CN109376777A - Cervical cancer tissues pathological image analysis method and equipment based on deep learning - Google Patents
Cervical cancer tissues pathological image analysis method and equipment based on deep learning Download PDFInfo
- Publication number
- CN109376777A CN109376777A CN201811212019.7A CN201811212019A CN109376777A CN 109376777 A CN109376777 A CN 109376777A CN 201811212019 A CN201811212019 A CN 201811212019A CN 109376777 A CN109376777 A CN 109376777A
- Authority
- CN
- China
- Prior art keywords
- trained
- full articulamentum
- image
- convolutional neural
- neural networks
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30096—Tumor; Lesion
Abstract
This method comprises: obtaining cervical cancer tissues pathological image, and image tag is arranged for every image in the invention discloses a kind of cervical cancer tissues pathological image analysis method and equipment based on deep learning；Based on being trained respectively to two convolutional neural networks for trained image, trained two convolutional neural networks are obtained；The parameter of fixed trained two convolutional neural networks obtains trained full articulamentum one based on being trained for trained image to full articulamentum one；Image to be tested is input in trained classifier, two convolutional neural networks extract feature vector from image respectively, the feature vector f1 and f2 of output are stitched together and are input to full articulamentum one, feature vector f3 is exported, classification results are determined by the maximum element of numerical value in feature vector f3.The differentiation degree for the original tissue pathological slice displaing micro picture that the present invention can automatically acquire doctor carries out Classification and Identification, and auxiliary doctor diagnoses.
Description
Technical field
The present invention relates to medical field more particularly to a kind of cervical cancer tissues pathological image analyses based on deep learning
Method and apparatus.
Background technique
Cervical carcinoma is common one of gynecologic malignant tumor, for now, cervical cancer tissues pathology micro-image
Computer-aided diagnosis research is focused primarily upon using classical image characteristic extracting method and machine learning classification method to uterine neck
Histopathology picture carries out image segmentation and pathological abnormalities screening, rarely has research for point of cervical cancer tissues pathology picture
Change degree carries out computer-aided diagnosis research.
The prior art according to use computer vision methods to the histopathology image of only single cervical cancer cell into
Then row feature extraction is classified using feature further progress of the conventional machines learning method to extraction.As shown in Figure 1, the party
Method from top to bottom includes five steps altogether:
(1) data prediction: becoming gray level image for color tissue pathological image, and image size is constant, carries out image
Enhancing filters out noise jamming, strengthens image edge information.
(2) Methods of Segmentation On Cell Images: carrying out image segmentation twice respectively, to distinguish cell and nucleus, is converted to binary map
As after, realized using the algorithm based on Threshold segmentation
(3) characteristics extraction: on the basis of carrying out threshold process to image, using 8 connection chain codes to cell and cell
Core carries out morphological feature extraction, and the feature of extraction has perimeter, area, circularity, rectangular degree, nucleocytoplasmic ratio etc., later to characteristic value
It is standardized.
(4) machine learning method (artificial neural network or support vector machines) learning functionality part: in the spy to cell
After value indicative is extracted and standardized, measured Standard Eigenvalue is learnt, training classifier parameters weight, so that instead
Reach required set value to error e.
(5) machine learning method classified part: this part is the most important part of system, and pervious various pieces are all these
Partial preparation makes higher point of final classification accuracy rate having learnt to have obtained after a series of parameter learning
Class weight.Finally can accurately it be classified to the cell image data for test.
However this method can only select that noise is smaller, clarity is high individual cells histopathology picture is handled,
Complete tissue pathology microsection image can not directly be handled, need manually to extract individual cells image；
This method has selected the morphological feature of 27 engineers as the input value of machine learning, and this transportable property of method is poor,
It is easy to generate over-fitting in study；And can only the cell categories of individual cells classify, can not globally diagnose disease
The cancer differentiation degree of people, thus the cancer state of an illness of auxiliary judgment patient.
Summary of the invention
The technical problems to be solved by the present invention are: in view of the problems of the existing technology, the present invention provides one kind and is based on
The cervical cancer tissues pathological image analysis method and equipment of deep learning can possess one the histopathology of many cells
Picture is classified, and the classification to cervical cancer tissues pathological image differentiation degree is passed through, it can be determined that the grade malignancy of cancer,
To help doctor preferably to formulate medical plan, cancer is timely and effectively treated.
A kind of cervical cancer tissues pathological image analysis method based on deep learning provided by the invention, comprising:
Step 1 obtains cervical cancer tissues pathological image, and image tag is arranged for every image；
Step 2 obtains trained classifier, the classification based on being trained for trained image to classifier
Device includes two convolutional neural networks and full articulamentum one, and the input terminal of the full articulamentum one connects described two convolutional Neurals
The output end of network, to the method that classifier is trained include: based on for trained image respectively to two convolutional Neurals
Network is trained, and obtains trained two convolutional neural networks；The parameter of fixed trained two convolutional neural networks,
Based on being trained for trained image to full articulamentum one, trained full articulamentum one is obtained；
Image to be tested is input in trained classifier by step 3, and two convolutional neural networks are respectively from figure
Feature vector is extracted as in, the feature vector f1 and f2 of output are stitched together and are input to full articulamentum one, exports feature
Vector f 3, classification results are determined by the maximum element of numerical value in feature vector f3.
Further, full articulamentum two and full articulamentum are connected separately with behind two convolutional neural networks in classifier
Three, the input terminal of full articulamentum one connects the output end of full articulamentum two and full articulamentum three at this time, is training two in step 2
Also full articulamentum two and full articulamentum three are trained while a convolutional neural networks.
Further, step 1 further include: the image x for obtaining eachiIt is divided into 16 equal-sized subgraph zij, make
With the mirror image edge filling picture z ' that subgraph filling growth is equal with widthij, for each picture z 'ij, carry out picture rotation
0 °, 90 °, 180 ° and 270 ° operation and flip horizontal, flip vertical and channel turning operation, i=1,2 ..., n, j=1,
2..., 16, n is total number of images.
Further, the training method of classifier is specifically included in step 2: be set separately two convolutional neural networks and
The hyper parameter of full articulamentum two and full articulamentum three；The model parameter downloaded from ImageNet is imported to two convolutional Neural nets
In network；It imports and two convolutional neural networks and full articulamentum two and full articulamentum three is trained for the image of training, obtain
To trained two convolutional neural networks and full articulamentum two and full articulamentum three；Reset two convolutional neural networks and
Hyper parameter and the fixation of full articulamentum two and full articulamentum three；It imports and full articulamentum one is trained for the image of training,
Obtain trained full articulamentum one.
Further, further including step 4, Performance Evaluation is carried out to classifier, evaluation index includes accuracy rate accuracy,
Accurate rate precision, recall rate recall and F1 estimate, and the calculation formula of each index is as follows:
Wherein, TP is the quantity for the positive sample that the convolutional neural networks prediction being trained to is positive, and FP is trained to
The quantity for the negative sample that convolutional neural networks prediction is positive, FN are the positive samples that the convolutional neural networks being trained to are predicted to be negative
Quantity, TN is the quantity of negative sample that the convolutional neural networks prediction being trained to is negative.
Further, described two convolutional neural networks are respectively VGG16 and Inception-V3.
A kind of cervical cancer tissues pathological image analytical equipment based on deep learning that another aspect of the present invention provides, packet
It includes:
Image tag is arranged for obtaining cervical cancer tissues pathological image, and for every image in image acquiring device；
Classifier training device, for obtaining trained point based on being trained for trained image to classifier
Class device, the classifier include two convolutional neural networks and full articulamentum one, and the input terminal of the full articulamentum one connects institute
The output end for stating two convolutional neural networks includes: based on for trained image difference to the method that classifier is trained
Two convolutional neural networks are trained, trained two convolutional neural networks are obtained；Fixed trained two convolution
The parameter of neural network obtains trained full articulamentum one based on being trained for trained image to full articulamentum one；
Classification results output device, for image to be tested to be input in trained classifier, two convolution minds
Feature vector is extracted from image respectively through network, the feature vector f1 and f2 of output are stitched together and are input to full connection
Layer one, exports feature vector f3, and classification results are determined by the maximum element of numerical value in feature vector f3.
Further, full articulamentum two and full articulamentum are connected separately with behind two convolutional neural networks in classifier
Three, the input terminal of full articulamentum one connects the output end of full articulamentum two and full articulamentum three at this time, and classifier training device exists
Also full articulamentum two and full articulamentum three are trained while two convolutional neural networks of training.
It further, further include classifier performance assessment device, for carrying out Performance Evaluation to classifier, evaluation index includes
Accuracy rate accuracy, accurate rate precision, recall rate recall and F1 estimate, and the calculation formula of each index is as follows:
Wherein, TP is the quantity for the positive sample that the convolutional neural networks prediction being trained to is positive, and FP is trained to
The quantity for the negative sample that convolutional neural networks prediction is positive, FN are the positive samples that the convolutional neural networks being trained to are predicted to be negative
Quantity, TN is the quantity of negative sample that the convolutional neural networks prediction being trained to is negative.
A kind of computer readable storage medium that another aspect of the present invention provides, is stored thereon with computer program, special
The step of sign is, the computer program realizes method as described above when being executed by processor.
Compared with prior art, present invention enhances the degree of intelligence of cervical cancer tissues pathology picture classification, can be certainly
Classification and Identification dynamicly is carried out to the differentiation degree of the original tissue pathological slice displaing micro picture of doctor's acquisition, auxiliary doctor carries out
Diagnosis.
Detailed description of the invention
Examples of the present invention will be described by way of reference to the accompanying drawings, in which:
Fig. 1 is the method flow diagram classified in the prior art to cervical cancer tissues pathological image；
Fig. 2 is that the image data of the embodiment of the present invention enhances schematic diagram；
Fig. 3 is the cervical cancer tissues pathological image analysis method schematic diagram of the embodiment of the present invention；
Fig. 4 is the scatter plot that the F1- of the classifier of training of the embodiment of the present invention estimates；
Fig. 5 is the cervical cancer tissues pathological image of successful classification of the embodiment of the present invention.
Specific embodiment
All features disclosed in this specification or disclosed all methods or in the process the step of, in addition to mutually exclusive
Feature and/or step other than, can combine in any way.
Any feature disclosed in this specification unless specifically stated can be equivalent or with similar purpose by other
Alternative features are replaced.That is, unless specifically stated, each feature is an example in a series of equivalent or similar characteristics
?.
The concrete scheme of cervical cancer tissues pathological image analysis method provided by the invention based on deep learning is as follows:
One, sample data obtains and enhances
The histopathology micro-image of cervical cancer tissues slice shooting is prepared by pathology department of Chinese Medical Sciences University, record
Cancer pathology type and differentiation degree, tumor size.Using the full-automatic immunohistochemical staining agent of Leica BOND-MAXTM
(Leica company) carries out immunohistochemical staining.AQP-1 monoclonal antibody (abcam company, Shanghai) stoste is diluted 1:300 work
Make liquid, injection dilutes in open reagent bottle.VEGF polyclonal antibody (abcam company, Shanghai) stoste is diluted to 1:50 work
Liquid.20min is repaired using antigen retrieval buffers ER1 later.It dewaxes, expose antigenic determinant, incubation I resists, closes, DAB is aoxidized and shown
Color, haematoxylin are redyed the processes such as dehydration and are automatically finished using computer, and artificial mounting is then carried out.Every slice randomly selects
3 are full of the high power lens cause (× 400) of cervical cancer tissues, carry out image by 3.2 image capture software of NIS-Elements F
Acquisition.
Since the histopathology micro-image data volume of existing cervical carcinoma only has 307 in total, it was easy to produce quasi-
It closes, meanwhile, the histopathology micro-image of cervical carcinoma has rotational invariance, and therefore, rotation and mirror image can be used in we
Method data enhancing is carried out to it, as shown in Figure 2.For each sample image xi, i=1,2 ..., n, n are a samples
Total number of images in this collection X, we are divided into 16 equal-sized subgraph zij, j=1,2... after 16, use mirror
As the edge filling picture z ' that subgraph filling growth is equal with widthij, for each picture z 'ij, we carry out two kinds of data
Enhancing operation, one is to 0 ° of picture rotation, 90 °, 180 °, 270 ° of operations, the second is carrying out flip horizontal to image, vertically turning over
Turn, channel turning operation, Zhang Zitu z ' each in this wayijCan produce 16 enhancing after picture, picture tag is still original sample figure
As xiLabel.Each sample image xi256 pictures are obtained after data enhancing, original data set size is 307, after data enhancing
Data set size is expanded to 78592, is 45824 for trained data set size in a specific embodiment.The present invention is real
Applying in example the results are shown in Table 1 after data set amplification, and AQP, HIF and VEGF represent three kinds of different cervical carcinoma staining pathologic section sides
Formula.In some embodiments, pre- picture can be carried out to histopathology image with K-means dividing method or Mask-RCNN
Plain grade segmentation, to remove the garbage in image.
1 cervical cancer tissues pathology micro-image data set of table
Two, classifier training
Based on being trained for trained image to classifier, trained classifier is obtained, the classifier includes
Two convolutional neural networks and full articulamentum one, the input terminal of full articulamentum one connect the output end of two convolutional neural networks,
Full articulamentum is common deep neural network (DNN).It include: based on for trained to the method that classifier is trained
Image is respectively trained two convolutional neural networks, obtains trained two convolutional neural networks；Fixation is trained
The parameter of two convolutional neural networks is obtained trained complete based on being trained for trained image to full articulamentum one
Articulamentum one.
In an embodiment of the present invention, will by the enhanced image segmentation of data at training set, verifying collection and test set,
Training set picture and the corresponding sample label (high, medium and low differentiation degree) Jing Guo binaryzation are input in classifier and are instructed
Practice, the feature vector for classification of output 1 × 3:
yi=[yI, 1 yI, 2 yI, 3] (1)
The corresponding input picture of each element of the feature vector of output is likely to be at the probability of high, medium and low differentiation degree
Value, final output are the maximum differentiation degree of corresponding element probability value.
It, can be using transfer learning (Transfer Learning) during building convolutional neural networks model
Method, transfer learning method can inhibit over-fitting, while can improve the performance of the lower classifier of small data quantity condition training.
The embodiment of the present invention realizes transfer learning by importing the parameter of other pre-training models.
In a specific embodiment of the present invention, the mould of two kinds of convolutional neural networks of VGG16 and Inception-V3 is had chosen
Type can also choose other kinds of neural network in other embodiments, such as use depth residual error neural network (ResNet) generation
For traditional VGG convolutional neural networks, lift scheme complexity improves ability in feature extraction.Before training, the embodiment of the present invention
The parameter of model pre-training is first imported, pre-training parameter is obtained by ImageNet data set pre-training, and the parameter of pre-training is usual
It cannot change or can only carry out fine-tune during training again, the embodiment of the present invention uses the side of fine-tune
Method, the fine-tune for carrying out learning rate to latter 8 layers of VGG16 and being 0.0001, learns latter 249 layers of Inception-V3
The fine-tune that habit rate is 0.0001.
Preferably, it in order to go on smoothly training process, can be added behind each convolutional neural networks model complete
Articulamentum, as shown in Figure 3.Full connection thereafter will can be inputted after the characteristic spectrum flaky process of convolutional neural networks output
Layer is handled, the effect of flaky process be the characteristic expansion of extracting convolutional neural networks as an one-dimensional feature to
Amount.In some embodiments, can be inserted among the subsequent full articulamentum of convolutional neural networks model batch normalization layer and
Drop-out layers come inhibit gradient to disappear, gradient explosion and overfitting problem, Drop-out rate are 0.5, last full articulamentum is defeated
Softmax layers to be arrived out to classify, the selection of target loss function intersects entropy function, and optimizer uses AdamOptimizer,
Learning rate is 0.0005, the batch size selected in training process (the size criticized disposably inputs the quantity of picture)
It is 64, (process that all pictures are all input into neural metwork training in training set is one epoch to training epochs, can
To be interpreted as the number being trained using pictures all in training set) it is 80.Finally, saving keeps verifying collection accuracy rate highest
Model parameter checkpoint, the last model parameter as two convolutional neural networks.
After being trained to full articulamentum, we carry out Fine-tune operation to reel product neural network model, and use is small
Learning rate the parameter of pre-training is trained.The convolutional neural networks of VGG16, Inception-v3 structure have been trained
Cheng Hou, the feature vector f of output 1 × 3VGG16, fInception-v3The as feature vector of deep learning method extraction, finally by this
Two feature vectors are stitched together, and are inputted in a new full Connection Neural Network again, to export final classification knot
Fruit, and before full Connection Neural Network new using this, also need using training set picture to the full Connection Neural Network into
Row training.
Specifically, the embodiment of the present invention includes: to the method that classifier is trained
Training convolutional neural networks VGG16, Inception-V3 is set separately and corresponds to the i.e. full connection of full articulamentum thereafter
The hyper parameter of layer two and full articulamentum three, hyper parameter are as shown in table 2.
The hyper parameter set before the training of table 2
| learning rate | fine-tune learning rate | epochs | batch size | drop-out rate |
| 0.0005 | 0.0001 | 80 | 64 | 0.5 |
It imports from the model parameter downloaded on ImageNet into VGG16 and Inception-V3 model.
The image for training is imported to VGG16 and Inception-V3 neural network and full articulamentum two and full articulamentum
Three are trained, and obtain trained two convolutional neural networks and full articulamentum two and full articulamentum three.
After training, the hyper parameter of two convolutional neural networks and full articulamentum two and full articulamentum three is reset simultaneously
Fixed, the hyper parameter reset is as shown in table 3.
The hyper parameter reset after the training of table 3
| learning rate | fine-tunelearning rate | epochs | batch size | drop-out rate |
| 0.0 | 0.000 | 80 | 64 | 0.5 |
It imports and full articulamentum one is trained for the image of training, obtain trained full articulamentum one.
Three, classifier is tested
Image to be tested is input in trained classifier, VGG16 and Inception-V3 convolutional neural networks
Feature vector is extracted from image respectively, the feature vector f1 and f2 of output are stitched together be input to full articulamentum one into
The further Feature Dimension Reduction of row, the feature vector f3 that output is one 1 × 3 are used for last classification, and classification results are by feature vector f3
The middle maximum element of numerical value determines.
Four, classifier performance is assessed
In machine learning field, the assessment to classifier performance is an important job, and its evaluation index often has
Following several points: accuracy rate (accuracy), accurate rate (precision), recall rate (recall) and F1- estimate.Wherein, accurately
Rate be for given test data set, the ratio between sample number and total number of samples that classifier is correctly classified, accurate rate reflect by
Classifier be determined as in positive sample be really positive sample specific gravity, recall rate reflects the positive example being appropriately determined and (referring to classification just
True sample) specific gravity of the total amount of positive sample in total is accounted for, F1- estimates, and is a finger for comprehensively considering accurate rate and recall rate
Mark, the calculation formula of four kinds of indexs are as follows:
Wherein, TP (True Positive) is the quantity for the positive sample that the convolutional neural networks prediction being trained to is positive,
FP (False Positive) is the quantity for the negative sample that the convolutional neural networks prediction being trained to is positive, FN (False
Negative) be the quantity of positive sample that the convolutional neural networks prediction being trained to is negative, TN (True Negative) be by
The quantity for the negative sample that trained convolutional neural networks prediction is negative.Multi-class (present invention is by cervical carcinoma point of the invention
High, normal, basic three phases are divided into, each stage is considered as a classification) in statistics, the sample for the classification studied at this time is positive sample
This, the sample standard deviation of other classifications is negative sample.
The classifier performance assessment result of the embodiment of the present invention is as shown in table 4.As shown in Table 4, finally for low differentiation journey
The classification accuracy of the cervical cancer tissues pathology picture of degree can achieve 96.05%, and middle differentiation can achieve 58.68%, height
Differentiation can achieve 85.39%.
4 classifier performance assessment result of table
Fig. 4 is the scatter plot that the F1- of the classifier of training of the embodiment of the present invention estimates, and F1-, which estimates, can represent classifier
Performance, F1- measure value is higher, and the robustness of classifier is stronger.
Fig. 5 illustrates the cervical cancer tissues pathological image of classifier successful classification, the cervical cancer tissues of low differentiation degree
The cell shape of pathological image is very irregular, and eucaryotic cell structure is difficult to differentiate, and the cell arrangement in middle differentiation degree image is not
Rule, but remain eucaryotic cell structure substantially, the cell arrangement neat compact in differentiated degree image, shape it is relatively full and
Rule.Since the feature of the histopathology image of middle differentiation degree is not obvious enough, feature is occupy between low differentiation and differentiated,
So being easily confused, classification accuracy is low compared to other two classes.
Another aspect of the present invention additionally provides a kind of cervical cancer tissues pathological image analytical equipment based on deep learning,
Including image acquiring device, classifier training device and classification results output device, it is preferable that further include classifier performance assessment
Device, each device are corresponded with the step of above-mentioned analysis method.
Those of ordinary skill in the art will appreciate that all or part of the steps in the various methods of above-described embodiment is can
To be done through the relevant hardware of the program instructions, which be can be stored in a computer readable storage medium, and storage is situated between
Matter may include: read-only memory (ROM, Read Only Memory), random access memory (RAM, Random Access
Memory), disk or CD etc..
The invention is not limited to specific embodiments above-mentioned.The present invention, which expands to, any in the present specification to be disclosed
New feature or any new combination, and disclose any new method or process the step of or any new combination.
Claims (10)
1. a kind of cervical cancer tissues pathological image analysis method based on deep learning characterized by comprising
Step 1 obtains cervical cancer tissues pathological image, and image tag is arranged for every image；
Step 2 obtains trained classifier, the classifier packet based on being trained for trained image to classifier
Two convolutional neural networks and full articulamentum one are included, the input terminal of the full articulamentum one connects described two convolutional neural networks
Output end, to the method that classifier is trained include: based on for trained image respectively to two convolutional neural networks
It is trained, obtains trained two convolutional neural networks；The parameter of fixed trained two convolutional neural networks, is based on
Full articulamentum one is trained for trained image, obtains trained full articulamentum one；
Image to be tested is input in trained classifier by step 3, and two convolutional neural networks are respectively from image
Feature vector is extracted, the feature vector f1 and f2 of output are stitched together and are input to full articulamentum one, exports feature vector
F3, classification results are determined by the maximum element of numerical value in feature vector f3.
2. a kind of cervical cancer tissues pathological image analysis method based on deep learning according to claim 1, special
Sign is, full articulamentum two and full articulamentum three is connected separately with behind two in classifier convolutional neural networks, at this time
The input terminal of full articulamentum one connects the output end of full articulamentum two and full articulamentum three, is training two convolution minds in step 2
Through being also trained to full articulamentum two and full articulamentum three while network.
3. a kind of cervical cancer tissues pathological image analysis method based on deep learning according to claim 1, special
Sign is, step 1 further include: the image x for obtaining eachiIt is divided into 16 equal-sized subgraph zij, use mirror image side
The edge filling picture z ' that subgraph filling growth is equal with widthij, for each picture z 'ij, carry out 0 ° of picture rotation, 90 °,
180 ° and 270 ° operations and flip horizontal, flip vertical and channel turning operation, i=1,2 ..., n, j=1,2..., 16,
N is total number of images.
4. a kind of cervical cancer tissues pathological image analysis method based on deep learning according to claim 2, special
Sign is, specifically includes in step 2 to the training method of classifier: two convolutional neural networks and full articulamentum is set separately
Two and full articulamentum three hyper parameter；It imports from the model parameter downloaded on ImageNet into two convolutional neural networks；It leads
Enter and two convolutional neural networks and full articulamentum two and full articulamentum three are trained for trained image, is trained
Two convolutional neural networks and full articulamentum two and full articulamentum three；Reset two convolutional neural networks and full articulamentum
Two and full articulamentum three hyper parameter and fixation；It imports and full articulamentum one is trained for the image of training, trained
Good full articulamentum one.
5. a kind of cervical cancer tissues pathological image analysis method based on deep learning according to claim 1, special
Sign is, further includes step 4, carries out Performance Evaluation to classifier, evaluation index includes accuracy rate accuracy, accurate rate
Precision, recall rate recall and F1- estimate, and the calculation formula of each index is as follows:
Wherein, TP is the quantity for the positive sample that the convolutional neural networks prediction being trained to is positive, and FP is the convolution being trained to
The quantity for the negative sample that neural network prediction is positive, FN are the numbers for the positive sample that the convolutional neural networks prediction being trained to is negative
Amount, TN are the quantity for the negative sample that the convolutional neural networks prediction being trained to is negative.
6. a kind of cervical cancer tissues pathological image analysis side based on deep learning according to claim 1-5
Method, which is characterized in that described two convolutional neural networks are respectively VGG16 and Inception-V3.
7. a kind of cervical cancer tissues pathological image analytical equipment based on deep learning characterized by comprising
Image tag is arranged for obtaining cervical cancer tissues pathological image, and for every image in image acquiring device；
Classifier training device, for obtaining trained classifier based on being trained for trained image to classifier,
The classifier includes two convolutional neural networks and full articulamentum one, and the input terminal connection of the full articulamentum one is described two
The output end of convolutional neural networks, to the method that classifier is trained include: based on for trained image respectively to two
Convolutional neural networks are trained, and obtain trained two convolutional neural networks；Fixed trained two convolutional Neural nets
The parameter of network obtains trained full articulamentum one based on being trained for trained image to full articulamentum one；
Classification results output device, for image to be tested to be input in trained classifier, two convolutional Neural nets
Network extracts feature vector from image respectively, and the feature vector f1 and f2 of output are stitched together and are input to full articulamentum
One, feature vector f3 is exported, classification results are determined by the maximum element of numerical value in feature vector f3.
8. a kind of cervical cancer tissues pathological image analytical equipment based on deep learning according to claim 7, special
Sign is, full articulamentum two and full articulamentum three is connected separately with behind two in classifier convolutional neural networks, at this time
The input terminal of full articulamentum one connects the output end of full articulamentum two and full articulamentum three, and classifier training device is in training two
Also full articulamentum two and full articulamentum three are trained while convolutional neural networks.
9. a kind of cervical cancer tissues pathological image analytical equipment based on deep learning according to claim 7, special
Sign is, further includes classifier performance assessment device, for carrying out Performance Evaluation to classifier, evaluation index includes accuracy rate
Accuracy, accurate rate precision, recall rate recall and F1- estimate, and the calculation formula of each index is as follows:
Wherein, TP is the quantity for the positive sample that the convolutional neural networks prediction being trained to is positive, and FP is the convolution being trained to
The quantity for the negative sample that neural network prediction is positive, FN are the numbers for the positive sample that the convolutional neural networks prediction being trained to is negative
Amount, TN are the quantity for the negative sample that the convolutional neural networks prediction being trained to is negative.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
The step of method described in any one of claims 1 to 6 is realized when being executed by processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811212019.7A CN109376777A (en) | 2018-10-18 | 2018-10-18 | Cervical cancer tissues pathological image analysis method and equipment based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811212019.7A CN109376777A (en) | 2018-10-18 | 2018-10-18 | Cervical cancer tissues pathological image analysis method and equipment based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109376777A true CN109376777A (en) | 2019-02-22 |
Family
ID=65400787
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811212019.7A Withdrawn CN109376777A (en) | 2018-10-18 | 2018-10-18 | Cervical cancer tissues pathological image analysis method and equipment based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109376777A (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109978771A (en) * | 2019-04-08 | 2019-07-05 | 哈尔滨理工大学 | Cell image rapid fusion method based on content analysis |
| CN110009097A (en) * | 2019-04-17 | 2019-07-12 | 电子科技大学 | The image classification method of capsule residual error neural network, capsule residual error neural network |
| CN110046670A (en) * | 2019-04-24 | 2019-07-23 | 北京京东尚科信息技术有限公司 | Feature vector dimension reduction method and device |
| CN110084270A (en) * | 2019-03-22 | 2019-08-02 | 上海鹰瞳医疗科技有限公司 | Pathological section image-recognizing method and equipment |
| CN110458223A (en) * | 2019-08-06 | 2019-11-15 | 湖南省华芯医疗器械有限公司 | Tumor of bronchus automatic testing method and detection system under a kind of scope |
| CN110472676A (en) * | 2019-08-05 | 2019-11-19 | 首都医科大学附属北京朝阳医院 | Stomach morning cancerous tissue image classification system based on deep neural network |
| CN111062411A (en) * | 2019-11-06 | 2020-04-24 | 北京大学 | Method, apparatus and device for identifying multiple compounds from mass spectrometry data |
| CN111134735A (en) * | 2019-12-19 | 2020-05-12 | 复旦大学附属中山医院 | Lung cell pathology rapid on-site evaluation system and method and computer readable storage medium |
| CN111242131A (en) * | 2020-01-06 | 2020-06-05 | 北京十六进制科技有限公司 | Method, storage medium and device for image recognition in intelligent marking |
| CN111274903A (en) * | 2020-01-15 | 2020-06-12 | 合肥工业大学 | Cervical cell image classification method based on graph convolution neural network |
| CN111462076A (en) * | 2020-03-31 | 2020-07-28 | 湖南国科智瞳科技有限公司 | Method and system for detecting fuzzy area of full-slice digital pathological image |
| 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 |
| CN111882001A (en) * | 2020-08-05 | 2020-11-03 | 武汉呵尔医疗科技发展有限公司 | Cervical cell image classification method based on cell biological characteristic-convolutional neural network |
| CN112309068A (en) * | 2020-10-29 | 2021-02-02 | 电子科技大学中山学院 | Forest fire early warning method based on deep learning |
| CN112861916A (en) * | 2021-01-13 | 2021-05-28 | 武汉希诺智能医学有限公司 | Invasive cervical carcinoma pathological image classification method and system |
| CN113408620A (en) * | 2021-06-21 | 2021-09-17 | 西安工业大学 | Classification method for breast tissue pathological images |
| CN113762379A (en) * | 2021-09-07 | 2021-12-07 | 福州迈新生物技术开发有限公司 | Method for generating training data based on immunohistochemistry and storage device |
| CN115908954A (en) * | 2023-03-01 | 2023-04-04 | 四川省公路规划勘察设计研究院有限公司 | Geological disaster hidden danger identification system and method based on artificial intelligence and electronic equipment |
| CN117173485A (en) * | 2023-09-18 | 2023-12-05 | 西安交通大学医学院第二附属医院 | Intelligent classification system method and system for lung cancer tissue pathological images |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6823129B1 (en) * | 2000-02-04 | 2004-11-23 | Quvis, Inc. | Scaleable resolution motion image recording and storage system |
| CN106548178A (en) * | 2016-09-26 | 2017-03-29 | 深圳大学 | A kind of semantic feature auto-scoring method and system based on Lung neoplasm CT images |
| CN107045720A (en) * | 2017-05-04 | 2017-08-15 | 深圳硅基智能科技有限公司 | Artificial neural network and system for recognizing eye fundus image lesion |
| CN107545302A (en) * | 2017-08-02 | 2018-01-05 | 北京航空航天大学 | A kind of united direction of visual lines computational methods of human eye right and left eyes image |
| KR20180092453A (en) * | 2017-02-09 | 2018-08-20 | 한국기술교육대학교 산학협력단 | Face recognition method Using convolutional neural network and stereo image |
-
2018
- 2018-10-18 CN CN201811212019.7A patent/CN109376777A/en not_active Withdrawn
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6823129B1 (en) * | 2000-02-04 | 2004-11-23 | Quvis, Inc. | Scaleable resolution motion image recording and storage system |
| CN106548178A (en) * | 2016-09-26 | 2017-03-29 | 深圳大学 | A kind of semantic feature auto-scoring method and system based on Lung neoplasm CT images |
| KR20180092453A (en) * | 2017-02-09 | 2018-08-20 | 한국기술교육대학교 산학협력단 | Face recognition method Using convolutional neural network and stereo image |
| CN107045720A (en) * | 2017-05-04 | 2017-08-15 | 深圳硅基智能科技有限公司 | Artificial neural network and system for recognizing eye fundus image lesion |
| CN107545302A (en) * | 2017-08-02 | 2018-01-05 | 北京航空航天大学 | A kind of united direction of visual lines computational methods of human eye right and left eyes image |
Non-Patent Citations (1)
| Title |
|---|
| AKILAN T ET AL: "《Effect of fusing features from multiple DCNN architectures in image classification》", 《IET IMAGE PROCESSING》 * |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110084270A (en) * | 2019-03-22 | 2019-08-02 | 上海鹰瞳医疗科技有限公司 | Pathological section image-recognizing method and equipment |
| CN109978771A (en) * | 2019-04-08 | 2019-07-05 | 哈尔滨理工大学 | Cell image rapid fusion method based on content analysis |
| CN110009097A (en) * | 2019-04-17 | 2019-07-12 | 电子科技大学 | The image classification method of capsule residual error neural network, capsule residual error neural network |
| CN110009097B (en) * | 2019-04-17 | 2023-04-07 | 电子科技大学 | Capsule residual error neural network and image classification method of capsule residual error neural network |
| CN110046670A (en) * | 2019-04-24 | 2019-07-23 | 北京京东尚科信息技术有限公司 | Feature vector dimension reduction method and device |
| CN110046670B (en) * | 2019-04-24 | 2021-04-30 | 北京京东尚科信息技术有限公司 | Feature vector dimension reduction method and device |
| CN110472676A (en) * | 2019-08-05 | 2019-11-19 | 首都医科大学附属北京朝阳医院 | Stomach morning cancerous tissue image classification system based on deep neural network |
| CN110458223A (en) * | 2019-08-06 | 2019-11-15 | 湖南省华芯医疗器械有限公司 | Tumor of bronchus automatic testing method and detection system under a kind of scope |
| CN111062411A (en) * | 2019-11-06 | 2020-04-24 | 北京大学 | Method, apparatus and device for identifying multiple compounds from mass spectrometry data |
| CN111134735A (en) * | 2019-12-19 | 2020-05-12 | 复旦大学附属中山医院 | Lung cell pathology rapid on-site evaluation system and method and computer readable storage medium |
| CN111242131A (en) * | 2020-01-06 | 2020-06-05 | 北京十六进制科技有限公司 | Method, storage medium and device for image recognition in intelligent marking |
| CN111274903A (en) * | 2020-01-15 | 2020-06-12 | 合肥工业大学 | Cervical cell image classification method based on graph convolution neural network |
| CN111274903B (en) * | 2020-01-15 | 2022-12-06 | 合肥工业大学 | Cervical cell image classification method based on graph convolution neural network |
| CN111462076A (en) * | 2020-03-31 | 2020-07-28 | 湖南国科智瞳科技有限公司 | Method and system for detecting fuzzy area of full-slice digital pathological image |
| 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 |
| CN111882001A (en) * | 2020-08-05 | 2020-11-03 | 武汉呵尔医疗科技发展有限公司 | Cervical cell image classification method based on cell biological characteristic-convolutional neural network |
| CN112309068B (en) * | 2020-10-29 | 2022-09-06 | 电子科技大学中山学院 | Forest fire early warning method based on deep learning |
| CN112309068A (en) * | 2020-10-29 | 2021-02-02 | 电子科技大学中山学院 | Forest fire early warning method based on deep learning |
| CN112861916A (en) * | 2021-01-13 | 2021-05-28 | 武汉希诺智能医学有限公司 | Invasive cervical carcinoma pathological image classification method and system |
| CN113408620A (en) * | 2021-06-21 | 2021-09-17 | 西安工业大学 | Classification method for breast tissue pathological images |
| CN113762379A (en) * | 2021-09-07 | 2021-12-07 | 福州迈新生物技术开发有限公司 | Method for generating training data based on immunohistochemistry and storage device |
| CN113762379B (en) * | 2021-09-07 | 2022-06-07 | 福州迈新生物技术开发有限公司 | Method for generating training data based on immunohistochemistry and storage device |
| WO2023035728A1 (en) * | 2021-09-07 | 2023-03-16 | 福州迈新生物技术开发有限公司 | Method for generating training data based on immunohistochemistry, and storage device |
| CN115908954A (en) * | 2023-03-01 | 2023-04-04 | 四川省公路规划勘察设计研究院有限公司 | Geological disaster hidden danger identification system and method based on artificial intelligence and electronic equipment |
| CN117173485A (en) * | 2023-09-18 | 2023-12-05 | 西安交通大学医学院第二附属医院 | Intelligent classification system method and system for lung cancer tissue pathological images |
| CN117173485B (en) * | 2023-09-18 | 2024-02-13 | 西安交通大学医学院第二附属医院 | Intelligent classification system method and system for lung cancer tissue pathological images |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109376777A (en) | Cervical cancer tissues pathological image analysis method and equipment based on deep learning | |
| Man et al. | Classification of breast cancer histopathological images using discriminative patches screened by generative adversarial networks | |
| Gertych et al. | Machine learning approaches to analyze histological images of tissues from radical prostatectomies | |
| Dov et al. | Weakly supervised instance learning for thyroid malignancy prediction from whole slide cytopathology images | |
| Beevi et al. | Automatic mitosis detection in breast histopathology images using convolutional neural network based deep transfer learning | |
| CN104933711B (en) | A kind of automatic fast partition method of cancer pathology image | |
| JP5315411B2 (en) | Mitotic image detection device and counting system, and method for detecting and counting mitotic images | |
| CN112215117A (en) | Abnormal cell identification method and system based on cervical cytology image | |
| CN111985536A (en) | Gastroscope pathological image classification method based on weak supervised learning | |
| Dov et al. | Thyroid cancer malignancy prediction from whole slide cytopathology images | |
| CN108416379A (en) | Method and apparatus for handling cervical cell image | |
| US20200372638A1 (en) | Automated screening of histopathology tissue samples via classifier performance metrics | |
| CN107871314A (en) | A kind of sensitive image discrimination method and device | |
| WO2013019856A1 (en) | Automated malignancy detection in breast histopathological images | |
| Zhang et al. | Automatic detection of invasive ductal carcinoma based on the fusion of multi-scale residual convolutional neural network and SVM | |
| CN111639697B (en) | Hyperspectral image classification method based on non-repeated sampling and prototype network | |
| BenTaieb et al. | Automatic diagnosis of ovarian carcinomas via sparse multiresolution tissue representation | |
| Rampun et al. | Breast density classification using local ternary patterns in mammograms | |
| Abbasi-Sureshjani et al. | Molecular subtype prediction for breast cancer using H&E specialized backbone | |
| Zhang et al. | Research on application of classification model based on stack generalization in staging of cervical tissue pathological images | |
| Tsaku et al. | Texture-based deep learning for effective histopathological cancer image classification | |
| CN113420793A (en) | Improved convolutional neural network ResNeSt 50-based gastric ring cell carcinoma classification method | |
| CN111680553A (en) | Pathological image identification method and system based on depth separable convolution | |
| Arar et al. | High-quality immunohistochemical stains through computational assay parameter optimization | |
| Guo et al. | Pathological Detection of Micro and Fuzzy Gastric Cancer Cells Based on Deep Learning. |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20191122 Address after: 610043 No.129, middle Juqiao street, Wuhou District, Chengdu, Sichuan Province Applicant after: Sichuan Smart Motion Muniu Intelligent Technology Co., Ltd. Address before: West high tech Zone Fucheng Road in Chengdu city of Sichuan province 610000 399 No. 7 Building 1 unit 11 floor No. 1107 Applicant before: SICHUAN MONIULIUMA INTELLIGENT TECHNOLOGY CO., LTD. |
|
| TA01 | Transfer of patent application right | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190222 |
|
| WW01 | Invention patent application withdrawn after publication |