CN109034045A - A kind of leucocyte automatic identifying method based on convolutional neural networks - Google Patents
A kind of leucocyte automatic identifying method based on convolutional neural networks Download PDFInfo
- Publication number
- CN109034045A CN109034045A CN201810802059.0A CN201810802059A CN109034045A CN 109034045 A CN109034045 A CN 109034045A CN 201810802059 A CN201810802059 A CN 201810802059A CN 109034045 A CN109034045 A CN 109034045A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- convolutional neural
- training
- net
- leucocyte
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/69—Microscopic objects, e.g. biological cells or cellular parts
- G06V20/698—Matching; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of leucocyte automatic identifying method based on convolutional neural networks belongs to and carries out automatic identification to medical microscope image using the machine vision scheme of deep learning.The present invention is first manually marked cell data set, make the data set of a standard, the method for reusing transfer learning, the model of VGG-Net and parameter are moved in WBC-Net convolutional neural networks, the identification function to leucocyte is completed as characteristic parameter training integrated classifier by extracting the wherein best characteristic layer of effect.
Description
Technical field
It the present invention relates to the use of computer vision scheme and automatic recognition classification field carried out to medical microscope image, especially
It is a kind of leucocyte automatic identifying method based on convolutional neural networks.
Background technique
With the rapid development of computer vision and artificial intelligence, image processing techniques has in terms of the medical treatment such as medical diagnosis on disease
Across the progress of property, traditional artificial treatment is all much surmounted in accuracy and processing speed.The detection pair of leucocyte
Great meaning is suffered from the diagnostic analysis of many diseases, wherein the detection of peripheral white blood cells can help virologist
Diagnosis such as leukaemia and other hematologic diseases;By the information of the white blood cell detection in assessment bone marrow smear, can be used to examine
The purpose of subsequent nursing after disconnected lymthoma, myeloma, bone marrow proliferative tumour and anaemia and chemotherapy.Acute lymphoblastic is white
Blood disease is a kind of serious hematologic disease, and diagnosis is extremely difficult, and it is deforming white that main inspection is that research is led due to cancer
Cell.Therefore, the further investigation of image processing techniques, it will help push the development of medical technology.
In general, the automatic measurement technique of leucocyte include Image Acquisition, image preprocessing and white blood cell detection, it is white thin
Five segmentation of born of the same parents' image, feature extraction and classifier design aspects.Many researchs are dedicated to the segmentation of Leukocyte Image, such as
Cluster, threshold method, edge detection, region increase, watershed, the methods of color and support vector machines, however due between cell
It is in contact with each other, otherness is small between background and cell, and leucocyte segmentation precision is caused to be difficult to improve.In feature extraction phases, lead to
The features such as extraction circularity, nucleocytoplasmic ratio, color and form, geometry, textural characteristics and gray level co-occurrence matrixes (GLCM) are crossed, to certain
Specific data can obtain huge success, but the adaptability of these low-level features feature new to new data is low, because greatly
Most manual features cannot be simply applied to New Terms.The thought of deep learning is to seek deeper feature, is passed through
Some nonlinear models are transformed into initial data higher level, more abstract expression.Deep learning compares in image procossing
Relatively successfully application is architecture using convolutional neural networks, and CNN can more preferably and fast and accurately understand analysis picture field
The regional aim formed in scape largely reduces the parameter of network training, is simplified by the shared setting with pond layer of weight
Network model, while improving trained efficiency.
Since convolutional neural networks all achieve significant raising in terms of accuracy rate and training effectiveness, much ground so having
The person of studying carefully has done many work to the recognition detection of the leucocyte based on convolutional neural networks, proposes the method for many practicals
Measure specifically includes that
(1) (Chinese patent discloses patent " a kind of the leukorrhea based on convolutional neural networks in leucocyte automatic identifying method "
Number 106897682 A of CN) a kind of method of leucocyte automatic identification in the leukorrhea based on convolutional neural networks is proposed, first
By artificial treatment, the segmented image of leucocyte is obtained, processing is zoomed in and out to segmented image with arest neighbors interpolation algorithm, is led to
It crosses nine layers of convolutional neural networks segmented image is trained and is tested, when meeting square error cost function, then completes to train,
The identification framework of leucocyte determines that, can divide picture by input, carry out the identification of leucocyte.
Disadvantage: the processing of picture is divided there is still a need for artificial, the accuracy of identification of leucocyte may be due to the mistake of segmentation
Difference and make a big difference;Lack the control to network over-fitting in the setting of convolutional neural networks, it may be due to data characteristics
Correlation cause the over-fitting of network.
(2) patent " a kind of leucocyte classification method based on nu- support vector machines " (China Patent Publication No. CN
107730499 A) a kind of leucocyte classification method based on nu- support vector machines is used, first to Color Blood microscope figure
As progress median filtering, image is then mapped to the color space HLS, nu- support vector machines is reused and tone images is divided,
It is detected from coarse segmentation image using Fuzzy Cellular Neural Networks (Fuzzy Cellular Neural Network --- FCNN)
Leucocyte area image out is split by clustering methodology threshold value in conjunction with Threshold segmentation and binary morphology method,
Nucleus topography, cytoplasm topography and background image are obtained, from nucleus topography and cytoplasm topography
It is middle to extract 47 representative features, using these feature vectors, completed using nu- support vector machines to leucocyte
Identification and classification.
Disadvantage: converted color space for cell image, may lose some significant color characteristics;Cell segmentation
There is error, has a great impact to feature extraction later;Due to segmentation extract be nucleus and cytoplasm local feature,
Global characteristics are not accounted for, have limitation when identification and classification, it is weaker to the adaptability of new leucocyte picture.
(3) patent " a kind of five classification method of leucocyte based on deep learning " (China Patent Publication No. CN
106248559 A) using a kind of method that the leucocyte five based on deep learning is classified, first by turning to the space RGB
It changes, obtained coarse segmentation image carries out morphological change and obtains complete nucleus figure, and then goes out leucocyte by apoptotic nueleolus
Image.Its textural characteristics (symbiosis LBP histogram feature) is extracted to the leucocyte of detection, separate accordingly basophilic granulocyte,
Eosinophil and other three classes cells: neutrophil leucocyte, lymphocyte, monocyte.It is automatic using convolutional neural networks
The feature for extracting other three classes cells carries out three classification by random forest.
Disadvantage: eosinophil and other three classes can not accurately only be distinguished by textural characteristics and SVM classifier
Cell, and extract the feature after being characterized in classification error using neural network, i.e., using part eosinophil as it
Its three classes cell has carried out the extraction of feature together, so that being compared with accuracy rate of the random forest to other three classes cell classifications
It is low.
Summary of the invention
The technical problem to be solved by the present invention is in view of the shortcomings of the prior art, provide a kind of based on convolutional neural networks
Leucocyte automatic identifying method,
In order to solve the above technical problems, the technical scheme adopted by the invention is that: it is a kind of based on the white of convolutional neural networks
New Approach for Cell Automatic, this method comprises the following steps:
Step 1: the processing and preparation of data set are used for pre-training network using ImageNet data set, to including white
The microscope photograph of cell carries out individual cells extraction and does classification annotation to it, obtains the standard data set with label and is used for
Later to the training and test of fine tuning convolutional neural networks;
1) multiple groups are extracted at random with the sample block for the NxN size that cell core is rough center, avoids and is determining cell
Divide error caused by when error when core position leads to extract entire cell, meanwhile, the multiple groups sample block of extraction can have
Effect realizes the enhancing to data set;
2) class label is marked to each sample block in step 1-1 by experienced professional, accurately separated different
Normal leucocyte and normal white cell;
Step 2: the leucocyte data set that step 1 obtains randomly is divided into training set and test set in appropriate proportion,
Training set is used for the fine tuning training process to convolutional neural networks, and test set is used to examine the efficiency and parameters weighting of entire algorithm
Update;
Step 3: image enhancement operation being carried out to the training set that step 2 obtains, specifically, using picture about Vertical Square
To mirror surface random reflected operation, and in [- 30,30] pixel coverage randomly left and right translation and upper and lower translation operation (survey
Examination collection does not do image enhancement operation);
Step 4: the structure setting of convolutional neural networks carries out pre-training to convolutional neural networks first, uses
ImageNet data set trains VGG-Net convolutional neural networks, the method for then using transfer learning, by the part after pre-training
Model and weight parameter are transferred on the WBC-Net for needing to finely tune, and are carried out using the enhancing training dataset obtained in step 3 micro-
Adjust operation;
1) the pre-training VGG-Net convolutional neural networks in step 4 are 16 layers, are input layer I1, convolutional layer C1, pond respectively
Change layer P1, convolutional layer C2, pond layer P2, convolutional layer C3, pond layer P3, convolutional layer C4, pond layer P4, convolutional layer C5, pond layer
P5, full articulamentum FC1, full articulamentum FC2, full articulamentum FC3, Softmax layers of S1, output layer O1;
2) the fine tuning training WBC-Net convolutional neural networks in step 4 are designed as in 14 layers, 1-13 layers and step 4-1
The 1-13 layer of pre-training convolutional neural networks VGG-Net be it is identical, provided to carry out shift learning and weight sharing operation
Network structure basis, 14 layers are full connection features extract layer FCL, and 15 layers are Softmax layers of S2, and 16 layers are classification output layer
CO1;
Step 5: the structural parameters of the 1-13 layer of the parameter setting of convolutional neural networks structure, VGG-Net and WBC-Net are set
Set identical, the setting of convolutional layer is that the size of convolution kernel is 3x3 respectively, and padding is set as 1, step-length 1;Convolution each time
An activation primitive processing is carried out after operation, is selected ReLU activation primitive here, gradient can not only be avoided to disappear, made simultaneously
The problem of network has sparsity, reduces the correlation of parameter, reduces over-fitting to a certain extent;Pond layer is using 2x2's
Maximum pond layer, step-length are set as 2;ReLU activation processing is carried out respectively after full articulamentum FC1, FC2, is arranged simultaneously
Dropout rate is 0.5, prevents network over-fitting, enhances the training effectiveness of network;The output of the full articulamentum FC3 of VGG-Net is
The feature output of 1000 dimensions, the output of the full connection features extract layer FCL of WBC-Net convolutional neural networks are set as 2;
Step 6: by the parameter of the convolutional neural networks structure of the structure setting and step 5 of the convolutional neural networks of step 4
It is set using ImageNet data set to be trained VGG-Net, after the completion of training, extracts 1-13 layers of network structure model
Shared with the weight of training parameter transfer learning for after, so far, the entire pre-training stage terminates;
Step 7: the trim process of convolutional neural networks, the pre-training network parameter that step 6 obtains is as fine tuning convolution mind
Initial weight through network, the cell enhancing training set obtained using step 3 are finely adjusted WBC-Net convolutional neural networks;
The cell picture that size is 224x224 is inputted, obtains the feature of 64 112x112 by first group of convolution pondization operation
Figure obtains the characteristic pattern of 128 56x56 by second group of convolution pondization operation, by third group, the 4th group, the 5th group of convolution
The characteristic pattern of 512 7x7 is finally obtained after pondization operation, and feature then is carried out to 512 characteristic patterns by three layers of full articulamentum and is melted
It closes, the probability of each classification is calculated by Softmax layers, finally export to obtain the classification results that two rows one arrange by output layer,
Complete calculating process from input layer to output layer is once to propagate forward;
Step 8: during finely tuning convolutional neural networks, using the method for supervised learning, in one Jing Guo step 7
After secondary propagation forward, the error in classification once propagated forward is calculated using cross entropy loss function, then selects random gradual change
Reduce loss function value to the Policy iteration of gradient decline, update the parameter value of every layer network, in order to accelerate the convergence speed of network
Degree, setting momentum are 0.9, and initial learning rate is 1.0e-4, and MiniBatchSize is set as 10, by weight parameter
It is updated to primary rear feedback propagation;
Step 9: WBC-Net convolutional neural networks are finely tuned using leucocyte data set, by the propagation forward of multiple step 7
Network architecture parameters are updated with the rear feedback propagation of step 8, when reaching the epoch of setting, convolutional neural networks convergence terminates
Training, so far, the trained completion of WBC-Net model;
Step 10: FCL layers of power is extracted from the trained WBC-Net convolutional neural networks model in step 9
Weight parameter is inputted as the feature of classifier, the stable integrated study classification of one group of accuracy height, strong robustness, performance of training
Device does final classification to leucocyte;
Step 11: the setting of integrated study classifier uses N number of decision stub classifier to be integrated as base classifier
The initial weight distribution of training, training data is equally distributed;
1) by training data training base classifier G (x) of initial weight, it is by what error in classification rate determined base classifier
Number;
2) one base classifier of repetitive exercise, the weight that a training dataset is updated after the completion of training are distributed every time, and
With the next base classifier of updated training set training;
3) N number of base classifier linear adder after training is combined into final leukocyte differential count device, in the training of classifier
Verifying assessment is carried out to it using five times of cross-validation methods in the process, so far, the project training of classifier has been completed;
Step 12: test of heuristics stage, the test set that step 3 is obtained (including normal white cell and abnormal white cell) are defeated
Enter trained WBC-Net convolutional neural networks in step 9, the eigenmatrix of extraction is input to rapid 11 as feature and is trained
Integrated classifier classify;
Step 13: the performance of verification algorithm, the predicted value obtained by step 12 and true value compare, and square is obscured in generation
Battle array, averages to confusion matrix, and the accuracy rate for obtaining leukocyte differential count detection algorithm of the invention is 98.72%;
Step 14: WBC-Net convolutional neural networks are obtained by being finely adjusted above to VGG-Net convolutional neural networks,
Extract convolutional neural networks layer in depth characteristic carry out integrated classifier training and test, finally obtained accuracy rate compared with
High, robustness is relatively strong, high stability leucocyte automatic detection algorithm;
Step 15: using the cell picture of no label as input, by the feature extraction of WBC-Net convolutional neural networks,
It recycles integrated classifier to classify, exports normal white cell or abnormal white cell classification.
Compared with prior art, the advantageous effect of present invention is that:
1, the present invention is operated by the enhancing of dialogue cell data set, convolutional neural networks is extracted richer
Characteristic parameter;
2, by the present invention in that having used the thinking of transfer learning, the ginseng of convolutional neural networks is updated with the method for fine tuning
Number, avoids the quantity of data set drawback less than normal, so that small data set can also train to obtain good classifying quality;
3, original Softmax classification layer in VGG-Net is substituted using integrated classifier in the present invention, makes classification accuracy
It is greatly improved.
Detailed description of the invention
Fig. 1 is the method for the present invention flow chart;
Fig. 2 is overall network structural block diagram of the present invention.
Specific embodiment
As shown in Figure 1, general steps of the invention are as follows:
Step 1: the processing and preparation of data set are used for pre-training network using ImageNet data set, to including white
The microscope photograph of cell carries out individual cells extraction and does classification annotation to it, obtains the standard data set with label and is used for
Later to the training and test of fine tuning convolutional neural networks;
1) multiple groups are extracted at random with the sample block for the 257x257 size that cell core is rough center, avoided in determination
Divide error caused by when error when cell core position leads to extract entire cell, meanwhile, the multiple groups sample block of extraction can
Effectively to realize the enhancing to data set;
2) class label is marked to each sample block in step 1-1 by experienced professional, accurately separated different
Normal leucocyte and normal white cell;
Step 2: the leucocyte data set that step 1 obtains randomly being divided into training set and test set in the ratio of 7:3, is instructed
Practice collection for the fine tuning training process to convolutional neural networks, test set is used to examine the efficiency and parameters weighting of entire algorithm
It updates;
Step 3: image enhancement operation being carried out to the training set that step 2 obtains, specifically, using picture about Vertical Square
To mirror surface random reflected operation, and in [- 30,30] pixel coverage uniformly left and right translation and upper and lower translation operation (survey
Examination collection does not do image enhancement operation);
Step 4: the structure setting of convolutional neural networks makes as shown in Fig. 2, first to convolutional neural networks progress pre-training
With ImageNet data set training VGG-Net convolutional neural networks, the method for then using transfer learning, by the portion after pre-training
Sub-model and weight parameter are transferred on the WBC-Net for needing to finely tune, and are carried out using the enhancing training dataset obtained in step 3
Fine tuning operation;
1) the pre-training VGG-Net convolutional neural networks in step 4 are 16 layers, are input layer I1, convolutional layer C1, pond respectively
Change layer P1, convolutional layer C2, pond layer P2, convolutional layer C3, pond layer P3, convolutional layer C4, pond layer P4, convolutional layer C5, pond layer
P5, full articulamentum FC1, full articulamentum FC2, full articulamentum FC3, Softmax layers of S1, output layer O1;
2) the fine tuning training WBC-Net convolutional neural networks in step 4 are designed as in 14 layers, 1-13 layers and step 4-1
The 1-13 layer of pre-training convolutional neural networks VGG-Net be it is identical, provided to carry out shift learning and weight sharing operation
Network structure basis, 14 layers are full connection features extract layer FCL, and 15 layers are Softmax layers of S2, and 16 layers are classification output layer
CO1;
Step 5: the structural parameters of the 1-13 layer of the parameter setting of convolutional neural networks structure, VGG-Net and WBC-Net are set
Set identical, the setting of convolutional layer is that the size of convolution kernel is 3x3 respectively, and padding is set as 1, step-length 1;Convolution each time
An activation primitive processing is carried out after operation, selects ReLU activation primitive here, the formula of activation primitive ReLU is
F (x)=max (0, x) (1)
From formula as can be seen that activation operation can not only avoid gradient from disappearing, while making network that there is sparsity, reduces
The correlation of parameter, to a certain extent the problem of reduction over-fitting;Pond layer uses the maximum pond layer of 2x2, step-length setting
It is 2;ReLU activation processing is carried out respectively after full articulamentum FC1, FC2, while it is 0.5 that dropout rate, which is arranged, prevents net
Network over-fitting enhances the training effectiveness of network;The output of the full articulamentum FC3 of VGG-Net is the feature output of 1000 dimensions, WBC-
The output of the full connection features extract layer FCL of Net convolutional neural networks is set as 2;
Step 6: by the parameter of the convolutional neural networks structure of the structure setting and step 5 of the convolutional neural networks of step 4
It is set using ImageNet data set to be trained VGG-Net, after the completion of training, extracts 1-13 layers of network structure model
Shared with the weight of training parameter transfer learning for after, so far, the entire pre-training stage terminates;
Step 7: the trim process of convolutional neural networks, the pre-training network parameter that step 6 obtains is as fine tuning convolution mind
Initial weight through network, the cell enhancing training set obtained using step 3 are finely adjusted WBC-Net convolutional neural networks;
The cell picture that size is 224x224 is inputted, obtains the feature of 64 112x112 by first group of convolution pondization operation
Figure obtains the characteristic pattern of 128 56x56 by second group of convolution pondization operation, by third group, the 4th group, the 5th group of convolution
The characteristic pattern of 512 7x7 is finally obtained after pondization operation, and feature then is carried out to 512 characteristic patterns by three layers of full articulamentum and is melted
It closes, the probability of each classification is calculated by Softmax layers, it is assumed that SjIndicate the probability of j-th of classification, ajIndicate jth class to
Magnitude (present invention in j=1,2), T be the classification number classified, and probability calculation formula is
Finally export to obtain the classification results that two rows one arrange by output layer, the complete calculating process from input layer to output layer
Once to propagate forward;
Step 8: during finely tuning convolutional neural networks, using the method for supervised learning, in one Jing Guo step 7
After secondary propagation forward, the error in classification once propagated forward, y are calculated using cross entropy loss functioniIndicate true classification
As a result, SjFor in step 7 Softmax layers calculating every one kind probability, then cross entropy loss function formula be
C=- ∑iyilnSi (3)
Then reduce loss function value with selecting the Policy iteration of random depth-graded decline, update the parameter of every layer network
Value, in order to accelerate the convergence rate of network, setting momentum is 0.9, and initial learning rate is 1.0e-4, MiniBatchSize setting
Be 10, by weight parameter be updated to it is primary after feedback propagation;
Step 9: WBC-Net convolutional neural networks are finely tuned using leucocyte data set, by the propagation forward of multiple step 7
Network architecture parameters are updated with the rear feedback propagation of step 8, when reaching the epoch of setting, convolutional neural networks convergence terminates
Training, so far, the trained completion of WBC-Net model;
Step 10: FCL layers of power is extracted from the trained WBC-Net convolutional neural networks model in step 9
Weight parameter is inputted as the feature of classifier, the stable integrated study classification of one group of accuracy height, strong robustness, performance of training
Device does final classification to leucocyte;
Step 11: the setting of integrated study classifier uses eight decision stub classifiers to be collected as base classifier
At training, it is assumed that training dataset
T={ (x1, y1), (x2, y2) ..., (xN, yN)},yi∈ { -1 ,+1 },
The initial weight distribution of training data is equally distributed;
1) error in classification rate of the base classifier G (x) on training dataset determines the coefficient of base classifier, error in classification rate
For
The coefficient of base classifier is
2) one base classifier of repetitive exercise, the weight that a training dataset is updated after the completion of training are distributed every time, and
With the next base classifier of updated training set training;
3) eight base classifier linear adders are combined into final leukocyte differential count device, make in the training process of classifier
Verifying assessment is carried out to it with five times of cross-validation methods, so far, the project training of classifier has been completed;
Step 12: test of heuristics stage, the test set that step 3 is obtained (including normal white cell and abnormal white cell) are defeated
Enter trained WBC-Net convolutional neural networks in step 9, is input to step 11 training for the eigenmatrix of extraction as feature
Good integrated classifier is classified;
Step 13: the performance of verification algorithm, the predicted value obtained by step 12 and true value compare, and square is obscured in generation
Battle array, averages to confusion matrix, and the accuracy rate for obtaining leukocyte differential count detection algorithm of the invention is 98.72%;
Step 14: WBC-Net convolutional neural networks are obtained by being finely adjusted above to VGG-Net convolutional neural networks,
Extract convolutional neural networks layer in depth characteristic carry out integrated classifier training and test, finally obtained accuracy rate compared with
High, robustness is relatively strong, high stability leucocyte automatic detection algorithm;
Step 15: using the cell picture of no label as input, by the feature extraction of WBC-Net convolutional neural networks,
It recycles integrated classifier to classify, exports normal white cell or abnormal white cell classification.
Claims (5)
1. a kind of leucocyte automatic identifying method based on convolutional neural networks, which comprises the following steps:
1) image recognition database is used, i.e. ImageNet database is used for pre-training network, to including the micro- of leucocyte
Mirror picture carries out individual cells extraction and does classification annotation to it, obtains the standard data set for having label;
2) standard data set is randomly divided into training set and test set in appropriate proportion;
3) image enhancement operation is carried out to the training set, obtains enhancing training dataset;
4) convolutional neural networks are set with the following method: using ImageNet database training classical neural network model VGG-
Net convolutional neural networks, then use transfer learning method, by after pre-training department pattern and weight parameter be transferred to need
The leukocyte differential count convolutional neural networks to be finely tuned that is, on WBC-Net convolutional neural networks, use the enhancing training dataset
It is finely adjusted operation;The parameter of convolutional neural networks structure: VGG-Net convolutional neural networks and WBC- is set as follows
The structural parameters setting of the 1-13 layer of Net convolutional neural networks is identical, and the setting of convolutional layer is that the size of convolution kernel is respectively
3x3, padding are set as 1, step-length 1;An activation primitive processing is carried out after convolution operation each time;Pond layer uses
The maximum pond layer of 2x2, step-length are set as 2;It is once corrected at linear unit activating respectively after full articulamentum FC1, FC2
Reason, while the method that partial nerve member is abandoned in application;The output of the full articulamentum FC3 of VGG-Net is the feature output of 1000 dimensions,
The output of the full connection features extract layer FCL of WBC-Net convolutional neural networks is set as 2;
5) pre-training is carried out to VGG-Net convolutional neural networks using ImageNet data set, after the completion of pre-training, extract 1~
13 layers of network structure model and training parameter;
6) training parameter for obtaining step 5) uses enhancing training number as the initial weight of fine tuning convolutional neural networks
WBC-Net convolutional neural networks are finely adjusted according to collection, this process is once to propagate forward;
7) method for using supervised learning is calculated after the primary propagation forward by step 6) using cross entropy loss function
The error in classification once propagated forward out, the method for then selecting random depth-graded decline iteratively reduce loss function value,
The parameter value of every layer network of WBC-Net convolutional neural networks is updated, this process is primary complete rear feedback propagation;
8) WBC-Net convolutional neural networks are finely tuned using leucocyte data set, by the propagation forward of multiple step 6) and step
7) rear feedback propagation updates network architecture parameters, when reaching the maximum update algebra of setting, convolutional neural networks convergence, and knot
Shu Xunlian obtains trained WBC-Net convolutional neural networks model;
9) feature of FCL layers of the weight parameter as classifier is extracted from trained WBC-Net convolutional neural networks model
Input, one group of integrated study classifier of training, final classification is done to leucocyte;
10) it uses N number of decision stub classifier as base classifier, integration trainingt is carried out to integrated study classifier;
11) test set is inputted into trained WBC-Net convolutional neural networks model, with the trained Ensemble classifier of step 10)
Device classifies to the FCL layer weight parameter of the WBC-Net convolutional neural networks extracted in step 9) as feature.
2. the leucocyte automatic identifying method according to claim 1 based on convolutional neural networks, which is characterized in that step
1) specific implementation process includes: random multiple groups of extracting with the sample block for the NxN size that cell core is rough center;To each
Sample block marks class label.
3. the leucocyte automatic identifying method according to claim 1 based on convolutional neural networks, which is characterized in that step
4) in, the WBC-Net convolutional neural networks are designed as the pre-training convolutional neural networks in 14 layers, 1-13 layers and VGG-Net
1-13 layer be it is identical, 14 layers be full connection features extract layer FCL, 15 layers be Softmax layer S2,16 layers for classify output layer
CO1。
4. the leucocyte automatic identifying method according to claim 1 based on convolutional neural networks, which is characterized in that step
4) in, the activation primitive is ReLU activation primitive.
5. the leucocyte automatic identifying method according to claim 1 based on convolutional neural networks, which is characterized in that step
10) specific implementation process includes:
1) by leucocyte training dataset training base classifier G (x) of initial weight, base classifier is determined by error in classification rate
Coefficient;
2) one base classifier of repetitive exercise, the weight that a leucocyte training dataset is updated after the completion of training are distributed every time,
And next base classifier is trained with updated training set;
3) N number of base classifier linear adder after training is combined into final leukocyte differential count device, i.e., trained Ensemble classifier
Device.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810802059.0A CN109034045A (en) | 2018-07-20 | 2018-07-20 | A kind of leucocyte automatic identifying method based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810802059.0A CN109034045A (en) | 2018-07-20 | 2018-07-20 | A kind of leucocyte automatic identifying method based on convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109034045A true CN109034045A (en) | 2018-12-18 |
Family
ID=64642554
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810802059.0A Pending CN109034045A (en) | 2018-07-20 | 2018-07-20 | A kind of leucocyte automatic identifying method based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109034045A (en) |
Cited By (33)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109740560A (en) * | 2019-01-11 | 2019-05-10 | 济南浪潮高新科技投资发展有限公司 | Human cellular protein automatic identifying method and system based on convolutional neural networks |
CN109740697A (en) * | 2019-03-05 | 2019-05-10 | 重庆大学 | Arena micro-image visible component recognition methods based on deep learning |
CN109948429A (en) * | 2019-01-28 | 2019-06-28 | 上海依智医疗技术有限公司 | Image analysis method, device, electronic equipment and computer-readable medium |
CN110033045A (en) * | 2019-04-17 | 2019-07-19 | 内蒙古工业大学 | A kind of method and apparatus of trained identification image atomization |
CN110059568A (en) * | 2019-03-21 | 2019-07-26 | 中南大学 | Multiclass leucocyte automatic identifying method based on deep layer convolutional neural networks |
CN110111310A (en) * | 2019-04-17 | 2019-08-09 | 广州思德医疗科技有限公司 | A kind of method and device of assessment tag picture |
CN110110807A (en) * | 2019-05-16 | 2019-08-09 | 福州大学 | A kind of leucocyte extraction and classification method based on improvement K-means and convolutional neural networks |
CN110136149A (en) * | 2019-05-21 | 2019-08-16 | 闽江学院 | Leucocyte positioning and dividing method based on deep neural network |
CN110414317A (en) * | 2019-06-12 | 2019-11-05 | 四川大学 | Full-automatic Arneth's count method based on capsule network |
CN110517272A (en) * | 2019-08-29 | 2019-11-29 | 电子科技大学 | Blood cell segmentation method based on deep learning |
CN110530779A (en) * | 2019-08-09 | 2019-12-03 | 上海依智医疗技术有限公司 | A kind of white blood cell detection method and device |
CN110580699A (en) * | 2019-05-15 | 2019-12-17 | 徐州医科大学 | Pathological image cell nucleus detection method based on improved fast RCNN algorithm |
CN110689051A (en) * | 2019-09-06 | 2020-01-14 | 北京市安全生产科学技术研究院 | Intelligent identification method for corrosion mode in gas pipeline based on transfer learning |
CN111062346A (en) * | 2019-12-21 | 2020-04-24 | 电子科技大学 | Automatic leukocyte positioning detection and classification recognition system and method |
CN111062296A (en) * | 2019-12-11 | 2020-04-24 | 武汉兰丁医学高科技有限公司 | Automatic white blood cell identification and classification method based on computer |
CN111079620A (en) * | 2019-12-10 | 2020-04-28 | 北京小蝇科技有限责任公司 | Leukocyte image detection and identification model construction method based on transfer learning and application |
CN111368873A (en) * | 2020-01-15 | 2020-07-03 | 江西理工大学 | Convolutional neural network model based on depth fusion and histological image classification method thereof |
CN111444829A (en) * | 2020-03-25 | 2020-07-24 | 南京航空航天大学 | Decision classification method for fusion reasoning and learning for liquid-based cytology examination |
CN111638465A (en) * | 2020-05-29 | 2020-09-08 | 浙大宁波理工学院 | Lithium battery health state estimation method based on convolutional neural network and transfer learning |
CN111680575A (en) * | 2020-05-19 | 2020-09-18 | 苏州大学 | Human epithelial cell staining and classifying device, equipment and storage medium |
CN111783802A (en) * | 2020-07-21 | 2020-10-16 | 南昌航空大学 | Metal fracture image identification method based on deep learning |
CN111783571A (en) * | 2020-06-17 | 2020-10-16 | 陕西中医药大学 | Cervical cell automatic classification model establishment and cervical cell automatic classification method |
CN111914902A (en) * | 2020-07-08 | 2020-11-10 | 南京航空航天大学 | Traditional Chinese medicine identification and surface defect detection method based on deep neural network |
CN112070772A (en) * | 2020-08-27 | 2020-12-11 | 闽江学院 | Blood leukocyte image segmentation method based on UNet + + and ResNet |
CN112201300A (en) * | 2020-10-23 | 2021-01-08 | 天津大学 | Protein subcellular localization method based on depth image features and threshold learning strategy |
CN112561863A (en) * | 2020-12-03 | 2021-03-26 | 吉林大学 | Medical image multi-classification recognition system based on improved ResNet |
CN112801212A (en) * | 2021-03-02 | 2021-05-14 | 东南大学 | White blood cell classification counting method based on small sample semi-supervised learning |
CN113011306A (en) * | 2021-03-15 | 2021-06-22 | 中南大学 | Method, system and medium for automatic identification of bone marrow cell images in continuous maturation stage |
CN113065430A (en) * | 2021-03-22 | 2021-07-02 | 天津大学 | Leukocyte classification method based on fusion of deep learning features and artificial extraction features |
CN113887503A (en) * | 2021-10-22 | 2022-01-04 | 大连工业大学 | Improved attention convolution neural network-based five-classification method for white blood cells |
CN113936165A (en) * | 2021-09-07 | 2022-01-14 | 上海商涌科技有限公司 | CT image processing method, terminal and computer storage medium |
CN114495097A (en) * | 2022-01-28 | 2022-05-13 | 陆建 | Multi-model-based urine cell identification method and system |
CN113936165B (en) * | 2021-09-07 | 2024-06-07 | 上海商涌科技有限公司 | CT image processing method, terminal and computer storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2402470A (en) * | 2003-04-30 | 2004-12-08 | Image Metrics Plc | Method and apparatus for classifying images |
CN1945326A (en) * | 2006-10-13 | 2007-04-11 | 江西特康科技有限公司 | Five classifying full blood cell analysis method based on vision shape |
US20170193336A1 (en) * | 2014-12-22 | 2017-07-06 | Yahoo! Inc. | Generating preference indices for image content |
CN107871136A (en) * | 2017-03-22 | 2018-04-03 | 中山大学 | The image-recognizing method of convolutional neural networks based on openness random pool |
CN108090510A (en) * | 2017-12-15 | 2018-05-29 | 北京大学 | A kind of integrated learning approach and device based on interval optimization |
-
2018
- 2018-07-20 CN CN201810802059.0A patent/CN109034045A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2402470A (en) * | 2003-04-30 | 2004-12-08 | Image Metrics Plc | Method and apparatus for classifying images |
CN1945326A (en) * | 2006-10-13 | 2007-04-11 | 江西特康科技有限公司 | Five classifying full blood cell analysis method based on vision shape |
US20170193336A1 (en) * | 2014-12-22 | 2017-07-06 | Yahoo! Inc. | Generating preference indices for image content |
CN107871136A (en) * | 2017-03-22 | 2018-04-03 | 中山大学 | The image-recognizing method of convolutional neural networks based on openness random pool |
CN108090510A (en) * | 2017-12-15 | 2018-05-29 | 北京大学 | A kind of integrated learning approach and device based on interval optimization |
Non-Patent Citations (1)
Title |
---|
王晓权 等: "基于迁移学习和SVM的糖网图像分类", 《信息通信》 * |
Cited By (45)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109740560A (en) * | 2019-01-11 | 2019-05-10 | 济南浪潮高新科技投资发展有限公司 | Human cellular protein automatic identifying method and system based on convolutional neural networks |
CN109740560B (en) * | 2019-01-11 | 2023-04-18 | 山东浪潮科学研究院有限公司 | Automatic human body cell protein identification method and system based on convolutional neural network |
CN109948429A (en) * | 2019-01-28 | 2019-06-28 | 上海依智医疗技术有限公司 | Image analysis method, device, electronic equipment and computer-readable medium |
CN109740697A (en) * | 2019-03-05 | 2019-05-10 | 重庆大学 | Arena micro-image visible component recognition methods based on deep learning |
CN109740697B (en) * | 2019-03-05 | 2023-04-14 | 重庆大学 | Urinary sediment microscopic image visible component identification method based on deep learning |
CN110059568A (en) * | 2019-03-21 | 2019-07-26 | 中南大学 | Multiclass leucocyte automatic identifying method based on deep layer convolutional neural networks |
CN110033045A (en) * | 2019-04-17 | 2019-07-19 | 内蒙古工业大学 | A kind of method and apparatus of trained identification image atomization |
CN110111310A (en) * | 2019-04-17 | 2019-08-09 | 广州思德医疗科技有限公司 | A kind of method and device of assessment tag picture |
CN110580699A (en) * | 2019-05-15 | 2019-12-17 | 徐州医科大学 | Pathological image cell nucleus detection method based on improved fast RCNN algorithm |
CN110110807A (en) * | 2019-05-16 | 2019-08-09 | 福州大学 | A kind of leucocyte extraction and classification method based on improvement K-means and convolutional neural networks |
CN110110807B (en) * | 2019-05-16 | 2023-01-03 | 福州大学 | Leukocyte extraction and classification method based on improved K-means and convolutional neural network |
CN110136149A (en) * | 2019-05-21 | 2019-08-16 | 闽江学院 | Leucocyte positioning and dividing method based on deep neural network |
CN110414317A (en) * | 2019-06-12 | 2019-11-05 | 四川大学 | Full-automatic Arneth's count method based on capsule network |
CN110414317B (en) * | 2019-06-12 | 2021-10-08 | 四川大学 | Full-automatic leukocyte classification counting method based on capsule network |
CN110530779A (en) * | 2019-08-09 | 2019-12-03 | 上海依智医疗技术有限公司 | A kind of white blood cell detection method and device |
CN110517272B (en) * | 2019-08-29 | 2022-03-25 | 电子科技大学 | Deep learning-based blood cell segmentation method |
CN110517272A (en) * | 2019-08-29 | 2019-11-29 | 电子科技大学 | Blood cell segmentation method based on deep learning |
CN110689051A (en) * | 2019-09-06 | 2020-01-14 | 北京市安全生产科学技术研究院 | Intelligent identification method for corrosion mode in gas pipeline based on transfer learning |
CN111079620A (en) * | 2019-12-10 | 2020-04-28 | 北京小蝇科技有限责任公司 | Leukocyte image detection and identification model construction method based on transfer learning and application |
CN111079620B (en) * | 2019-12-10 | 2023-10-17 | 北京小蝇科技有限责任公司 | White blood cell image detection and identification model construction method and application based on transfer learning |
CN111062296A (en) * | 2019-12-11 | 2020-04-24 | 武汉兰丁医学高科技有限公司 | Automatic white blood cell identification and classification method based on computer |
CN111062346A (en) * | 2019-12-21 | 2020-04-24 | 电子科技大学 | Automatic leukocyte positioning detection and classification recognition system and method |
CN111368873A (en) * | 2020-01-15 | 2020-07-03 | 江西理工大学 | Convolutional neural network model based on depth fusion and histological image classification method thereof |
CN111444829B (en) * | 2020-03-25 | 2023-08-18 | 南京航空航天大学 | Fusion reasoning and learning decision classification method for liquid-based cytology examination |
CN111444829A (en) * | 2020-03-25 | 2020-07-24 | 南京航空航天大学 | Decision classification method for fusion reasoning and learning for liquid-based cytology examination |
CN111680575A (en) * | 2020-05-19 | 2020-09-18 | 苏州大学 | Human epithelial cell staining and classifying device, equipment and storage medium |
CN111638465A (en) * | 2020-05-29 | 2020-09-08 | 浙大宁波理工学院 | Lithium battery health state estimation method based on convolutional neural network and transfer learning |
CN111638465B (en) * | 2020-05-29 | 2023-02-28 | 浙大宁波理工学院 | Lithium battery health state estimation method based on convolutional neural network and transfer learning |
CN111783571A (en) * | 2020-06-17 | 2020-10-16 | 陕西中医药大学 | Cervical cell automatic classification model establishment and cervical cell automatic classification method |
CN111914902A (en) * | 2020-07-08 | 2020-11-10 | 南京航空航天大学 | Traditional Chinese medicine identification and surface defect detection method based on deep neural network |
CN111914902B (en) * | 2020-07-08 | 2024-03-26 | 南京航空航天大学 | Traditional Chinese medicine identification and surface defect detection method based on deep neural network |
CN111783802A (en) * | 2020-07-21 | 2020-10-16 | 南昌航空大学 | Metal fracture image identification method based on deep learning |
CN112070772B (en) * | 2020-08-27 | 2024-01-12 | 闽江学院 | Blood leukocyte image segmentation method based on UNet++ and ResNet |
CN112070772A (en) * | 2020-08-27 | 2020-12-11 | 闽江学院 | Blood leukocyte image segmentation method based on UNet + + and ResNet |
CN112201300B (en) * | 2020-10-23 | 2022-05-13 | 天津大学 | Protein subcellular localization method based on depth image features and threshold learning strategy |
CN112201300A (en) * | 2020-10-23 | 2021-01-08 | 天津大学 | Protein subcellular localization method based on depth image features and threshold learning strategy |
CN112561863A (en) * | 2020-12-03 | 2021-03-26 | 吉林大学 | Medical image multi-classification recognition system based on improved ResNet |
CN112801212A (en) * | 2021-03-02 | 2021-05-14 | 东南大学 | White blood cell classification counting method based on small sample semi-supervised learning |
CN113011306A (en) * | 2021-03-15 | 2021-06-22 | 中南大学 | Method, system and medium for automatic identification of bone marrow cell images in continuous maturation stage |
CN113065430A (en) * | 2021-03-22 | 2021-07-02 | 天津大学 | Leukocyte classification method based on fusion of deep learning features and artificial extraction features |
CN113936165A (en) * | 2021-09-07 | 2022-01-14 | 上海商涌科技有限公司 | CT image processing method, terminal and computer storage medium |
CN113936165B (en) * | 2021-09-07 | 2024-06-07 | 上海商涌科技有限公司 | CT image processing method, terminal and computer storage medium |
CN113887503A (en) * | 2021-10-22 | 2022-01-04 | 大连工业大学 | Improved attention convolution neural network-based five-classification method for white blood cells |
CN113887503B (en) * | 2021-10-22 | 2022-06-14 | 大连工业大学 | Improved attention convolution neural network-based five-classification method for white blood cells |
CN114495097A (en) * | 2022-01-28 | 2022-05-13 | 陆建 | Multi-model-based urine cell identification method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109034045A (en) | A kind of leucocyte automatic identifying method based on convolutional neural networks | |
Ma et al. | Combining DC-GAN with ResNet for blood cell image classification | |
CN109658411A (en) | A kind of correlation analysis based on CT images feature Yu Patients with Non-small-cell Lung prognosis situation | |
Albayrak et al. | Mitosis detection using convolutional neural network based features | |
Beevi et al. | Automatic mitosis detection in breast histopathology images using convolutional neural network based deep transfer learning | |
Bai et al. | Liver tumor segmentation based on multi-scale candidate generation and fractal residual network | |
Beevi et al. | A multi-classifier system for automatic mitosis detection in breast histopathology images using deep belief networks | |
CN109447998B (en) | Automatic segmentation method based on PCANet deep learning model | |
CN109635846A (en) | A kind of multiclass medical image judgment method and system | |
Pan et al. | Mitosis detection techniques in H&E stained breast cancer pathological images: A comprehensive review | |
CN109902736A (en) | A kind of Lung neoplasm image classification method indicated based on autocoder construction feature | |
CN109977955A (en) | A kind of precancerous lesions of uterine cervix knowledge method for distinguishing based on deep learning | |
CN111767952A (en) | Interpretable classification method for benign and malignant pulmonary nodules | |
CN110264454A (en) | Cervical cancer tissues pathological image diagnostic method based on more hidden layer condition random fields | |
CN115100474B (en) | Thyroid gland puncture image classification method based on topological feature analysis | |
Aziz et al. | Channel boosted convolutional neural network for classification of mitotic nuclei using histopathological images | |
CN112183237A (en) | Automatic white blood cell classification method based on color space adaptive threshold segmentation | |
Nneji et al. | A Super-Resolution Generative Adversarial Network with Siamese CNN Based on Low Quality for Breast Cancer Identification | |
Liu et al. | Deep multiscale multi-instance networks with regional scoring for mammogram classification | |
Poornam et al. | Detection of Brain Tumor in MRI Images using Deep Learning Method | |
Pushpa et al. | Deep learning model for plant species classification using leaf vein features | |
Ramalingam et al. | Analysis of oral squamous cell carcinoma into various stages using pre-trained convolutional neural networks | |
Rathore et al. | CBISC: a novel approach for colon biopsy image segmentation and classification | |
Sun et al. | Detection of breast tumour tissue regions in histopathological images using convolutional neural networks | |
Yancey | Deep Feature Fusion for Mitosis Counting |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181218 |