CN109300530A - The recognition methods of pathological picture and device - Google Patents
The recognition methods of pathological picture and device Download PDFInfo
- Publication number
- CN109300530A CN109300530A CN201810896157.5A CN201810896157A CN109300530A CN 109300530 A CN109300530 A CN 109300530A CN 201810896157 A CN201810896157 A CN 201810896157A CN 109300530 A CN109300530 A CN 109300530A
- Authority
- CN
- China
- Prior art keywords
- neural network
- deep neural
- pathological picture
- picture
- model
- 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
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H30/00—ICT specially adapted for the handling or processing of medical images
- G16H30/20—ICT specially adapted for the handling or processing of medical images for handling medical images, e.g. DICOM, HL7 or PACS
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
Landscapes
- Health & Medical Sciences (AREA)
- Engineering & Computer Science (AREA)
- Medical Informatics (AREA)
- Public Health (AREA)
- Epidemiology (AREA)
- Biomedical Technology (AREA)
- Primary Health Care (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Pathology (AREA)
- Databases & Information Systems (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of recognition methods of pathological picture and devices, wherein this method comprises: obtaining pathological picture to be identified；The deep neural network model that pathological picture to be identified is inputted to multiple and different types that training generates in advance, identifies pathological picture to be identified, the deep neural network model of each type obtains a preliminary recognition result；According to multiple pathological picture samples, training generates the deep neural network model of multiple and different types in advance；The preliminary recognition result that the deep neural network model of multiple and different types obtains is merged, the final recognition result of pathological picture to be identified is obtained.Above-mentioned technical proposal improves the efficiency and accuracy rate of pathological picture identification.
Description
Technical field
The present invention relates to the recognition methods of field of medical technology more particularly to pathological picture and devices.
Background technique
Lymphatic metastasis is the most common branch mode of tumour.The radical excision operation of advanced gastric carcinoma includes thoroughly cutting
Except gastric cancer primary lesion, metastatic lymph node and the tissue invaded, internal organs.The pathological diagnosis of postoperative gastric cancer is the gold mark of diagnosing gastric cancer
Standard provides important evidence with treatment by stages for patient.And the assessment that whether lymph node shifts in pathological diagnosis is even more in diagnosing
Key, need pathologist to each lymph node carry out one by one it is conscientious, examine, whole process not only takes time and effort, according to
Rely experience, accuracy rate is not satisfactory, and there may be different identification conclusions to same pathological picture by the doctor of different experiences
Risk.
Summary of the invention
The embodiment of the present invention provides a kind of recognition methods of pathological picture, to improve the efficiency and standard of pathological picture identification
True rate, this method comprises:
Obtain pathological picture to be identified；
The deep neural network model that pathological picture to be identified is inputted to multiple and different types that training generates in advance, is treated
Identification pathological picture is identified that the deep neural network model of each type obtains a preliminary recognition result；It is the multiple not
According to multiple pathological picture samples, training generates the deep neural network model of same type in advance；
The preliminary recognition result that the deep neural network model of multiple and different types obtains is merged, obtain it is described to
Identify the final recognition result of pathological picture.
The embodiment of the present invention also provides a kind of identification device of pathological picture, to improve pathological picture identification efficiency and
Accuracy rate, the device include:
Acquiring unit, for obtaining pathological picture to be identified；
Recognition unit, for pathological picture to be identified to be inputted to the depth nerve for multiple and different types that training generates in advance
Network model identifies that the deep neural network model of each type obtains a preliminary identification knot to pathological picture to be identified
Fruit；According to multiple pathological picture samples, training generates the multiple different types of deep neural network model in advance；
Integrated unit, the preliminary recognition result obtained for the deep neural network model to multiple and different types melt
It closes, obtains the final recognition result of the pathological picture to be identified.
The embodiment of the present invention also provides a kind of computer equipment, including memory, processor and storage are on a memory simultaneously
The computer program that can be run on a processor, the processor realize above-mentioned pathological picture when executing the computer program
Recognition methods.
The embodiment of the present invention also provides a kind of computer readable storage medium, and the computer-readable recording medium storage has
Execute the computer program of the recognition methods of above-mentioned pathological picture.
Technical solution provided in an embodiment of the present invention first obtains pathological picture to be identified, then pathological picture to be identified is defeated
The deep neural network model for entering multiple and different types that training generates in advance, identifies pathological picture to be identified, each
The deep neural network model of type obtains a preliminary recognition result；Multiple different types of deep neural network model according to
Training generates multiple pathological picture samples in advance；Finally, being obtained to the deep neural network model of multiple and different types preliminary
Recognition result is merged, and the final recognition result of the pathological picture to be identified is obtained, due to trained depth nerve net
Network model has pathological picture automatic identification function, and pathological picture is inputted the deep neural network model, that is, may recognize that disease
There may be the region of malignant change on reason picture, realize the good pernicious classification to pathological picture, whole process is time saving, laborsaving, no
But improve the efficiency of pathological picture identification, and the personal experience independent of doctor, final recognition result be to it is multiple not
The preliminary recognition result that the deep neural network model of same type obtains merges to obtain, and substantially increases the standard of pathological picture identification
True rate.
Technical solution provided in an embodiment of the present invention can be applied not only to the identification of lymphonodi gastrici metastasis of cancer pathological picture,
It can also be applied to the identification of other cancer pathological pictures.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.In the accompanying drawings:
Fig. 1 is the flow diagram of the recognition methods of pathological picture in the embodiment of the present invention；
Fig. 2 is a specific example figure of the recognition methods of pathological picture in the embodiment of the present invention；
Fig. 3 is more GPU parallel training schematic illustrations in the embodiment of the present invention；
Fig. 4 is the schematic diagram of the identification device of pathological picture in the embodiment of the present invention.
Specific embodiment
Understand in order to make the object, technical scheme and advantages of the embodiment of the invention clearer, with reference to the accompanying drawing to this hair
Bright embodiment is described in further details.Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but simultaneously
It is not as a limitation of the invention.
Before introducing the embodiment of the present invention, technical term of the present invention is introduced first.
1, false positive rate: false positive rate, practical is that the number of samples negative, model prediction is the positive exists
Ratio in all negative samples.
2, false negative rate: false negative rate, practical is that the number of samples positive, model prediction is feminine gender exists
Ratio in all positive samples.
3, accuracy rate: Accuracy=(predicting correct sample number)/(total number of samples).
4, training set: the lymphonodi gastrici metastasis of cancer with mark cancer cell region (lesion region) of model training is inputed to
Digital pathological image.
5, test set: the lymphonodi gastrici cancer with mark cancer cell region (lesion region) for not inputing to model training turns
Move digital pathological image.
6, verifying collection: without the lymphonodi gastrici metastasis of cancer digital pathological image of mark cancer cell region (lesion region)
7, transfer learning (Transfer learning):
Trained model parameter is moved to new model, to accelerate new model training.
8, scn file: a kind of medical treatment picture storage format, when reading, need specially treated.
9, Top5 error rate: imagenet image usually has 1000 possible classifications, can be pre- simultaneously to each image
Survey 5 class labels, when wherein have it is any once predict right, as a result all calculate pair, when 5 times are wrong in every respect, just calculate in advance
Sniffing misses, and classification error rate at this time is just top5 error rate.
10, HE is dyed: hematoxylin eosin staining method (hematoxylin-eosin staining), hematoxylin dye liquor are
Alkalinity mainly makes endonuclear chromatin and intracytoplasmic nucleic acid hyacinthine；Yihong is acid dyes, mainly makes cytoplasm
With the ingredient red coloration in extracellular matrix.
11, negative sample: the sample not comprising cancer cell, negative sample can be done by also crying: normal or benign lesion pathology figure
Piece.
12, positive sample: the sample comprising cancer cell can also be called positive sample: malignant change pathological picture.
The goal of the invention of the embodiment of the present invention is: for patients with gastric cancer operation clean lymph node pathological section, according to
The various features of histopathologic slide's superior gluteal lymph node make the diagnosis of cancer-free cell, if there is cancer cell then shows its position.
As China 60 years old and the above population account for the ratio of total population and is continuously increased, according to disease incidence of the cancer in population
Measuring and calculating, the cancered size of population will quickly increase.It is more nervous that this will lead to medical resource.During cancer diagnosis, disease
Reason diagnosis is the goldstandard finally made a definite diagnosis.The diagnosis of traditional lymph node metastasis of cancer, needs Pathologis, under the microscope repeatedly
Lymph node is observed, whether there is or not metastasis of cancer for the number for determining lymph node and each lymph node.It is limited to doctors experience and doctor's fatigue shape
State, it may occur that the mistaken diagnosis of certain probability and fail to pinpoint a disease in diagnosis.Accuracy rate of the present invention has reached 99.80% in patch rank, and false positive rate exists
Patch rank is lower than 0.06%, effectively assists diagnosis, reduces the misdiagnosis rate and rate of missed diagnosis of doctor, and final promoted is suffered from
The medical treatment of person is experienced.
So then, inventor is introduced from discovery technique problem to the process for proposing the embodiment of the present invention.
So far from 1970s, machine learning techniques are constantly fast-developing, improve the production efficiency of the mankind.In machine
Device learns in development process, and hardware performance and valid data amount constrain always the development of machine learning.Before and after 2010, firmly
Part performance is significantly promoted and the accumulation of a large amount of quality datas promotes an important component-depth of machine learning
It practises, above has very big breakthrough in algorithm and application.On image procossing, deep learning model is appointed in classification, detection, segmentation etc.
The progress of great-leap-forward is all achieved in business.In the biggish situation of valid data amount, deep learning model logarithm appropriate is used
According to modeling, effect often due to conventional machines study effect, and its to different data collection have transfer learning ability,
Significantly reduce the cost that Feature Engineering is carried out in conventional machines study.So inventor mainly uses nerve in the present invention
Network model models data.
The diagnosis of medical digital pathological image at present has generallyd use and has cut patch/ train classification models/prediction substantially
Process.In this process, it cuts patch and mainly uses 256*256, tri- kinds of specifications of 512*512,1024*1024.Disaggregated model
It is main to use the preferable model of Top5 accuracy rate effect on Academic Data collection ImageNet, such as Inception Model Series,
ResNet Model Series, VGG Model Series.Result is obtained to new data with single trained model when prediction.Due to
ImageNet data set and digital pathological image have many similarities in terms of feature extraction, so on ImageNet data set
Preferable result can be also obtained on effect good model use to digital pathological image data set.Model is trained/is tested
In the process, existing implementation is constantly adjusted model according to available data and has reached better effect.
However, it is found by the inventors that existing scheme has the following technical problems, and phase is proposed for the technical issues of discovery
The solution answered:
A. it when the model used in determining, is limited and hardware limitation by code realizations, only selects 1 depth model progress
Training and test.It is limited to the ability to express of same depth model, model is unilateral to be optimized a certain performance indicator and make other one
A little performance indicators are relatively low.Since inventor considers this technical problem, the technical solution of proposition is: having selected 6 first
Model is trained, and analyzes its feature, then final result is obtained using the result integrated treatment of 3 models, in patch rank
On, false positive control 0.06% while, guarantee false negative rate 0.3% hereinafter, overall accuracy is 99.8% or more.
B. when carrying out model training, the training speed using single GPU is slower, does not make full use of the parallel meter of more GPU
Calculation advantage.It since inventor considers this technical problem, proposes and is trained using more GPU, in same training data
In the case where amount, reduce the model training time, shorten the time of training stage model debugging, save developer when
Between.
C. model prediction improves the stage, in existing implementation, excavates more information from data with existing to improve mould
Type.If encountering model does not have processed cancer cell classification, model can not be identified, the direction of model iteration is not pasted examines with pathology
It is disconnected practical.Since inventor considers this technical problem, technical solution is proposed: being known by continuous integrative medicine profession
Know, supplementary model does not have processed digital pathological picture, so that model is identified more cancer cell forms, missing inspection is effectively reduced
Rate, the direction of iterative model more meet the identification needs of pathological picture, it may also be said to be pathological diagnosis actual demands.
The scheme that inventor proposes is related to the cancer cell based on deep learning disaggregated model (deep neural network model) and examines
It surveys.It may include: 1. acquisitions positive patch sample (positive sample) from digital pathology scn formatted file according to process this method
With negative patch sample (negative sample)；2. the case where comprehensive analysis feminine gender and positive patch sample, selects 6 kinds of suitable depth
Learning classification model makes full use of the advantage of different models；3. with training set data to 6 kinds of deep neural networks of above-mentioned determination
Model is trained under more GPU environments, and having each model in predicted figure pathological image, patch is good on each lymph node
Pernicious ability；4. carrying out detailed test to each model performance with test set, the performance of each model is tested, including, it is false
Positive rate, false negative rate, accuracy rate；5. merging 3 models according to comprehensive preferably 3 models of the performance of each model and existing
Purpose of the result to improve accuracy rate, reduce false positive rate, false negative rate on verifying collection.6. analyzing final prediction result, fill
Divide the professional knowledge in conjunction with doctor, supplements new training data, targetedly iterative model, further increase recognition accuracy.
It describes in detail below to the identifying schemes of the pathological picture as follows.
Fig. 1 is the flow diagram of the recognition methods of pathological picture in the embodiment of the present invention, as shown in Figure 1, this method packet
Include following steps:
Step 101: obtaining pathological picture to be identified；
Step 102: pathological picture to be identified is inputted to the deep neural network for multiple and different types that training generates in advance
Model identifies that the deep neural network model of each type obtains a preliminary recognition result to pathological picture to be identified；It should
According to multiple pathological picture samples, training generates the deep neural network model of multiple and different types in advance；
Step 103: the preliminary recognition result that the deep neural network model of multiple and different types obtains being merged, is obtained
To the final recognition result of pathological picture to be identified.
Technical solution provided in an embodiment of the present invention first obtains pathological picture to be identified, then pathological picture to be identified is defeated
The deep neural network model for entering multiple and different types that training generates in advance, identifies pathological picture to be identified, each
The deep neural network model of type obtains a preliminary recognition result；Multiple different types of deep neural network model according to
Training generates multiple pathological picture samples in advance；Finally, being obtained to the deep neural network model of multiple and different types preliminary
Recognition result is merged, and the final recognition result of the pathological picture to be identified is obtained, due to trained depth nerve net
Network model has pathological picture automatic identification function, and pathological picture is inputted the deep neural network model, that is, may recognize that disease
There may be the region of malignant change on reason picture, realize the good pernicious classification to pathological picture, whole process is time saving, laborsaving, no
But improve the efficiency of pathological picture identification, and the personal experience independent of doctor, final recognition result be to it is multiple not
The preliminary recognition result that the deep neural network model of same type obtains merges to obtain, and substantially increases the standard of pathological picture identification
True rate.
With reference to the accompanying drawing 2, for the recognition methods of pathological picture in the embodiment of the present invention each step carry out it is detailed
It is described below.
First, introduce the process that training in advance generates the deep neural network model of multiple and different types.
In one embodiment, training generates the multiple different types of deep neural network mould in advance as follows
Type:
Sample data is obtained, the sample data includes positive sample and negative sample, and the positive sample is malignant change pathology
Picture, the negative sample are normal or benign lesion pathological picture, mark lesion region on the malignant change pathological picture；
The sample data is divided into training set, test set and verifying collection；
It is trained using deep neural network model of the training set to multiple and different types in first set；
Using the test set to the deep neural network models of multiple and different types in trained first set into
Row test；
According to test result, filtered out from the deep neural network model of multiple and different types in the first set
Multiple deep neural network models are as second set；
Fusion is carried out to the deep neural network model of multiple and different types in second set using verifying collection to test
Card obtains the deep neural network model for multiple and different types that the preparatory training generates.
When it is implemented, the process for obtaining training data (sample data) is introduced first, lymphonodi gastrici metastasis of cancer number disease
Manage storage and the mark of picture:
According to the standard operation on clinical medicine, first the lymph node tissue of patients with gastric cancer is contaminated with the mode that HE is dyed
Color is made into pathological section.Then digital pathological picture is obtained using 40 times of digital pathology scanner scannings, is stored in magnetic disk media
On, obtain the picture of scn format.Physical size ranges 0.4GB-the 8GB that one scn picture stores on disk, the number of pixel
Magnitude is 10^9-10^10.By having the Pathologis of good professional ability using ImageScope software to digital pathology figure
Piece is labeled, and cancerous area is sketched out to come.The data delineated save as the xml label file of specific format so as to program reading
It takes.
Secondly, to the preprocessing process of sample data, obtaining male/female patch after introducing acquisition sample data
(positive/negative sample):
In one embodiment, after obtaining sample data, further the sample data is carried out as follows pre-
Processing:
For each positive sample, pre-processed as follows:
Hsv color format is converted from RGB color format by normal or benign lesion pathological picture；
Normal or benign lesion pathological picture the saturation degree of hsv color format is adjusted to preset threshold；
Normal or benign lesion pathological picture prospect cell compartment after saturation degree is adjusted to preset threshold extracts more
The block patch picture of a presetted pixel size；
The first ratio for judging whole patch picture shared by the prospect for including in the patch picture of presetted pixel size,
When first ratio is less than the first preset ratio value, the patch picture of the presetted pixel size is deleted；
For each negative sample, pre-processed as follows:
The patch picture of multiple preset step-lengths is extracted in the lesion region marked on malignant change pathological picture；
The second ratio for judging whole patch picture shared by the prospect for including in the patch picture of preset step-length, described
When second ratio is less than the second preset ratio value, the patch picture of the preset step-length is deleted；The patch picture of preset step-length
In include prospect be lesion region.
When it is implemented, above-mentioned first preset ratio value and the second preset ratio value can need to carry out according to real work
Flexible setting, the two can be identical, can not also be identical, such as 0.85 be mentioned below.Above-mentioned presetted pixel size can be
The 224*224 pixel being mentioned below, naturally it is also possible to be 112*112,128*128,256*256,512*512 pixel or close big
Small.Above-mentioned preset step-length can be the 112*112 being mentioned below, naturally it is also possible to be 128*128 or close size.
When it is implemented, the pixel quantity of digital pathological picture is very big, this will lead to, and model is subsequent to be built and training process
In, the problem of low memory.In order to solve this problem, present invention employs the modes of patch grade classification to solve hardware
The limitation of low memory, this module is cut into digital pathological picture the patch of 224*224 pixel as a result,.Specific practice is such as
Under:
It is negative sample for whole picture, picture is converted into hsv color format from RGB color format first, so
In saturation degree H, this layer determines suitable threshold value afterwards, for distinguishing picture cell prospect and not at the background blank of cell
It opens.In prospect cell compartment, according to from left to right, from top to bottom, nonoverlapping patch picture for taking 224*224 pixel guarantees
Each piece of foreground area has corresponding patch.Then judge the ratio of whole patch shared by the prospect for including in patch, such as
The fruit ratio illustrates to include that background is more in patch less than 0.85, deletes this patch.The patch finally stayed is
Treated data, the input as model in next module of this pathological picture.To each negative pathological image sample weight
This multiple operation.
For positive pathological picture sample, xml label file is parsed with program, when reading label, it should be noted that distinguish different
Closed area and describe region coordinate points.At this point, other parts are as background delineating as prospect inside region.
Since the area of all in all positive region is less than the region on negative picture, more positive patch are in order to obtain to put down
Weigh Yin/Yang sample size, in the positive region of positive pathological picture, by the way of overlapping 1/2, takes 224*224 picture
The patch of element uses the step-length of 112*112 from left to right, from top to bottom, successively to extract pa tch in positive region.In sun
Property edges of regions part, judge in patch include prospect ratio, delete this patch if this ratio is less than 0.85.It is right
Each positive pathological image sample repeats this operation, obtains the data needed in next module.
Then, then after introducing and being pre-processed to sample data, training generates the multiple different types of depth mind
Process through network model.
(1) training/test data is divided, determines 6 models to be trained:
In one embodiment, the deep neural network model of multiple and different types in the first set includes:
Inception v3 model, resnet18 model, resnet34 model, resnet50 model, VGG16 model and VGG19 model.
When it is implemented, the present invention wherein 80% will be used as training set, and 20% is used as to obtained positive negative sample patch
Test set, test set are used for continuous correction model during model training, model performance are made to reach desirable level.By to point
The investigation of class model, in conjunction with the characteristics of notebook data, the present invention have selected 6 models (multiple and different types in first set
Deep neural network model) training on identical training set, select model to be respectively as follows: inception v3, resnet18,
Resnet34, resnet50, VGG16 and VGG19.The model chosen is extensive in Academic Data collection and true classification task
Using, however inventor has selected above-mentioned 6 models according to many experiments.Each model has the characteristics that different: inception
V3 is mixed with different network minor structures with the versatility of lift scheme；Resnet18 network is shallower, or number small to data volume
It is preferable according to the little situation effect of difference；Resnet34 balances model descriptive power and training speed；Resnet50 with it is identical
The network structure of the number of plies is compared, and trained speed is relatively fast；VGG16 and VGG19 Model Parameter is more, occupies computing resource
More, ability to express is abundant.Performance of the different models in data set of the present invention is different, passes through the reality to a variety of models
Verifying, the further preferred preferable model of effect promote indicators of overall performance.6 models determined in this module, after being
Model preferably provide prerequisite with result fusion, lay the foundation for subsequent raising recognition accuracy.
When it is implemented, the model quantity of model is selected to can be other numbers, it is also possible to other kinds of model.
(2) model training and test iteration:
In present example, the realization of model refers to correlative theses and existing implementation, and special according to patch data
Point selects initial network parameter: learning rate and its tendency planning, patch size, predicts classification number, and gradient decline optimization is calculated
Method initializes weight.Each model is trained and is tested, according to the test result performance current come analysis model, adjustment
Training parameter is to promote the model properties.By 5-10 training and test iteration, the ability orientation of each model
To ultimate attainment, the model optimum performance under notebook data collection is obtained.
In one embodiment, using the training set to the deep neural network mould of multiple and different types in first set
Type is trained, comprising: parallel training is carried out to the deep neural network model of multiple and different types in first set,
In, each model is trained using 2 graphics processor GPU.
When it is implemented, more GPU computing resources are utilized to accelerate training speed in the embodiment of the present invention, shorten on the whole
Period of model training test iteration.In more GPU realization, the gradient that different GPU are calculated is needed to merge, this
The amalgamation mode used is invented as summation, or is averaged.On notebook data collection, the better effect of summation, so of the invention
More GPU have been selected to calculate the mode of gradient summation.It is largely tested by inventor, proposes scheme: determined to single mould
The number of GPU used in type training can be reduced for 80% training time with 2 GPU for each model, such as use more GPU,
Since more GPU stepped costs are higher, acceleration effect is not obviously improved.In the training process, each model uses 2
GPU carries out acceleration training, and the training simultaneously of 6 models takes full advantage of the computing resource of 12 GPU.
(3) detailed process of parallel training multi-model is described below.
In one embodiment, each model is trained using 2 graphics processor GPU, may include:
Training set data is equally divided into the first training data stream and the second training data stream not overlapped；
The pathological picture that preset data amount is obtained in the first training data stream is input to "current" model, and model is by calculating
First-loss functional value is obtained, loss function seeks partial derivative to each variable, obtains the first gradient value of variable；
The pathological picture that preset data amount is obtained in the second training data stream is input to "current" model, and model is by calculating
The second loss function value is obtained, loss function seeks partial derivative to each variable, obtains the second gradient value of variable；
CPU waits the first GPU (GPU1) and the 2nd GPU (GPU2) to calculate gradient value and completes, and sums to two gradient values, then
Corresponding variable is updated with obtained gradient value, obtains the new value of variable, updated variate-value is passed to the first GPU by CPU
(GPU1) and the 2nd GPU (GPU2) original model in the first GPU (GPU1) and the 2nd GPU (GPU2), is covered (such as in Fig. 3
Shown in VGG16 model) variate-value, until training complete.
When it is implemented, since calculating of each model in training is mutually indepedent, here with VGG16 model at 2
On GPU for the process of training, process is as shown in Figure 3:
Training data is equally divided into first the two data streams not overlapped, respectively (the first training of training data stream 1
Data flow) and training data stream 2 (the second training data stream).The patch of fixed quantity in data flow 1 is taken to be input on GPU1
In model, this quantity is 64 patch in this example.Model loses by loss function value first-loss functional value is calculated
Function seeks partial derivative to each variable to obtain the gradient value 1 (first gradient value) of variable.Similar operations are carried out on GPU2,
Obtain variable gradient value 2 (the second gradient value).Hereafter control has given CPU, and it is complete that CPU waits GPU1 and GPU2 to calculate gradient value
At to the summation of two gradient values, the gradient value that then use obtains updates corresponding variable, obtains the new value of variable.CPU can be incited somebody to action every time
GPU1 is synchronous with the model variable on GPU2: updated variate-value is passed to GPU1 and GPU2 by CPU, covers GPU1 and GPU2
In original VGG16 model variable value so that the model variable value in GPU1 and GPU2 is consistent.Then model is again from number
According to data are read in stream, loss function is calculated, calculates variable gradient value, and so on.The above mechanism ensure that data flow 1 and 2
In data can be finished simultaneously, at this moment with training data expanding data stream again, likewise, data flow 1 and data flow 2 do not weigh mutually
Folded and data volume is identical.
After GPU1 and GPU2 calculating gradient value being waited to be completed due to CPU, next step gradient value summation behaviour can be just
Make, the process of waiting causes waste of time, so training speed cannot be made to promote 100% using 2 GPU.In the present invention
In, enable training speed improve 80% using 2 GPU.
(4) 3 models that the fusion after introducing trained multi-model again below is selected carry out prediction process on verifying collection.
In an example, the deep neural network model of multiple and different types in the second set includes:
Resnet34 model, VGG16 model and VGG19 model
When it is implemented, compared the performance of 6 models, the poor resnet18 model of experiment effect is eliminated,
Resnet50 model and inceptionv3 model, the present invention have selected the lower resnet34 model of false negative rate, false positive rate
Lower VGG16 model and VGG19 model.This 3 model result fusion methods are as follows: all predict patch for yin on 3 models
Property the case where, the final prediction result of this patch be feminine gender；In the case of 3 models all predict that patch is positive, this patch
Final prediction result is the positive；In the case of 3 models are inconsistent to same patch prediction result, this patch is finally predicted
It as a result is feminine gender.
It is predicted using data of the above method to verifying collection, and patch prediction result is corresponded to digital pathology figure
On piece goes out the position of model Predict masculine gender patch in pathological picture subscript, carries out visualization presentation, is the result of next module
Ready for analysis.
Second, it introduces after the deep neural network model for the multiple and different types trained in advance, utilizes the model
The process predicted.
After the step of obtaining pathological picture to be identified, for example above-mentioned pretreated mistake to sample data can also be carried out
Journey further increases the efficiency and accuracy rate of identification after to pathological picture to be identified pretreatment.
Third, what the preliminary recognition result that the deep neural network model of multiple and different types obtains was merged in introduction
Process, the process of the fusion, after referring to above-mentioned trained multi-model, merge 3 selected models verifying collection on carry out it is pre-
Survey process.
When it is implemented, the advantageous effects obtained according to abundant experimental results are as follows: iteration 3-5 times according to the actual situation
Afterwards, the embodiment of the present invention patch rank false positive control 0.06% while, guarantee false negative rate 0.3% hereinafter,
Patch rank overall accuracy is 99.8% or more.
4th, the step of introducing model optimization, can also include to model optimization in the use process of following model
Process, further increase identification pathological picture accuracy rate in one embodiment, the process of the model optimization may include:
When the false positive for judging pathological picture to be identified according to recognition result is higher, the yin of the type pathological picture is supplemented
Property sample is into pathological picture sample database；
When the false negative for judging pathological picture to be identified according to recognition result is higher, the sun of the type pathological picture is supplemented
Property sample is into pathological picture sample database；
According to the pathological picture sample database after supplement, to the depth for multiple and different types that the preparatory training generates
Neural network model optimizes training, obtains the deep neural network model of updated multiple and different types；
The deep neural network model that pathological picture to be identified is inputted to multiple and different types that training generates in advance, is treated
Identification pathological picture is identified, may include:
The deep neural network model that pathological picture to be identified is inputted to updated multiple and different types, to disease to be identified
Reason picture is identified.
When it is implemented, data characteristic has very important influence to modelling effect.It, can for above-mentioned model prediction mistake
It is divided into false positive and false negative.From the angle analysis of data, the general character of error prediction data is found, i.e., mistake occurs in a few classes
On patch.If false positive is higher on certain one kind patch, corresponding this kind patch occur in negative sample compared with
It is few, require supplementation with this kind of pathological pictures with special negative mark；If false negative is higher, illustrate model to this kind
Positive patch does not learn well, requires supplementation with delineating for this kind of positive pathological pictures.To model prediction mistake
Patch is sorted out, and the professional knowledge of binding of pathological section doctor and algorithm engineering teacher are needed, and common analysis determination requires supplementation with assorted
The data of sample.In addition, the more diversified pathological pictures of supplement have modelling effect aobvious due to the form of diverse of cancer cell
It writes and is promoted.
Based on the same inventive concept, a kind of identification device of pathological picture is additionally provided in the embodiment of the present invention, it is such as following
Embodiment described in.Since the principle that the device solves the problems, such as is similar to the recognition methods of pathological picture, the reality of the device
The implementation that may refer to the recognition methods of pathological picture is applied, overlaps will not be repeated.
Fig. 4 is the schematic diagram of the identification device of pathological picture in the embodiment of the present invention, as shown in figure 4, the device can wrap
It includes:
Acquiring unit 02, for obtaining pathological picture to be identified；
Recognition unit 04, for pathological picture to be identified to be inputted to the depth mind for multiple and different types that training generates in advance
Through network model, pathological picture to be identified is identified, the deep neural network model of each type obtains a preliminary identification
As a result；According to multiple pathological picture samples, training generates the multiple different types of deep neural network model in advance；
Integrated unit 06, the preliminary recognition result obtained for the deep neural network model to multiple and different types carry out
Fusion, obtains the final recognition result of the pathological picture to be identified.
In one embodiment, the identification device of above-mentioned pathological picture can also include: training unit, for according to as follows
Training generates the multiple different types of deep neural network model to method in advance:
Sample data is obtained, the sample data includes positive sample and negative sample, and the positive sample is malignant change pathology
Picture, the negative sample are normal or benign lesion pathological picture, mark lesion region on the malignant change pathological picture；
The sample data is divided into training set, test set and verifying collection；
It is trained using deep neural network model of the training set to multiple and different types in first set；
Using the test set to the deep neural network models of multiple and different types in trained first set into
Row test；
According to test result, filtered out from the deep neural network model of multiple and different types in the first set
Multiple deep neural network models are as second set；
Fusion is carried out to the deep neural network model of multiple and different types in second set using verifying collection to test
Card obtains the deep neural network model for multiple and different types that the preparatory training generates.
In one embodiment, the identification device of above-mentioned pathological picture further includes pretreatment unit, the pretreatment unit
For:
For each positive sample, pre-processed as follows:
Hsv color format is converted from RGB color format by normal or benign lesion pathological picture；
Normal or benign lesion pathological picture the saturation degree of hsv color format is adjusted to preset threshold；
Normal or benign lesion pathological picture prospect cell compartment after saturation degree is adjusted to preset threshold extracts more
The block patch picture of a presetted pixel size；
The first ratio for judging whole patch picture shared by the prospect for including in the patch picture of presetted pixel size,
When first ratio is less than the first preset ratio value, the patch picture of the presetted pixel size is deleted；
For each negative sample, pre-processed as follows:
The patch picture of multiple preset step-lengths is extracted in the lesion region marked on malignant change pathological picture；
The second ratio for judging whole patch picture shared by the prospect for including in the patch picture of preset step-length, described
When second ratio is less than the second preset ratio value, the patch picture of the preset step-length is deleted；The patch picture of preset step-length
In include prospect be lesion region.
In one embodiment, the deep neural network model of multiple and different types in the first set includes:
Inception v3 model, resnet18 model, resnet34 model, resnet50 model, VGG16 model and VGG19 model；
The deep neural network model of multiple and different types in the second set includes: resnet34 model, VGG16
Model and VGG19 model.
In one embodiment, using the training set to the deep neural network of multiple and different types in first set
Model is trained, comprising: parallel training is carried out to the deep neural network model of multiple and different types in first set,
In, each model is trained using 2 graphics processor GPU.
In one embodiment, each model is trained using 2 graphics processor GPU, may include:
Training set data is equally divided into the first training data stream and the second training data stream not overlapped；
The pathological picture that preset data amount is obtained in the first training data stream is input to "current" model, and model is by calculating
First-loss functional value is obtained, loss function seeks partial derivative to each variable, obtains the first gradient value of variable；
The pathological picture that preset data amount is obtained in the second training data stream is input to "current" model, and model is by calculating
The second loss function value is obtained, loss function seeks partial derivative to each variable, obtains the second gradient value of variable；
CPU waits the first GPU and the 2nd GPU to calculate gradient value and completes, and sums to two gradient values, then with obtained gradient
Value updates corresponding variable, obtains the new value of variable, updated variate-value is passed to the first GPU and the 2nd GPU, covered by CPU
Original model variable value in lid the first GPU and the 2nd GPU, until training is completed.
In one embodiment, the identification device of above-mentioned pathological picture can also include optimization unit, the optimization unit
For:
When the false positive for judging pathological picture to be identified according to recognition result is higher, the yin of the type pathological picture is supplemented
Property sample is into pathological picture sample database；
When the false negative for judging pathological picture to be identified according to recognition result is higher, the sun of the type pathological picture is supplemented
Property sample is into pathological picture sample database；
According to the pathological picture sample database after supplement, to the depth for multiple and different types that the preparatory training generates
Neural network model optimizes training, obtains the deep neural network model of updated multiple and different types；
The deep neural network model that pathological picture to be identified is inputted to multiple and different types that training generates in advance, is treated
Identification pathological picture is identified, comprising:
The deep neural network model that pathological picture to be identified is inputted to updated multiple and different types, to disease to be identified
Reason picture is identified.
The embodiment of the present invention also provides a kind of computer equipment, including memory, processor and storage are on a memory simultaneously
The computer program that can be run on a processor, the processor realize above-mentioned pathological picture when executing the computer program
Recognition methods.
The embodiment of the present invention also provides a kind of computer readable storage medium, and the computer-readable recording medium storage has
Execute the computer program of the recognition methods of above-mentioned pathological picture.
In conclusion nowadays, artificial intelligence technology is risen again, along with big data, new algorithm, the hair of cloud computing
Exhibition, training deep neural network model have become possibility, and artificial intelligence will bring far-reaching influence, people to various industries
Work intelligence+medical treatment is certainly also wherein.Advantage of the artificial intelligence on picture classification is utilized in the embodiment of the present invention, in conjunction with tradition
Medical treatment allows to make pathological picture correct classification.
Technical solution provided in an embodiment of the present invention can be applied not only to the identification of lymphonodi gastrici metastasis of cancer pathological picture,
It can also be applied to the identification of other cancer pathological pictures.
The advantageous effects that technical solution provided in an embodiment of the present invention reaches are as follows:
1. the training stage chooses 6 models, and preferably goes out 3 models, its result is merged.6 models are chosen to be trained,
The characteristics of fully considering each model preferably goes out 3 models according to false positive rate and false negative rate.It is pre- to the positive of 3 models
Result is surveyed to seek common ground to obtain final result.The fusion of multi-model has effectively played the advantage of each model, to promote totality
Performance.
2. in training, using the implementation of more GPU acceleration.With more GPU when parallel, gradient data is distributed to
Different GPU is calculated, and then carries out the synchronization of more GPU results, synchronous method is by the way of adduction in the present invention.Using more
GPU is calculated, and improves training speed, thus the whole time for reducing iterative model.
3. the analysis model result stage focuses on to combine the professional knowledge of doctor, fitting is practical.In practical applications, not only
It only needs the accuracy rate of lift scheme, and needs to supplement new data from the angle of medicine and model is iterated, allow model
Generalization ability is stronger, more practical.Physician specialty knowledge and algorithm knowledge are sufficiently combined in the present invention, the direction of model iteration is more
Closing to reality.
It should be understood by those skilled in the art that, the embodiment of the present invention can provide as method, system or computer program
Product.Therefore, complete hardware embodiment, complete software embodiment or reality combining software and hardware aspects can be used in the present invention
Apply the form of example.Moreover, it wherein includes the computer of computer usable program code that the present invention, which can be used in one or more,
The computer program implemented in usable storage medium (including but not limited to magnetic disk storage, CD-ROM, optical memory etc.) produces
The form of product.
The present invention be referring to according to the method for the embodiment of the present invention, the process of equipment (system) and computer program product
Figure and/or block diagram describe.It should be understood that every one stream in flowchart and/or the block diagram can be realized by computer program instructions
The combination of process and/or box in journey and/or box and flowchart and/or the block diagram.It can provide these computer programs
Instruct the processor of general purpose computer, special purpose computer, Embedded Processor or other programmable data processing devices to produce
A raw machine, so that being generated by the instruction that computer or the processor of other programmable data processing devices execute for real
The device for the function of being specified in present one or more flows of the flowchart and/or one or more blocks of the block diagram.
These computer program instructions, which may also be stored in, is able to guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works, so that it includes referring to that instruction stored in the computer readable memory, which generates,
Enable the manufacture of device, the command device realize in one box of one or more flows of the flowchart and/or block diagram or
The function of being specified in multiple boxes.
These computer program instructions also can be loaded onto a computer or other programmable data processing device, so that counting
Series of operation steps are executed on calculation machine or other programmable devices to generate computer implemented processing, thus in computer or
The instruction executed on other programmable devices is provided for realizing in one or more flows of the flowchart and/or block diagram one
The step of function of being specified in a box or multiple boxes.
Particular embodiments described above has carried out further in detail the purpose of the present invention, technical scheme and beneficial effects
Describe in detail it is bright, it should be understood that the above is only a specific embodiment of the present invention, the guarantor being not intended to limit the present invention
Range is protected, all within the spirits and principles of the present invention, any modification, equivalent substitution, improvement and etc. done should be included in this
Within the protection scope of invention.
Claims (10)
1. a kind of recognition methods of pathological picture characterized by comprising
Obtain pathological picture to be identified；
The deep neural network model that pathological picture to be identified is inputted to multiple and different types that training generates in advance, to be identified
Pathological picture is identified that the deep neural network model of each type obtains a preliminary recognition result；The multiple inhomogeneity
According to multiple pathological picture samples, training generates the deep neural network model of type in advance；
The preliminary recognition result that the deep neural network model of multiple and different types obtains is merged, is obtained described to be identified
The final recognition result of pathological picture.
2. the recognition methods of pathological picture as described in claim 1, which is characterized in that training generates in advance as follows
The multiple different types of deep neural network model:
Sample data is obtained, the sample data includes positive sample and negative sample, and the positive sample is malignant change pathological picture,
The negative sample is normal or benign lesion pathological picture, marks lesion region on the malignant change pathological picture；
The sample data is divided into training set, test set and verifying collection；
It is trained using deep neural network model of the training set to multiple and different types in first set；
It is surveyed using deep neural network model of the test set to multiple and different types in trained first set
Examination；
According to test result, filtered out from the deep neural network model of multiple and different types in the first set multiple
Deep neural network model is as second set；
Fusion verifying is carried out to the deep neural network model of multiple and different types in second set using verifying collection, is obtained
The deep neural network model of the multiple and different types generated to the preparatory training.
3. the recognition methods of pathological picture as claimed in claim 2, which is characterized in that after obtaining sample data, further
The sample data is pre-processed as follows:
For each positive sample, pre-processed as follows:
Hsv color format is converted from RGB color format by normal or benign lesion pathological picture；
Normal or benign lesion pathological picture the saturation degree of hsv color format is adjusted to preset threshold；
Normal or benign lesion pathological picture prospect cell compartment after saturation degree is adjusted to preset threshold extracts multiple pre-
If the block patch picture of pixel size；
The first ratio for judging whole patch picture shared by the prospect for including in the patch picture of presetted pixel size, described
When first ratio is less than the first preset ratio value, the patch picture of the presetted pixel size is deleted；
For each negative sample, pre-processed as follows:
The patch picture of multiple preset step-lengths is extracted in the lesion region marked on malignant change pathological picture；
The second ratio for judging whole patch picture shared by the prospect for including in the patch picture of preset step-length, described second
When ratio is less than the second preset ratio value, the patch picture of the preset step-length is deleted；It is wrapped in the patch picture of preset step-length
The prospect contained is lesion region.
4. the recognition methods of pathological picture as claimed in claim 2, which is characterized in that multiple and different in the first set
The deep neural network model of type include: inception v3 model, resnet18 model, resnet34 model,
Resnet50 model, VGG16 model and VGG19 model；
The deep neural network model of multiple and different types in the second set includes: resnet34 model, VGG16 model
With VGG19 model.
5. the recognition methods of pathological picture as claimed in claim 2, which is characterized in that using the training set to first set
In the deep neural network models of multiple and different types be trained, comprising: to multiple and different types in first set
Deep neural network model carries out parallel training, wherein is trained using 2 graphics processor GPU to each model.
6. the recognition methods of pathological picture as claimed in claim 2, which is characterized in that further include:
When the false positive for judging pathological picture to be identified according to recognition result is higher, the negative sample of the type pathological picture is supplemented
This is into pathological picture sample database；
When the false negative for judging pathological picture to be identified according to recognition result is higher, the positive sample of the type pathological picture is supplemented
This is into pathological picture sample database；
According to the pathological picture sample database after supplement, to the depth nerve for multiple and different types that the preparatory training generates
Network model optimizes training, obtains the deep neural network model of updated multiple and different types；
The deep neural network model that pathological picture to be identified is inputted to multiple and different types that training generates in advance, to be identified
Pathological picture is identified, comprising:
The deep neural network model that pathological picture to be identified is inputted to updated multiple and different types, to pathology figure to be identified
Piece is identified.
7. a kind of identification device of pathological picture characterized by comprising
Acquiring unit, for obtaining pathological picture to be identified；
Recognition unit, for pathological picture to be identified to be inputted to the deep neural network for multiple and different types that training generates in advance
Model identifies that the deep neural network model of each type obtains a preliminary recognition result to pathological picture to be identified；Institute
The deep neural network model of multiple and different types is stated according to the training generation in advance of multiple pathological picture samples；
Integrated unit, the preliminary recognition result obtained for the deep neural network model to multiple and different types merge,
Obtain the final recognition result of the pathological picture to be identified.
8. the identification device of pathological picture as claimed in claim 7, which is characterized in that further include: training unit, for according to
Training generates the multiple different types of deep neural network model to following method in advance:
Sample data is obtained, the sample data includes positive sample and negative sample, and the positive sample is malignant change pathological picture,
The negative sample is normal or benign lesion pathological picture, marks lesion region on the malignant change pathological picture；
The sample data is divided into training set, test set and verifying collection；
It is trained using deep neural network model of the training set to multiple and different types in first set；
It is surveyed using deep neural network model of the test set to multiple and different types in trained first set
Examination；
According to test result, filtered out from the deep neural network model of multiple and different types in the first set multiple
Deep neural network model is as second set；
Fusion verifying is carried out to the deep neural network model of multiple and different types in second set using verifying collection, is obtained
The deep neural network model of the multiple and different types generated to the preparatory training.
9. a kind of computer equipment including memory, processor and stores the meter that can be run on a memory and on a processor
Calculation machine program, which is characterized in that the processor realizes any side of claim 1 to 6 when executing the computer program
Method.
10. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has perform claim
It is required that the computer program of 1 to 6 any the method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810896157.5A CN109300530B (en) | 2018-08-08 | 2018-08-08 | Pathological picture identification method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810896157.5A CN109300530B (en) | 2018-08-08 | 2018-08-08 | Pathological picture identification method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109300530A true CN109300530A (en) | 2019-02-01 |
| CN109300530B CN109300530B (en) | 2020-02-21 |
Family
ID=65168188
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810896157.5A Active CN109300530B (en) | 2018-08-08 | 2018-08-08 | Pathological picture identification method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109300530B (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109961423A (en) * | 2019-02-15 | 2019-07-02 | 平安科技（深圳）有限公司 | A kind of pulmonary nodule detection method based on disaggregated model, server and storage medium |
| CN110097564A (en) * | 2019-04-04 | 2019-08-06 | 平安科技（深圳）有限公司 | Image labeling method, device, computer equipment and storage medium based on multi-model fusion |
| CN110335668A (en) * | 2019-05-22 | 2019-10-15 | 台州市中心医院（台州学院附属医院） | Thyroid cancer cell pathological map auxiliary analysis method and system based on deep learning |
| CN110706812A (en) * | 2019-09-29 | 2020-01-17 | 医渡云（北京）技术有限公司 | Medical index time sequence prediction method, device, medium and electronic equipment |
| CN111276254A (en) * | 2020-01-13 | 2020-06-12 | 印迹信息科技（北京）有限公司 | Medical open platform system and diagnosis and treatment data processing method |
| CN111325103A (en) * | 2020-01-21 | 2020-06-23 | 华南师范大学 | Cell labeling system and method |
| CN111710394A (en) * | 2020-06-05 | 2020-09-25 | 沈阳智朗科技有限公司 | Artificial intelligence assisted early gastric cancer screening system |
| CN111815609A (en) * | 2020-07-13 | 2020-10-23 | 北京小白世纪网络科技有限公司 | Pathological image classification method and system based on context awareness and multi-model fusion |
| CN112348059A (en) * | 2020-10-23 | 2021-02-09 | 北京航空航天大学 | Deep learning-based method and system for classifying multiple dyeing pathological images |
| CN112507801A (en) * | 2020-11-14 | 2021-03-16 | 武汉中海庭数据技术有限公司 | Lane road surface digital color recognition method, speed limit information recognition method and system |
| CN112734707A (en) * | 2020-12-31 | 2021-04-30 | 重庆西山科技股份有限公司 | Auxiliary detection method, system and device for 3D endoscope and storage medium |
| CN113222928A (en) * | 2021-05-07 | 2021-08-06 | 北京大学第一医院 | Artificial intelligent urothelial cancer recognition system for urocytology |
| WO2022028127A1 (en) * | 2020-08-06 | 2022-02-10 | 腾讯科技（深圳）有限公司 | Artificial intelligence-based pathological image processing method and apparatus, electronic device, and storage medium |
| CN118378726A (en) * | 2024-06-25 | 2024-07-23 | 之江实验室 | Model training system, method, storage medium and electronic equipment |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1815399A2 (en) * | 2004-11-19 | 2007-08-08 | Koninklijke Philips Electronics N.V. | A stratification method for overcoming unbalanced case numbers in computer-aided lung nodule false positive reduction |
| WO2008035276A2 (en) * | 2006-09-22 | 2008-03-27 | Koninklijke Philips Electronics N.V. | Methods for feature selection using classifier ensemble based genetic algorithms |
| CN205015889U (en) * | 2015-09-23 | 2016-02-03 | 北京科技大学 | Definite system of traditional chinese medical science lingual diagnosis model based on convolution neuroid |
| CN107564580A (en) * | 2017-09-11 | 2018-01-09 | 合肥工业大学 | Gastroscope visual aids processing system and method based on integrated study |
| CN107909566A (en) * | 2017-10-28 | 2018-04-13 | 杭州电子科技大学 | A kind of image-recognizing method of the cutaneum carcinoma melanoma based on deep learning |
| CN108288506A (en) * | 2018-01-23 | 2018-07-17 | 雨声智能科技（上海）有限公司 | A kind of cancer pathology aided diagnosis method based on artificial intelligence technology |
| CN108364293A (en) * | 2018-04-10 | 2018-08-03 | 复旦大学附属肿瘤医院 | A kind of on-line training thyroid tumors Ultrasound Image Recognition Method and its device |
-
2018
- 2018-08-08 CN CN201810896157.5A patent/CN109300530B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1815399A2 (en) * | 2004-11-19 | 2007-08-08 | Koninklijke Philips Electronics N.V. | A stratification method for overcoming unbalanced case numbers in computer-aided lung nodule false positive reduction |
| WO2008035276A2 (en) * | 2006-09-22 | 2008-03-27 | Koninklijke Philips Electronics N.V. | Methods for feature selection using classifier ensemble based genetic algorithms |
| CN205015889U (en) * | 2015-09-23 | 2016-02-03 | 北京科技大学 | Definite system of traditional chinese medical science lingual diagnosis model based on convolution neuroid |
| CN107564580A (en) * | 2017-09-11 | 2018-01-09 | 合肥工业大学 | Gastroscope visual aids processing system and method based on integrated study |
| CN107909566A (en) * | 2017-10-28 | 2018-04-13 | 杭州电子科技大学 | A kind of image-recognizing method of the cutaneum carcinoma melanoma based on deep learning |
| CN108288506A (en) * | 2018-01-23 | 2018-07-17 | 雨声智能科技（上海）有限公司 | A kind of cancer pathology aided diagnosis method based on artificial intelligence technology |
| CN108364293A (en) * | 2018-04-10 | 2018-08-03 | 复旦大学附属肿瘤医院 | A kind of on-line training thyroid tumors Ultrasound Image Recognition Method and its device |
Non-Patent Citations (1)
| Title |
|---|
| 崔泳琳等: "气象与环境因子影响心脑血管疾病急诊数预测模型研究", 《环境与健康杂志》 * |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109961423A (en) * | 2019-02-15 | 2019-07-02 | 平安科技（深圳）有限公司 | A kind of pulmonary nodule detection method based on disaggregated model, server and storage medium |
| CN109961423B (en) * | 2019-02-15 | 2024-05-31 | 平安科技（深圳）有限公司 | Lung nodule detection method based on classification model, server and storage medium |
| CN110097564A (en) * | 2019-04-04 | 2019-08-06 | 平安科技（深圳）有限公司 | Image labeling method, device, computer equipment and storage medium based on multi-model fusion |
| CN110097564B (en) * | 2019-04-04 | 2023-06-16 | 平安科技（深圳）有限公司 | Image labeling method and device based on multi-model fusion, computer equipment and storage medium |
| WO2020199477A1 (en) * | 2019-04-04 | 2020-10-08 | 平安科技（深圳）有限公司 | Image labeling method and apparatus based on multi-model fusion, and computer device and storage medium |
| CN110335668A (en) * | 2019-05-22 | 2019-10-15 | 台州市中心医院（台州学院附属医院） | Thyroid cancer cell pathological map auxiliary analysis method and system based on deep learning |
| CN110706812A (en) * | 2019-09-29 | 2020-01-17 | 医渡云（北京）技术有限公司 | Medical index time sequence prediction method, device, medium and electronic equipment |
| CN111276254A (en) * | 2020-01-13 | 2020-06-12 | 印迹信息科技（北京）有限公司 | Medical open platform system and diagnosis and treatment data processing method |
| CN111325103B (en) * | 2020-01-21 | 2020-11-03 | 华南师范大学 | Cell labeling system and method |
| CN111325103A (en) * | 2020-01-21 | 2020-06-23 | 华南师范大学 | Cell labeling system and method |
| CN111710394A (en) * | 2020-06-05 | 2020-09-25 | 沈阳智朗科技有限公司 | Artificial intelligence assisted early gastric cancer screening system |
| CN111815609A (en) * | 2020-07-13 | 2020-10-23 | 北京小白世纪网络科技有限公司 | Pathological image classification method and system based on context awareness and multi-model fusion |
| CN111815609B (en) * | 2020-07-13 | 2024-03-01 | 北京小白世纪网络科技有限公司 | Pathological image classification method and system based on context awareness and multi-model fusion |
| WO2022028127A1 (en) * | 2020-08-06 | 2022-02-10 | 腾讯科技（深圳）有限公司 | Artificial intelligence-based pathological image processing method and apparatus, electronic device, and storage medium |
| CN112348059A (en) * | 2020-10-23 | 2021-02-09 | 北京航空航天大学 | Deep learning-based method and system for classifying multiple dyeing pathological images |
| CN112507801A (en) * | 2020-11-14 | 2021-03-16 | 武汉中海庭数据技术有限公司 | Lane road surface digital color recognition method, speed limit information recognition method and system |
| CN112734707B (en) * | 2020-12-31 | 2023-03-24 | 重庆西山科技股份有限公司 | Auxiliary detection method, system and device for 3D endoscope and storage medium |
| CN112734707A (en) * | 2020-12-31 | 2021-04-30 | 重庆西山科技股份有限公司 | Auxiliary detection method, system and device for 3D endoscope and storage medium |
| CN113222928A (en) * | 2021-05-07 | 2021-08-06 | 北京大学第一医院 | Artificial intelligent urothelial cancer recognition system for urocytology |
| CN113222928B (en) * | 2021-05-07 | 2023-09-19 | 北京大学第一医院 | Urine cytology artificial intelligence urothelial cancer identification system |
| CN118378726A (en) * | 2024-06-25 | 2024-07-23 | 之江实验室 | Model training system, method, storage medium and electronic equipment |
| CN118378726B (en) * | 2024-06-25 | 2024-09-20 | 之江实验室 | Model training system, method, storage medium and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109300530B (en) | 2020-02-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109300530A (en) | The recognition methods of pathological picture and device | |
| CN108596882B (en) | The recognition methods of pathological picture and device | |
| WO2020182710A1 (en) | Multiple instance learner for prognostic tissue pattern identification | |
| CN108446730A (en) | A kind of CT pulmonary nodule detection methods based on deep learning | |
| CN108257135A (en) | The assistant diagnosis system of medical image features is understood based on deep learning method | |
| CN111047594A (en) | Tumor MRI weak supervised learning analysis modeling method and model thereof | |
| CN109410219A (en) | A kind of image partition method, device and computer readable storage medium based on pyramid fusion study | |
| CN114693933A (en) | Medical image segmentation device based on generation of confrontation network and multi-scale feature fusion | |
| CN110175998A (en) | Breast cancer image-recognizing method, device and medium based on multiple dimensioned deep learning | |
| CN104484886B (en) | A kind of dividing method and device of MR images | |
| Öztürk et al. | Cell‐type based semantic segmentation of histopathological images using deep convolutional neural networks | |
| CN108986132A (en) | A method of certificate photo Trimap figure is generated using full convolutional neural networks | |
| Sikder et al. | Supervised learning-based cancer detection | |
| Yonekura et al. | Improving the generalization of disease stage classification with deep CNN for glioma histopathological images | |
| Almufareh et al. | Automated brain tumor segmentation and classification in MRI using YOLO-based Deep Learning | |
| Saueressig et al. | Exploring graph-based neural networks for automatic brain tumor segmentation | |
| CN108564582A (en) | A kind of MRI brain tumor automatic identifying methods based on deep neural network | |
| Barrera et al. | Automatic normalized digital color staining in the recognition of abnormal blood cells using generative adversarial networks | |
| Renugadevi et al. | Machine Learning Empowered Brain Tumor Segmentation and Grading Model for Lifetime Prediction | |
| Tang et al. | Consistency and adversarial semi-supervised learning for medical image segmentation | |
| Huang et al. | Flip learning: Erase to segment | |
| Sun et al. | Detection of breast tumour tissue regions in histopathological images using convolutional neural networks | |
| CN113689950B (en) | Method, system and storage medium for identifying blood vessel distribution pattern of liver cancer IHC staining pattern | |
| CN113393445B (en) | Breast cancer image determination method and system | |
| Inamdar et al. | A Novel Attention based model for Semantic Segmentation of Prostate Glands using Histopathological Images |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |