result of nervus opticus network model is 79, it is known that 79<
80, that is to say that test result is not more than default passing score, then first nerves network model can be defined as latest edition
Neural network model.
Alternatively, the multiple client for being included in medical image segmentation system is schemed using local neural network model to CT
After as being split, local neural network model can be given a mark, that is to say, even if certain client is not test client
End, the client can also give a mark to local neural network model, and the marking can give a mark for conventional, and client can lead to
Cross the segmentation accuracy of the marking its local neural network model for using to server feedback.Its scoring process can be：It is aobvious
Representation model marking interface, receives marking of the user to local neural network model, and marking is sent by model marking interface
To server.Test client may be referred to nervus opticus to the associated description of the marking and the process that implements of the marking
The marking of network model, here is omitted.
It should be noted that the multiple client included in medical image segmentation system is to the normal of first nerves network model
Rule marking for can selection operation, Compulsory Feature is not done to it, that is to say, CT images are being entered using local neural network model
After row segmentation, user can select to give a mark local neural network model or not to the marking of local neural network model.
Alternatively, may be stored with the neural network model of multiple versions in server, wherein, the nerve net of latest edition
Network model is the currently used neural network model of client, and the neural network model of other versions is for background maintenance personnel
Recorded or referred to.
Due to that in medical domain, can set up corresponding neural network model for Different Organs, client is based on should
Neural network model carries out medical image segmentation, and the label image obtained after segmentation includes designated organ, above-mentioned implementation of the invention
Example is mainly illustrated so that the designated organ is as liver as an example, and in practical application, the designated organ can also be heart or brain
The training method of the neural network model of liver is may be referred to Deng, the training method of corresponding neural network model, the present invention is real
Example is applied not repeat this.
It should be noted that the sequencing of neural network model training method step provided in an embodiment of the present invention can be with
Suitably adjusted, step according to circumstances can also accordingly be increased and decreased, any one skilled in the art is at this
Apply in the technical scope of exposure, the method that can readily occur in change should all cover within the protection domain of the application, therefore
Repeat no more.
In sum, neural network model training method provided in an embodiment of the present invention, by server in multiple clients
After each client in end carries out CT image segmentations using newest neural network model, the bag that multiple client sends is received
Medical science sample containing multiple CT images and label image, the marking according to test client determines the neutral net mould of latest edition
Type, and neural network model to latest edition is trained, and realizes the on-line training of neural network model, shortens nerve net
The training time of network model, it is effectively improved the accuracy of neural network model.
Fig. 8 is a kind of structured flowchart of neural network model trainer 800 provided in an embodiment of the present invention, the nerve net
Network model training apparatus 800 are applied to the server of medical image segmentation system, as shown in figure 8, the neural network model is trained
Device 800 can include：
First receiver module 801, the medical science sample for receiving multiple client transmission, wherein, the first client sends
Medical science sample include：Multiple CT images and corresponding first label image of every CT image, the first label image are used to identify
The designated organ that CT images are included, the first label image is that the first client uses local neural network model to multiple CT images
Carry out splitting what is obtained, the first client is any client in multiple client.
Training module 802, for the medical science sample sent according to multiple client, the first nerves net in training server
Network model, first nerves network model is the neural network model of latest edition in server.
In sum, neural network model trainer provided in an embodiment of the present invention, each in multiple client
After client carries out CT image segmentations using newest neural network model, multiple client is received by the first receiver module and is sent out
The medical science sample comprising multiple CT images and label image for sending, training module is instructed to the neural network model of latest edition
Practice, realize the on-line training of neural network model, shorten the training time of neural network model, be effectively improved nerve net
The accuracy of network model training.
As shown in figure 9, training module 802 can include：
Submodule 8021 is deleted, for medical science sample inaccurate in the medical science sample for deleting multiple client transmission, is obtained
To training sample.
Training submodule 8022, for using the first nerves network model in training sample training server.
Alternatively, delete submodule 8021 specifically for：
Multiple mask images of first sample are shown successively, and each mask image is by a label image in first sample
It is superimposed upon on corresponding CT images and is formed, first sample is any sample in the medical science sample that multiple client sends.
Receive the artificial interface where any mask image or interface triggering where the first sample to described
The deletion action of first sample.
Deleted first sample as inaccurate sample according to deletion action.
Alternatively, delete submodule 8021 specifically for：
Every CT image in first sample is split using default standard neural network model in server, is obtained
To the second label image corresponding with every CT image, second label image is used to identify the designated organ that CT images are included,
First sample is any sample in the medical science sample that multiple client sends.
Whether the segmentation image difference of CT images is judged more than preset difference value threshold value, and segmentation image difference is corresponding to same
First label image of CT images and the image difference of the second label image.
It is more than default ratio when segmentation image difference is more than accounting of the CT images of preset difference value threshold value in first sample
When, first sample is stored to manual confirmation database, whether the sample in manual confirmation database is used to be deleted by manual confirmation
Remove.
It is not more than default ratio when segmentation image difference is more than accounting of the CT images of preset difference value threshold value in first sample
During value, first sample is stored to training sample database, the sample in training sample database is used to train neutral net mould
Type.
Sample in the sample for confirming not delete in manual confirmation database and/or training sample database is defined as instruction
Practice sample.
As shown in Figure 10, the neural network model trainer 800 can also include：
Sending module 803, for nervus opticus network model to be sent to multiple test clients, nervus opticus network mould
Type is the nerve net obtained by the first nerves network model in the medical science sample training server sent according to multiple client
Network model.
Second receiver module 804, for receiving marking of multiple test clients to nervus opticus network model.
First determining module 805, for the marking according to multiple test clients to nervus opticus network model, it is determined that surveying
Examination fraction.
Judge module 806, for judging test result whether more than default passing score.
Second determining module 807, it is for being more than default passing score when test result, nervus opticus network model is true
It is set to the neural network model of latest edition.
3rd determining module 808, for being not more than default passing score when test result, by first nerves network model
It is defined as the neural network model of latest edition.
In sum, neural network model trainer provided in an embodiment of the present invention, each in multiple client
After client carries out CT image segmentations using newest neural network model, multiple client is received by the first receiver module and is sent out
The medical science sample comprising multiple CT images and label image for sending, the marking according to test client determines the nerve of latest edition
Network model, training module is trained to the neural network model of latest edition, realizes the on-line training of neural network model,
The training time of neural network model is shortened, the accuracy of neural network model training is effectively improved.
Figure 11 is a kind of structured flowchart of neural network model trainer 900 provided in an embodiment of the present invention, the nerve
Network model trainer 900 is applied to the first client of medical image segmentation system, as shown in figure 11, the neutral net mould
Type trainer 900 can include：
Acquisition module 901, includes the sample to be split of multiple computer tomography CT images, the CT images for obtaining
Image comprising designated organ.
Segmentation module 902, for being divided every CT image using the local neural network model in the first client
Cut, obtain the first label image corresponding with every CT image, first label image is used to identifying that CT images to include specifies
Organ.
Sending module 903, for medical science sample to be sent to server, takes in order to server according to medical science sample training
First nerves network model in business device, the first nerves network model is the neural network model of latest edition in server,
Medical science sample includes：Every CT image and corresponding first label image of every CT image.
In sum, neural network model trainer provided in an embodiment of the present invention, each in multiple client
After client carries out CT image segmentations using newest neural network model, sent to server by sending module and include multiple
The medical science sample of CT images and label image, enables the server to enter the neural network model of latest edition according to medical science sample
Row training, realizes the on-line training of neural network model, shortens the training time of neural network model, is effectively improved god
Through the accuracy that network model is trained.
It should be noted that the first client can also be default test client.When the first client is default
During test client, as shown in figure 12, the neural network model trainer 900 can also include：
First receiver module 904, for the nervus opticus network model that the reception server sends.
Display module 905, for using after nervus opticus network model, display model marking interface.
Second receiver module 906, for receiving marking of the user to nervus opticus network model by model marking interface.
Marking sending module 907, for marking to be sent to server.
In sum, neural network model trainer provided in an embodiment of the present invention, each in multiple client
After client carries out CT image segmentations using newest neural network model, sent to server by sending module and include multiple
The medical science sample of CT images and label image, marking sending module sends to server the marking of test client, in order to
Server determines the neural network model of latest edition, and the neural network model of latest edition is instructed according to medical science sample
Practice, realize the on-line training of neural network model, shorten the training time of neural network model, be effectively improved nerve net
The accuracy of network model training.
It is apparent to those skilled in the art that, for convenience and simplicity of description, the device of foregoing description,
The specific work process of module and submodule, may be referred to the corresponding process in preceding method embodiment, will not be repeated here.
The embodiment of the present invention additionally provides a kind of medical image segmentation system, and the medical image segmentation system includes：At least
One server and at least one first clients being connected with server, the server are trained including above-mentioned neural network model
Device 800, first client includes above-mentioned neural network model trainer 900.
One of ordinary skill in the art will appreciate that realizing that all or part of step of above-described embodiment can be by hardware
To complete, it is also possible to instruct the hardware of correlation to complete by program, described program can be stored in a kind of computer-readable
In storage medium, storage medium mentioned above can be read-only storage, disk or CD etc..
The foregoing is only presently preferred embodiments of the present invention, be not intended to limit the invention, it is all it is of the invention spirit and
Within principle, any modification, equivalent substitution and improvements made etc. should be included within the scope of the present invention.
Claims (10)
1. a kind of neural network model training method, it is characterised in that be applied to the server of medical image segmentation system, it is described
Method includes：
The medical science sample that multiple client sends is received, wherein, the medical science sample that the first client sends includes：Multiple computers
Tomoscan CT images and corresponding first label image of every CT image, first label image are used to identify the CT figures
As comprising designated organ, first label image is first client using local neural network model to described many
Opening CT images carries out splitting what is obtained, and first client is any client in the multiple client；
According to the medical science sample that the multiple client sends, the first nerves network model in the server is trained, it is described
First nerves network model is the neural network model of latest edition in the server.
2. method according to claim 1, it is characterised in that the medical science sample sent according to the multiple client
This, trains the first nerves network model in the server, including：
Inaccurate medical science sample in the medical science sample that the multiple client sends is deleted, training sample is obtained；
The first nerves network model in the server is trained using the training sample.
3. method according to claim 2, it is characterised in that the medical science sample that the multiple client of deletion sends
In inaccurate medical science sample, obtain training sample, including：
Multiple mask images of first sample are shown successively, and each mask image is by a label image in the first sample
It is superimposed upon on corresponding CT images and is formed, the first sample is any sample in the medical science sample that the multiple client sends
This；
Receive the artificial interface where any mask image or interface triggering where the first sample to described first
The deletion action of sample；
Deleted the first sample as inaccurate sample according to the deletion action.
4. method according to claim 2, it is characterised in that the medical science sample that the multiple client of deletion sends
In inaccurate medical science sample, obtain training sample, including：
Every CT image in the first sample is divided using default standard neural network model in the server
Cut, obtain the second label image corresponding with every CT images, second label image is used to identify the CT images
Comprising designated organ, the first sample is any sample in the medical science sample that the multiple client sends；
Whether judge the segmentation image difference of the CT images more than preset difference value threshold value, the segmentation image difference be corresponding to
First label image of same CT images and the image difference of second label image；
It is more than default ratio when segmentation image difference is more than accounting of the CT images of preset difference value threshold value in the first sample
When, the first sample is stored to manual confirmation database, the sample in the manual confirmation database is used for by artificial true
Recognize and whether delete；
It is not more than default ratio when segmentation image difference is more than accounting of the CT images of preset difference value threshold value in the first sample
During value, the first sample is stored to training sample database, the sample in the training sample database is used to train god
Through network model；
Sample in the sample for confirming not delete in the manual confirmation database and/or the training sample database is determined
It is the training sample.
5. according to any described method of Claims 1-4, it is characterised in that sent according to the multiple client described
Medical science sample, train after the first nerves network model in the server, methods described also includes：
Nervus opticus network model is sent to multiple test clients, the nervus opticus network model is according to the multiple
The neural network model obtained by first nerves network model in server described in the medical science sample training that client sends；
Receive marking of the multiple test client to the nervus opticus network model；
Marking according to the multiple test client to the nervus opticus network model, determines test result；
Judge the test result whether more than default passing score；
When the test result is more than default passing score, the nervus opticus network model is defined as the god of latest edition
Through network model；
When the test result is not more than default passing score, the first nerves network model is defined as latest edition
Neural network model.
6. a kind of neural network model training method, it is characterised in that be applied to the first client of medical image segmentation system,
Methods described includes：
Acquisition includes the sample to be split of multiple computer tomography CT images, figure of the CT images comprising designated organ
Picture；
Every CT image is split using the local neural network model in first client, obtain with it is described every
Corresponding first label image of CT images, first label image is used to identify the designated organ that the CT images are included；
Medical science sample is sent to the server, in order to the server according to the medical science sample training server
In first nerves network model, the first nerves network model is the neutral net mould of latest edition in the server
Type, the medical science sample includes：Every CT images and corresponding first label image of every CT images.
7. a kind of neural network model trainer, it is characterised in that be applied to the server of medical image segmentation system, it is described
Device includes：
First receiver module, the medical science sample for receiving multiple client transmission, wherein, the medical science sample that the first client sends
Originally include：Multiple CT images and corresponding first label image of every CT image, first label image are used to identifying described
The designated organ that CT images are included, first label image is that first client uses local neural network model to institute
Stating multiple CT images carries out splitting what is obtained, and first client is any client in the multiple client；
Training module, for the medical science sample sent according to the multiple client, trains the first nerves in the server
Network model, the first nerves network model is the neural network model of latest edition in the server.
8. device according to claim 7, it is characterised in that the training module, including：
Submodule is deleted, for medical science sample inaccurate in the medical science sample for deleting the multiple client transmission, is instructed
Practice sample；
Training submodule, for training the first nerves network model in the server using the training sample.
9. a kind of neural network model trainer, it is characterised in that be applied to the first client of medical image segmentation system,
Described device includes：
Acquisition module, the sample to be split of multiple computer tomography CT images is included for obtaining, and the CT images are included
The image of designated organ；
Segmentation module, for being split to every CT image using the local neural network model in first client,
The first label image corresponding with every CT images is obtained, first label image is included for identifying the CT images
Designated organ；
Sending module, for medical science sample to be sent to the server, in order to the server according to the medical science sample
The first nerves network model in the server is trained, the first nerves network model is latest edition in the server
Neural network model, the medical science sample includes：Every CT images and every CT images corresponding described first
Label image.
10. a kind of medical image segmentation system, it is characterised in that the medical image segmentation system includes：At least one service
Device and at least one first clients being connected with the server；
The server includes the neural network model trainer described in claim 7 or 8；
Each described first client includes the neural network model trainer described in claim 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710002230.5A CN106709917B (en) | 2017-01-03 | 2017-01-03 | Neural network model training method, device and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710002230.5A CN106709917B (en) | 2017-01-03 | 2017-01-03 | Neural network model training method, device and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106709917A true CN106709917A (en) | 2017-05-24 |
| CN106709917B CN106709917B (en) | 2020-09-11 |
Family
ID=58905798
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710002230.5A Active CN106709917B (en) | 2017-01-03 | 2017-01-03 | Neural network model training method, device and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106709917B (en) |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107817204A (en) * | 2017-11-01 | 2018-03-20 | 中国科学院地质与地球物理研究所 | A kind of shale micro-void structures analysis method and device |
| CN108205527A (en) * | 2016-12-16 | 2018-06-26 | 深圳联友科技有限公司 | A kind of method for drafting and device of engine data balance |
| CN108875508A (en) * | 2017-11-23 | 2018-11-23 | 北京旷视科技有限公司 | In vivo detection algorithm update method, device, client, server and system |
| CN109166107A (en) * | 2018-04-28 | 2019-01-08 | 北京市商汤科技开发有限公司 | A kind of medical image cutting method and device, electronic equipment and storage medium |
| CN109215094A (en) * | 2017-09-22 | 2019-01-15 | 上海联影医疗科技有限公司 | phase contrast image generating method and system |
| CN109214343A (en) * | 2018-09-14 | 2019-01-15 | 北京字节跳动网络技术有限公司 | Method and apparatus for generating face critical point detection model |
| CN109492698A (en) * | 2018-11-20 | 2019-03-19 | 腾讯科技（深圳）有限公司 | A kind of method of model training, the method for object detection and relevant apparatus |
| CN109544550A (en) * | 2018-12-05 | 2019-03-29 | 易必祥 | Based on CT images intelligent measurement and recognition methods and system |
| CN109615058A (en) * | 2018-10-24 | 2019-04-12 | 上海新储集成电路有限公司 | A kind of training method of neural network model |
| CN109727663A (en) * | 2018-09-30 | 2019-05-07 | 上海联影医疗科技有限公司 | The system and method for generating the neural network model for image procossing |
| CN109805950A (en) * | 2017-10-06 | 2019-05-28 | 佳能医疗系统株式会社 | Medical image-processing apparatus and medical image processing system |
| CN109977957A (en) * | 2019-03-04 | 2019-07-05 | 苏宁易购集团股份有限公司 | A kind of invoice recognition methods and system based on deep learning |
| CN110123249A (en) * | 2019-04-09 | 2019-08-16 | 苏州西能捷科技发展有限公司 | A kind of nasosinusitis detection device and application method |
| CN110232411A (en) * | 2019-05-30 | 2019-09-13 | 北京百度网讯科技有限公司 | Model distills implementation method, device, system, computer equipment and storage medium |
| CN110321864A (en) * | 2019-07-09 | 2019-10-11 | 西北工业大学 | Remote sensing images explanatory note generation method based on multiple dimensioned cutting mechanism |
| CN110353707A (en) * | 2018-03-26 | 2019-10-22 | 通用电气公司 | The training method and system of collimator boundary detection method |
| CN110503151A (en) * | 2019-08-26 | 2019-11-26 | 北京推想科技有限公司 | A kind of processing method and system of image |
| CN110570419A (en) * | 2019-09-12 | 2019-12-13 | 杭州依图医疗技术有限公司 | Method and device for acquiring characteristic information and storage medium |
| CN110766693A (en) * | 2018-09-06 | 2020-02-07 | 北京连心医疗科技有限公司 | Method for jointly predicting radiotherapy structure position based on multi-model neural network |
| CN110766694A (en) * | 2019-09-24 | 2020-02-07 | 清华大学 | Interactive segmentation method of three-dimensional medical image |
| CN110853024A (en) * | 2019-11-14 | 2020-02-28 | 北京推想科技有限公司 | Medical image processing method, medical image processing device, storage medium and electronic equipment |
| CN110969072A (en) * | 2018-09-30 | 2020-04-07 | 杭州海康威视系统技术有限公司 | Model optimization method and device and image analysis system |
| CN110969622A (en) * | 2020-02-28 | 2020-04-07 | 南京安科医疗科技有限公司 | Image processing method and system for assisting pneumonia diagnosis |
| CN111247556A (en) * | 2017-10-24 | 2020-06-05 | 爱克发医疗保健公司 | Training artificial neural networks on additional tasks while avoiding catastrophic interference |
| CN111665177A (en) * | 2020-06-11 | 2020-09-15 | 太原理工大学 | Laboratory protection system based on object recognition, toxic gas and heat source detection |
| CN111882048A (en) * | 2020-09-28 | 2020-11-03 | 深圳追一科技有限公司 | Neural network structure searching method and related equipment |
| CN112350995A (en) * | 2020-09-30 | 2021-02-09 | 山东众阳健康科技集团有限公司 | Image processing method, device, equipment and storage medium |
| CN113009077A (en) * | 2021-02-18 | 2021-06-22 | 南方电网数字电网研究院有限公司 | Gas detection method, gas detection device, electronic apparatus, and storage medium |
| CN113223101A (en) * | 2021-05-28 | 2021-08-06 | 支付宝(杭州)信息技术有限公司 | Image processing method, device and equipment based on privacy protection |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102955946A (en) * | 2011-08-18 | 2013-03-06 | 刘军 | Two-stage fast classifier based on linear classification tree and neural network |
| CN103489009A (en) * | 2013-09-17 | 2014-01-01 | 北方信息控制集团有限公司 | Pattern recognition method based on self-adaptation correction neural network |
| CN105160361A (en) * | 2015-09-30 | 2015-12-16 | 东软集团股份有限公司 | Image identification method and apparatus |
| CN105447498A (en) * | 2014-09-22 | 2016-03-30 | 三星电子株式会社 | A client device configured with a neural network, a system and a server system |
-
2017
- 2017-01-03 CN CN201710002230.5A patent/CN106709917B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102955946A (en) * | 2011-08-18 | 2013-03-06 | 刘军 | Two-stage fast classifier based on linear classification tree and neural network |
| CN103489009A (en) * | 2013-09-17 | 2014-01-01 | 北方信息控制集团有限公司 | Pattern recognition method based on self-adaptation correction neural network |
| CN105447498A (en) * | 2014-09-22 | 2016-03-30 | 三星电子株式会社 | A client device configured with a neural network, a system and a server system |
| CN105160361A (en) * | 2015-09-30 | 2015-12-16 | 东软集团股份有限公司 | Image identification method and apparatus |
Cited By (40)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108205527B (en) * | 2016-12-16 | 2022-01-18 | 深圳联友科技有限公司 | Drawing method and device for engine data balance |
| CN108205527A (en) * | 2016-12-16 | 2018-06-26 | 深圳联友科技有限公司 | A kind of method for drafting and device of engine data balance |
| CN109215094A (en) * | 2017-09-22 | 2019-01-15 | 上海联影医疗科技有限公司 | phase contrast image generating method and system |
| CN109805950A (en) * | 2017-10-06 | 2019-05-28 | 佳能医疗系统株式会社 | Medical image-processing apparatus and medical image processing system |
| CN111247556A (en) * | 2017-10-24 | 2020-06-05 | 爱克发医疗保健公司 | Training artificial neural networks on additional tasks while avoiding catastrophic interference |
| CN107817204B (en) * | 2017-11-01 | 2018-12-28 | 中国科学院地质与地球物理研究所 | A kind of shale micro-void structures analysis method and device |
| CN107817204A (en) * | 2017-11-01 | 2018-03-20 | 中国科学院地质与地球物理研究所 | A kind of shale micro-void structures analysis method and device |
| CN108875508B (en) * | 2017-11-23 | 2021-06-29 | 北京旷视科技有限公司 | Living body detection algorithm updating method, device, client, server and system |
| CN108875508A (en) * | 2017-11-23 | 2018-11-23 | 北京旷视科技有限公司 | In vivo detection algorithm update method, device, client, server and system |
| CN110353707A (en) * | 2018-03-26 | 2019-10-22 | 通用电气公司 | The training method and system of collimator boundary detection method |
| CN109166107A (en) * | 2018-04-28 | 2019-01-08 | 北京市商汤科技开发有限公司 | A kind of medical image cutting method and device, electronic equipment and storage medium |
| CN110766693B (en) * | 2018-09-06 | 2022-06-21 | 北京连心医疗科技有限公司 | Method for jointly predicting radiotherapy structure position based on multi-model neural network |
| CN110766693A (en) * | 2018-09-06 | 2020-02-07 | 北京连心医疗科技有限公司 | Method for jointly predicting radiotherapy structure position based on multi-model neural network |
| CN109214343A (en) * | 2018-09-14 | 2019-01-15 | 北京字节跳动网络技术有限公司 | Method and apparatus for generating face critical point detection model |
| CN109727663A (en) * | 2018-09-30 | 2019-05-07 | 上海联影医疗科技有限公司 | The system and method for generating the neural network model for image procossing |
| US11599796B2 (en) | 2018-09-30 | 2023-03-07 | Shanghai United Imaging Healthcare Co., Ltd. | Systems and methods for generating a neural network model for image processing |
| CN110969072B (en) * | 2018-09-30 | 2023-05-02 | 杭州海康威视系统技术有限公司 | Model optimization method, device and image analysis system |
| US11907852B2 (en) | 2018-09-30 | 2024-02-20 | Shanghai United Imaging Healthcare Co., Ltd. | Systems and methods for generating a neural network model for image processing |
| CN110969072A (en) * | 2018-09-30 | 2020-04-07 | 杭州海康威视系统技术有限公司 | Model optimization method and device and image analysis system |
| WO2020062262A1 (en) * | 2018-09-30 | 2020-04-02 | Shanghai United Imaging Healthcare Co., Ltd. | Systems and methods for generating a neural network model for image processing |
| CN109615058A (en) * | 2018-10-24 | 2019-04-12 | 上海新储集成电路有限公司 | A kind of training method of neural network model |
| CN109492698A (en) * | 2018-11-20 | 2019-03-19 | 腾讯科技（深圳）有限公司 | A kind of method of model training, the method for object detection and relevant apparatus |
| CN109544550A (en) * | 2018-12-05 | 2019-03-29 | 易必祥 | Based on CT images intelligent measurement and recognition methods and system |
| CN109544550B (en) * | 2018-12-05 | 2021-10-22 | 易必祥 | CT image-based intelligent detection and identification method and system |
| CN109977957A (en) * | 2019-03-04 | 2019-07-05 | 苏宁易购集团股份有限公司 | A kind of invoice recognition methods and system based on deep learning |
| CN110123249A (en) * | 2019-04-09 | 2019-08-16 | 苏州西能捷科技发展有限公司 | A kind of nasosinusitis detection device and application method |
| CN110123249B (en) * | 2019-04-09 | 2022-02-01 | 苏州西能捷科技发展有限公司 | Nasosinusitis detection device and use method thereof |
| CN110232411A (en) * | 2019-05-30 | 2019-09-13 | 北京百度网讯科技有限公司 | Model distills implementation method, device, system, computer equipment and storage medium |
| CN110232411B (en) * | 2019-05-30 | 2022-08-23 | 北京百度网讯科技有限公司 | Model distillation implementation method, device, system, computer equipment and storage medium |
| CN110321864A (en) * | 2019-07-09 | 2019-10-11 | 西北工业大学 | Remote sensing images explanatory note generation method based on multiple dimensioned cutting mechanism |
| CN110503151A (en) * | 2019-08-26 | 2019-11-26 | 北京推想科技有限公司 | A kind of processing method and system of image |
| CN110570419A (en) * | 2019-09-12 | 2019-12-13 | 杭州依图医疗技术有限公司 | Method and device for acquiring characteristic information and storage medium |
| CN110766694A (en) * | 2019-09-24 | 2020-02-07 | 清华大学 | Interactive segmentation method of three-dimensional medical image |
| CN110853024A (en) * | 2019-11-14 | 2020-02-28 | 北京推想科技有限公司 | Medical image processing method, medical image processing device, storage medium and electronic equipment |
| CN110969622A (en) * | 2020-02-28 | 2020-04-07 | 南京安科医疗科技有限公司 | Image processing method and system for assisting pneumonia diagnosis |
| CN111665177A (en) * | 2020-06-11 | 2020-09-15 | 太原理工大学 | Laboratory protection system based on object recognition, toxic gas and heat source detection |
| CN111882048A (en) * | 2020-09-28 | 2020-11-03 | 深圳追一科技有限公司 | Neural network structure searching method and related equipment |
| CN112350995A (en) * | 2020-09-30 | 2021-02-09 | 山东众阳健康科技集团有限公司 | Image processing method, device, equipment and storage medium |
| CN113009077A (en) * | 2021-02-18 | 2021-06-22 | 南方电网数字电网研究院有限公司 | Gas detection method, gas detection device, electronic apparatus, and storage medium |
| CN113223101A (en) * | 2021-05-28 | 2021-08-06 | 支付宝(杭州)信息技术有限公司 | Image processing method, device and equipment based on privacy protection |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106709917B (en) | 2020-09-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106709917A (en) | Neural network model training method, device and system | |
| DE102019106666A1 (en) | Dental CAD automation using deep learning | |
| CN103222876B (en) | Medical image-processing apparatus, image diagnosing system, computer system and medical image processing method | |
| WO2018205232A1 (en) | Method for automatically and accurately positioning reference line according to spliced result | |
| CN109146899A (en) | CT image jeopardizes organ segmentation method and device | |
| Vučinić et al. | Automatic landmarking of cephalograms using active appearance models | |
| CN107665492A (en) | Colon and rectum panorama numeral pathological image tissue segmentation methods based on depth network | |
| CN107220933A (en) | A kind of reference line determines method and system | |
| CN1985258A (en) | Workflow optimization for high thoughput imaging enviroments | |
| DE102018128806A1 (en) | MEDICAL IMAGE ANALYSIS PROCESS USING MACHINE LEARNING AND SYSTEM THEREFOR | |
| CN103959358A (en) | Medical training method and system thereof | |
| CN104720835A (en) | Display device, image displaying method and computerized tomography apparatus | |
| CN107610675A (en) | A kind of image processing method and device based on dynamic level | |
| CN111383328B (en) | 3D visualization method and system for breast cancer focus | |
| CN109241867A (en) | Using the method and device of intelligent algorithm identification digital cores image | |
| DE102021210435A1 (en) | Process and device for the visualization of three-dimensional objects | |
| CN111882559B (en) | ECG signal acquisition method and device, storage medium and electronic device | |
| WO2019146358A1 (en) | Learning system, method, and program | |
| Daemen et al. | Photographic documentation and severity quantification of pectus excavatum through three-dimensional optical surface imaging | |
| CN109523507B (en) | Method and device for generating lesion image and computer readable storage medium | |
| CN117012345A (en) | Semi-automatic labeling method, system, equipment and medium for medical image | |
| US20180365827A1 (en) | Creation of a decision support material indicating damage to an anatomical joint | |
| CN108665386B (en) | Subjective scoring examination system and method based on mobile platform and image processing | |
| CN104915989A (en) | CT image-based blood vessel three-dimensional segmentation method | |
| CN111523507A (en) | Artificial intelligence wound assessment area measuring and calculating method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |