CN108647731A - Cervical carcinoma identification model training method based on Active Learning - Google Patents
Cervical carcinoma identification model training method based on Active Learning Download PDFInfo
- Publication number
- CN108647731A CN108647731A CN201810456532.4A CN201810456532A CN108647731A CN 108647731 A CN108647731 A CN 108647731A CN 201810456532 A CN201810456532 A CN 201810456532A CN 108647731 A CN108647731 A CN 108647731A
- Authority
- CN
- China
- Prior art keywords
- cell
- cervical carcinoma
- identification model
- layer
- cervical
- 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
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/03—Recognition of patterns in medical or anatomical images
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The cervical carcinoma identification model training method based on Active Learning that the invention discloses a kind of, includes the following steps:Prepare multiple cervical carcinomas slice digital picture and build a sample database, digital image training identification model is sliced using the cervical carcinoma in the sample database;It is sliced digital picture for every cervical carcinoma in sample database, by exporting corresponding recognition result after the identification of identification model, the recognition result includes the cell that the identification model is difficult to;In the recognition result, the cell being difficult to is labeled using artificial notation methods, and the cervical carcinoma digital picture after mark is updated in sample database, and re -training identification model is identified, until in all recognition results there is no the cell being difficult to, compared with prior art, it is fast that the beneficial effects of the invention are as follows recognition speeds, accuracy of identification is high, can automatic fitration fall the apparent cervical cancer cell of big measure feature.
Description
Technical field
The present invention relates to cancer identification field more particularly to a kind of cervical carcinoma identification model training sides based on Active Learning
Method.
Background technology
Cervical carcinoma is a kind of malignant tumour seriously endangering female reproductive health, according to incompletely statistics, the annual uterine neck in the whole world
The new cases of cancer nearly 500,000 people, and cervical carcinoma annual new cases in China's are about 13.2 ten thousand people, account for about the world total
28%.In order to improve the accuracy rate of diagnosis of cervical cancer, in recent years, domestic and international many medical research team start to be dedicated to research respectively
The recognition methods of kind cervical cancer cell, but recognizer is still using the identification model for having supervision mostly both at home and abroad at present, also
It is to say, this identification model, which still needs largely manually to mark just by mark doctor, can obtain cervical cancer cell knowledge higher
Not rate.
But since the digital image resolution of current cervical carcinoma is very high, mark doctor marks sick cell needs and goes through
Each corner of image, this has taken prodigious work load to doctor.And, it is generally the case that each cervical carcinoma number
All there is the apparent sick cell of big measure feature in word image, if mark doctor is required for the apparent lesion of these features every time
Cell marks out one by one to be come, and can increase the multiplicity of their routine works, and the apparent sick cell of these features is to training
Identification model has little significance.Therefore, it is desirable to study it is a kind of can be based on the cervical carcinoma identification model training method of Active Learning, with drop
The working strength of low mark doctor, improves the discrimination of cervical cancer cell.
Invention content
The cervical carcinoma identification model training method based on Active Learning that the object of the present invention is to provide a kind of, solve with
Upper technical problem.
The present invention, which solves its technical problem, to be achieved through the following technical solutions:
A kind of cervical carcinoma identification model training method based on Active Learning, prepares multiple cervical carcinoma number of slices in advance first
Word image simultaneously builds a sample database, then trains to form a knowledge for cervical cancer cell to be identified using the sample database
Other model;
The process of the training identification model specifically includes:
Step 1:The identification model is trained using the cervical carcinoma slice digital picture in the sample database;
Step 2:It is sliced digital picture for every cervical carcinoma in the sample database, passes through the identification model
Corresponding recognition result is exported after identification, the recognition result includes the cell that the identification model is difficult to;
Step 3:In the recognition result, the cell being difficult to is labeled by the way of manually marking, and
By in the digital image update to the sample database of cervical carcinoma slice after mark, it is then returned to the step 1, and all
Step 4 is turned to when the cell being difficult to being not present in the recognition result;
Step 4:The identification model training is finished and is preserved, with backed off after random.
Further, in the step 2, judge whether the cell in the cervical carcinoma slice digital picture is to be difficult to
Cell the step of specifically include:
Step 21:Each cell that the identification model is sliced in digital picture the cervical carcinoma is identified, and
Export the confidence level set of each cell, the confidence level set include corresponding cell belong to it is different classes of each
Confidence level;
Step 22:It handles to obtain corresponding cell according to highest two confidence levels of numerical value in the confidence level set
Identification difficulty;
Step 23:Judge whether the identification difficulty is more than a predetermined threshold value:
If so, indicating that corresponding cell can be identified;
If not, then it represents that corresponding cell is difficult to be identified.
Further, the predetermined threshold value is 0.4.
Further, the identification model is sliced digitized map using the object detection method based on SSD to the cervical carcinoma
As the cell in 1 is identified, specific identification step includes:
Step S1 builds cervical carcinoma feature extraction network according to the sample database in the identification model, needs to illustrate
, it is sliced in digital picture in the cervical carcinoma comprising the sick cell that may meet cervical cancer cell feature and normally thin
Born of the same parents, the cervical carcinoma feature extraction network identify that output includes the disease for meeting cervical cancer cell feature by way of successively convolution
The Minimum Area figure for becoming cell and part normal cell filters out in the cervical carcinoma slice digital picture and sick cell is not present
Region;
Step S2 builds cervical cancer cell extraction network, the uterine neck in the identification model according to the sample database
Cancer cell extracts network and carries out further convolution identification to the Minimum Area figure that the step S1 is extracted, and exports corresponding
Recognition result, the recognition result have filtered out the normal cell in the Minimum Area figure, in the Minimum Area figure
The sick cell for meeting cervical cancer cell feature is retained.
Step S3 builds cervical cancer cell sorter network, the uterine neck in the identification model according to the sample database
Cancer cell sorter network carries out further convolution identification to the recognition result that the step S2 is exported, and according to the knowledge
The other preset cell category of model, specific kind of the final sick cell for judging to meet described in identification output cervical cancer cell feature
Class, the cell category are divided into two types, positive cell and non-positive cell, and the positive cell represents cervical cancer cell,
Non- positive cell is other kinds of sick cell.
Further, the cervical carcinoma feature extraction network includes cascade nine feature extraction layers successively, the uterine neck
Cancer is sliced input data of the digital picture as first feature extraction layer, the 9th feature extraction layer it is defeated
It is that the cervical carcinoma is sliced in digital picture comprising the sick cell and normal cell for meeting cervical cancer cell feature to go out data
The Minimum Area figure.
Further, the cervical cancer cell extraction network includes five cell extraction layers being arranged in order, the cell
5th to the 9th feature extraction layer of extract layer and the cervical carcinoma feature extraction network corresponds, and the 5th extremely
Correspondence is input to the corresponding cell to the Minimum Area figure of each layer of output of the 9th feature extraction layer respectively
Extract layer, the Minimum Area figure is after the cervical cancer cell extracts the further convolution identification of network, the smallest region
The normal cell on the figure of domain is filtered, and identification exports the lesion for meeting cervical cancer cell feature on the Minimum Area figure
Cell.
Further, the cervical cancer cell classification extraction network includes five cell classification layers being arranged in order, described
Cell classification layer and the 5th to the 9th feature extraction layer of the cervical carcinoma feature extraction network correspond, and the 5th
Correspondence is input to the cervical carcinoma to the Minimum Area figure of each layer of output of a to the 9th feature extraction layer respectively
The corresponding cell classification layer of cell classification network, the Minimum Area figure pass through the further volume of the cell classification layer
Product identification, identification export the specific type for the sick cell for meeting cervical cancer cell feature on the Minimum Area figure, and by its
Finally it is confirmed as cervical cancer cell.
The innovation of the invention consists in that the cervical carcinoma identification model training method based on Active Learning to cervical cancer cell into
Row feature recognition under the premise of ensureing recognition speed, improves the identification essence of cervical cancer cell using above-mentioned recognition methods
Degree, but also can automatic fitration fall the apparent cervical cancer cell of big measure feature, leave behind a small number of sick cells being difficult to and hand over
By marking doctor's mark, the workload of mark doctor is greatly reduced, actual demand is preferably met.
Description of the drawings
Fig. 1 is the implementation flow chart of the present invention;
Fig. 2 is the method flow diagram of present invention training identification model;
Fig. 3 is that the present invention is based on the method flow diagrams that confidence level method identifies cervical cancer cell;
Fig. 4 is that the present invention is based on the method flow diagrams that the object detection method of SSD identifies cervical cancer cell;
Fig. 5 is that the present invention is based on the network structures of SSD object detection methods;
Fig. 6 is the structure chart of cervical carcinoma feature extraction network in the present invention;
Fig. 7 is the structure chart of FA type features extract layer in the present invention;
Fig. 8 is the structure chart of FB type features extract layer in the present invention;
Fig. 9 is the structure chart of FC type features extract layer in the present invention;
Figure 10 is the structure chart of FD type features extract layer in the present invention;
Figure 11 is the structure chart of FE type features extract layer in the present invention;
Figure 12 is the structure chart of FF type features extract layer in the present invention;
Figure 13 is the structure chart of cervical cancer cell extraction network and cervical cancer cell sorter network of the present invention.
Specific implementation mode
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation describes, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art obtained under the premise of not making creative work it is all its
His embodiment, shall fall within the protection scope of the present invention.
It should be noted that in the absence of conflict, the feature in embodiment and embodiment in the present invention can phase
Mutually combination.
The invention will be further described in the following with reference to the drawings and specific embodiments, but not as limiting to the invention.One
Cervical carcinoma identification model training method of the kind based on Active Learning, referring to Fig.1 and 2, first by preparing multiple palaces in advance
Neck cancer be sliced digital picture simultaneously build a sample database, then using the sample database train to be formed one for cervical cancer cell into
The identification model of row identification;
The process of the training identification model specifically includes:
Step 1:The identification model is trained using the cervical carcinoma slice digital picture in the sample database,
Obtain training data;
Step 2:It is sliced digital picture for every cervical carcinoma in the sample database, passes through the identification model
Corresponding recognition result is exported after identification, the recognition result includes the cell that the identification model is difficult to;
Step 3:In the recognition result, marking platform and using thin to what is be difficult to by way of manually marking
Born of the same parents are labeled, and by after mark annotation results i.e. the cervical carcinoma digital picture be updated in the sample database, with
After return to the step 1, and when cell being difficult to is not present in all recognition results, turns to step 4;
Step 4:The identification model training is finished and is preserved, with backed off after random.
Please refer to Fig. 3, a kind of cervical carcinoma identification model training method based on Active Learning of the invention, the step 2
In, judge that the step of whether cell in the cervical carcinoma slice digital picture is the cell being difficult to specifically includes:
Step 21:Each cell that the identification model is sliced in digital picture the cervical carcinoma is identified, and
The confidence level set for exporting each cell, is denoted as conf={ conf1,conf2,…,confn, n is the confidence level set
Include that corresponding cell belongs to different classes of each confidence level;
Step 22:It handles to obtain corresponding cell according to highest two confidence levels of numerical value in the confidence level set
Identification difficulty, highest two confidence levels of the numerical value are denoted as conf respectivelybest, confsec-best;
Step 23:Judge whether the identification difficulty is more than a predetermined threshold value:
If so, indicating that corresponding cell can be identified;
If not, then it represents that corresponding cell is difficult to be identified.
A kind of cervical carcinoma identification model training method based on Active Learning of the present invention, the predetermined threshold value are 0.4,
That is working as confbest-confsec-best≤ 0.4, indicate that the corresponding cell is difficult to identified need using artificial mark
Mode carry out further mark identification.
Fig. 4 and Fig. 5 are please referred to, a kind of cervical carcinoma identification model training method based on Active Learning of the invention is described
Identification model is identified using the cell that the object detection method based on SSD is sliced in digital picture the cervical carcinoma,
Specifically identification step includes:
Step S1 builds cervical carcinoma feature extraction network 01 according to the sample database in the identification model, needs
It is bright, it is sliced in digital picture in the cervical carcinoma comprising the sick cell that may meet cervical cancer cell feature and normally thin
Born of the same parents, the cervical carcinoma feature extraction network 01 identify that output includes to meet cervical cancer cell feature by way of successively convolution
The Minimum Area figure of sick cell and part normal cell, filters out in cervical carcinoma slice digital picture that there is no lesion is thin
The region of born of the same parents;
Step S2 builds cervical cancer cell extraction network 02, the palace in the identification model according to the sample database
Neck cancer cell extraction network 02 carries out further convolution identification, and the output phase to the Minimum Area figure that the step S1 is extracted
The recognition result answered, the recognition result have filtered out the normal cell in the Minimum Area figure, the Minimum Area
The sick cell for meeting cervical cancer cell feature in figure is retained.
Step S3 builds cervical cancer cell sorter network 03, the palace in the identification model according to the sample database
Neck cancer cell classification network 03 carries out further convolution identification to the recognition result that the step S2 is exported, and according to institute
The preset cell category of identification model is stated, the tool of the final sick cell for judging to meet described in identification output cervical cancer cell feature
Body type, the cell category are divided into two types, positive cell and non-positive cell, and it is thin that the positive cell represents cervical carcinoma
Born of the same parents, non-positive cell are other kinds of sick cell.
Please refer to Fig. 6, a kind of cervical carcinoma identification model training method based on Active Learning of the invention, the cervical carcinoma
Feature extraction network 01 includes cascade nine feature extraction layers successively, and nine feature extraction layers are respectively that fisrt feature carries
Take layer 011, second feature extract layer 012, third feature extract layer 013, fourth feature extract layer 014, fifth feature extract layer
015, sixth feature extract layer 016, seventh feature extract layer 017, eighth feature extract layer 018 and ninth feature extract layer
019, input data of the cervical carcinoma slice digital picture as the fisrt feature extract layer 011, the ninth feature carries
Take layer 019 output data be the cervical carcinoma be sliced digital picture on comprising meet cervical cancer cell feature sick cell and
The Minimum Area figure of normal cell.The program extracts the information that the cervical carcinoma is sliced digital picture by multilayer, effectively
It is extracted the Global Information of input picture.
A kind of cervical carcinoma identification model training method based on Active Learning of the present invention, the cervical carcinoma feature extraction net
Nine feature extraction layers of network 01 include six kinds of different types of feature extraction layers, six kinds of different types of feature extractions altogether
Layer is respectively FA type features extract layer 100, FB type features extract layer 101, FC type features extract layer 102, FD type spies
Levy extract layer 103, FE type features extract layer 104 and FF type features extract layer 105;The feature extraction of each type
Layer is including different types of convolutional layer or including different types of convolutional layer and different types of pond layer.The program is rolled up with multilayer
Input picture convolution is generated feature map by product mode, can effectively extract characteristics of image.
A kind of cervical carcinoma identification model training method based on Active Learning of the present invention, the convolutional layer includes CA types
Convolutional layer 200, CB types convolutional layer 201, CC types convolutional layer 202, CD types convolutional layer 203 and CE types convolutional layer 204, institute
The convolution kernel size for stating CA types convolutional layer 200 is 3*3, and step-length 1 is filled with 1;The convolution kernel of CB types convolutional layer 201 is big
Small is 3*3, and step-length 1 is filled with 6;The convolution kernel size of CC types convolutional layer 202 is 1*1, and step-length 1 is filled with 0;CD classes
The convolution kernel size of type convolutional layer 203 is 3*3, and step-length 2 is filled with 1;The convolution kernel size of CE types convolutional layer 204 is 4*
4, step-length 1 is filled with 1.
A kind of cervical carcinoma identification model training method based on Active Learning of the present invention, the pond layer includes PA types
The core size of pond layer 300 and PB types pond layer 301, the PA types Chi Huaceng300Chiization is 2*2, step-length 2, filling
It is 0;The core size of the PB types Chi Huaceng301Chiization is 3*3, and step-length 1 is filled with 1.
Fig. 7 to Figure 12 is please referred to, a kind of cervical carcinoma identification model training method based on Active Learning of the invention is described
FA type features extract layer 100 includes two CA types convolutional layers 200 and a PA types pond layer 300;The FB types are special
It includes three CA types convolutional layers 200 and a PA types pond layer 300 to levy extract layer 101;The FC type features extract layer
102 include three CA types convolutional layers 200 and a PB types pond layer 301;The FD type features extract layer 103 includes one
201, four CC types convolutional layers 202 of a CB types convolutional layer and a CD types convolutional layer 203;The FE type features extraction
Layer 104 includes a CC types convolutional layer 202 and a CD types convolutional layer 203;The FF type features extract layer 105 includes
One CC types convolutional layer 202 and a CE types convolutional layer 204.
A kind of cervical carcinoma identification model training method based on Active Learning of the present invention, the cervical carcinoma feature extraction net
The fisrt feature extract layer 011 and second feature extract layer 012 of network 01 are FA type features extract layer 100;The third is special
It is FB type features extract layer 101 to levy extract layer 013 and fourth feature extract layer 014;The fifth feature extract layer 015 is
FC type features extract layer 102;The sixth feature extract layer 016 is FD type features extract layer 103;The seventh feature carries
It is FE type features extract layer 104 to take layer 017 and eighth feature extract layer 018;The ninth feature extract layer 019 is FF types
Feature extraction layer 105, input data of the cervical carcinoma slice digital picture as the fisrt feature extract layer 011, successively
By the convolution identification of nine feature extraction layers finally the cervical carcinoma number is exported from the ninth feature extract layer 019
The Minimum Area figure of the sick cell for meeting cervical cancer cell feature and normal cell that include on sectioning image.The program is abundant
Processing capacity of the cyclic convolution network to time series, flexible Application and the feature extraction to single image is utilized, improves
The discrimination of cervical carcinoma.
A kind of cervical carcinoma identification model training method based on Active Learning of the present invention, the fisrt feature extract layer
The convolution nuclear volume of two CA types convolutional layers 200 in 011 is 64;Two in the second feature extract layer 012
The convolution nuclear volume of CA types convolutional layer 200 is 128;The volume of three convolutional layers in the third feature extract layer 013
Product nuclear volume is 256;The convolution nuclear volume of three convolutional layers in the fourth feature extract layer 014 is 512;Institute
The convolution nuclear volume for stating three in fifth feature extract layer 015 convolutional layers is 512;The sixth feature extract layer 016
In six convolutional layers, the wherein convolution nuclear volume of CB types convolutional layer 201 is 1024, four CC types convolutional layers 202
Convolution nuclear volume is respectively 1024,256,512 and 128, and the convolution nuclear volume of CD types convolutional layer 203 is 256, and described the
The convolution nuclear volume of seven and the CC types convolutional layer 202 in the 8th feature extraction layer 31 is 128, CD type convolutional layers
203 convolution nuclear volume is 256;The convolution check figure of CC types convolutional layer 202 in 9th feature extraction layer 31
Amount is 128, and the convolution nuclear volume of CE types convolutional layer 204 is 256.The program is fully extracted the feature of image.
A kind of cervical carcinoma identification model training method based on Active Learning of the present invention, PA types pond layer 300
Pond mode with PB types pond layer 301 is maximum pond.The program can reduce output feature, but also can reduce
The offset that mean value is estimated caused by convolutional layer parameter error, more retains texture information.
Please refer to Figure 13, a kind of cervical carcinoma identification model training method based on Active Learning of the invention, the uterine neck
Cancer cell extraction network 02 includes five cell extraction layers being arranged in order, respectively the first cell extraction layer 021, the second cell
Extract layer 022, third cell extraction layer 023, the 4th cell extraction layer 024 and the 5th cell extraction layer 025, the cell extraction
The fifth feature extract layer 015 of layer and the cervical carcinoma feature extraction network 01 is a pair of to the difference of ninth feature extract layer 019 one
It answers, the Minimum Area figure of each layer of fifth feature extract layer 015 to ninth feature extract layer 019 output corresponds to defeated respectively
Enter 021 to the 5th cell extraction layer 025 of the first cell extraction layer to the corresponding cell extraction layer, the Minimum Area figure
After the cervical cancer cell extracts the further convolution identification of network 02, the normal cell on the Minimum Area figure is by mistake
It filters, identification exports the sick cell for meeting cervical cancer cell feature on the Minimum Area figure.
A kind of cervical carcinoma identification model training method based on Active Learning of the present invention, the cervical cancer cell classification carry
It includes five cell classification layers being arranged in order, respectively the first cell classification layer 031, the second cell classification layer to take network 03
032, third cell classification layer 033, the 4th cell classification layer 034 and the 5th cell classification layer 035, the first cell classification layer
The fifth feature extract layer 015 of 031 to the 5th cell classification layer 035 and the cervical carcinoma feature extraction network 01 is to ninth feature
Extract layer 019 corresponds, the minimum of each layer of output of fifth feature extract layer 015 to ninth feature extract layer 019
Administrative division map corresponds to the first cell of the corresponding cell classification layer for being input to the cervical cancer cell sorter network 03 respectively
Classification 031 to the 5th cell classification layer 035 of layer, the Minimum Area figure are known by the further convolution of the cell classification layer
Not, identification exports the specific type for the sick cell for meeting cervical cancer cell feature on the Minimum Area figure, and it is final
It is confirmed as cervical cancer cell.
A kind of cervical carcinoma identification model training method based on Active Learning of the present invention, the cervical cancer cell extract net
Network 02 and the convolution kernel size of the use in the cervical cancer cell sorter network 03 are 3*3, and step-length is 1, and filling is
1。
In conclusion object detection method of the present invention using confidence level method or based on the study of SSD even depth can be
Under the premise of ensureing recognition speed, improve the accuracy of identification of cervical cancer cell, and can automatic fitration to fall big measure feature apparent
Cervical cancer cell, leave behind minority be difficult to cell be submitted to mark platform transfer to mark doctor mark, effectively reduce mark
The workload for noting doctor, can more preferably meet actually detected demand.
The foregoing is merely preferred embodiments of the present invention, are not intended to limit embodiments of the present invention and protection model
It encloses, to those skilled in the art, should can appreciate that all with made by description of the invention and diagramatic content
Equivalent replacement and obviously change obtained scheme, should all be included within the scope of the present invention.
Claims (7)
1. a kind of cervical carcinoma identification model training method based on Active Learning, which is characterized in that prepare multiple cervical carcinomas in advance
It is sliced digital picture and simultaneously builds a sample database, train using the sample database to form one for cervical cancer cell to be identified
Identification model;
The process of the training identification model specifically includes:
Step 1:The identification model is trained using the cervical carcinoma slice digital picture in the sample database;
Step 2:It is sliced digital picture for every cervical carcinoma in the sample database, passes through the identification of the identification model
After export corresponding recognition result, the recognition result includes the cell that the identification model is difficult to;
Step 3:In the recognition result, the cell being difficult to is labeled by the way of manually marking, and will mark
The cervical carcinoma digital picture after note is updated in the sample database, is then returned to the step 1, and in all identifications
Step 4 is turned to when the cell being difficult to being not present in as a result;
Step 4:The identification model training is finished and is preserved, with backed off after random.
2. the cervical carcinoma identification model training method based on Active Learning as described in claim 1, it is characterised in that:The step
In rapid 2, judge that the step of whether cell in the cervical carcinoma slice digital picture is the cell being difficult to specifically includes:
Step 21:Each cell that the identification model is sliced in digital picture the cervical carcinoma is identified, and exports
The confidence level set of each cell, the confidence level set include that corresponding cell belongs to different classes of each confidence
Degree;
Step 22:It handles to obtain the knowledge of corresponding cell according to highest two confidence levels of numerical value in the confidence level set
Other difficulty;
Step 23:Judge whether the identification difficulty is more than a predetermined threshold value:
If so, indicating that corresponding cell can be identified;
If not, then it represents that corresponding cell is difficult to be identified.
3. the cervical carcinoma identification model training method based on Active Learning as claimed in claim 2, which is characterized in that described pre-
If threshold value is 0.4.
4. the cervical carcinoma identification model training method based on Active Learning as described in claim 1, it is characterised in that:The knowledge
Other model is identified using the cell that the object detection method based on SSD is sliced in digital picture the cervical carcinoma, is had
Body identification step includes:
Step S1 builds cervical carcinoma feature extraction network, in the cervical carcinoma in the identification model according to the sample database
Include the sick cell and normal cell for meeting cervical cancer cell feature, the cervical carcinoma feature extraction net in slice digital picture
Network identifies the Minimum Area figure that output includes the sick cell and the normal cell by way of successively convolution, and filters
Fall the region of the sick cell and the normal cell is not present;
Step S2, the cervical cancer cell extraction network carries out convolution identification to the Minimum Area figure, and exports corresponding knowledge
Not as a result, having filtered out the normal cell in the Minimum Area figure in the recognition result, and retain the lesion
Cell.
Step S3, the cervical cancer cell sorter network are directed to the recognition result, according to preset thin in the identification model
Born of the same parents' type, identifies the specific type of the sick cell, and the cell category includes positive cell and non-positive cell, described
Positive cell is used to indicate the sick cell of cervical cancer cell type, and non-positive cell is for indicating other kinds of described
Sick cell.
5. the cervical carcinoma identification model training method based on Active Learning as claimed in claim 4, it is characterised in that:The palace
Neck cancer feature extraction network includes cascade nine feature extraction layers successively, and the cervical carcinoma slice digital picture is as described the
The output data of the input data of one feature extraction layer, the 9th feature extraction layer is the Minimum Area figure.
6. the cervical carcinoma identification model training method based on Active Learning as claimed in claim 5, it is characterised in that:The palace
Neck cancer cell extraction network includes five cell extraction layers being arranged in order, and the cell extraction layer is carried with the cervical carcinoma feature
Take the 5th to the 9th of the network feature extraction layer to correspond, the 5th to the 9th feature extraction layer it is every
The Minimum Area figure of one layer of output is separately input to the corresponding cell extraction layer, described in the Minimum Area figure passes through
The sick cell on the Minimum Area figure is exported after the convolution identification of cell extraction layer.
7. the cervical carcinoma identification model training method based on Active Learning as claimed in claim 6, it is characterised in that:The palace
Neck cancer cell classification extraction network includes five cell classification layers being arranged in order, and the cell classification layer is special with the cervical carcinoma
5th to the 9th feature extraction layer of sign extraction network corresponds, the 5th to the 9th feature extraction layer
The Minimum Area figure of each layer of output corresponding be respectively input to the corresponding described of the cervical cancer cell sorter network
Cell classification layer, the Minimum Area figure export specific kind of the sick cell after the identification of the cell classification layer
Class.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810456532.4A CN108647731A (en) | 2018-05-14 | 2018-05-14 | Cervical carcinoma identification model training method based on Active Learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810456532.4A CN108647731A (en) | 2018-05-14 | 2018-05-14 | Cervical carcinoma identification model training method based on Active Learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108647731A true CN108647731A (en) | 2018-10-12 |
Family
ID=63755236
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810456532.4A Pending CN108647731A (en) | 2018-05-14 | 2018-05-14 | Cervical carcinoma identification model training method based on Active Learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108647731A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109800805A (en) * | 2019-01-14 | 2019-05-24 | 上海联影智能医疗科技有限公司 | Image processing system and computer equipment based on artificial intelligence |
CN112334990A (en) * | 2019-06-04 | 2021-02-05 | 艾多特公司 | Automatic cervical cancer diagnosis system |
US12087445B2 (en) | 2019-06-04 | 2024-09-10 | Aidot Inc. | Automatic cervical cancer diagnosis system |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104503960A (en) * | 2015-01-07 | 2015-04-08 | 渤海大学 | Text data processing method for English translation |
CN105224947A (en) * | 2014-06-06 | 2016-01-06 | 株式会社理光 | Sorter training method and system |
CN107067025A (en) * | 2017-02-15 | 2017-08-18 | 重庆邮电大学 | A kind of data automatic marking method based on Active Learning |
CN107704878A (en) * | 2017-10-09 | 2018-02-16 | 南京大学 | A kind of high-spectral data storehouse semi-automation method for building up based on deep learning |
CN107832780A (en) * | 2017-10-17 | 2018-03-23 | 北京木业邦科技有限公司 | Low confidence sample processing method and system are sorted based on artificial intelligence plank |
CN107832838A (en) * | 2017-11-28 | 2018-03-23 | 北京羽医甘蓝信息技术有限公司 | The method and apparatus for evaluating cell smear sample satisfaction |
-
2018
- 2018-05-14 CN CN201810456532.4A patent/CN108647731A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105224947A (en) * | 2014-06-06 | 2016-01-06 | 株式会社理光 | Sorter training method and system |
CN104503960A (en) * | 2015-01-07 | 2015-04-08 | 渤海大学 | Text data processing method for English translation |
CN107067025A (en) * | 2017-02-15 | 2017-08-18 | 重庆邮电大学 | A kind of data automatic marking method based on Active Learning |
CN107704878A (en) * | 2017-10-09 | 2018-02-16 | 南京大学 | A kind of high-spectral data storehouse semi-automation method for building up based on deep learning |
CN107832780A (en) * | 2017-10-17 | 2018-03-23 | 北京木业邦科技有限公司 | Low confidence sample processing method and system are sorted based on artificial intelligence plank |
CN107832838A (en) * | 2017-11-28 | 2018-03-23 | 北京羽医甘蓝信息技术有限公司 | The method and apparatus for evaluating cell smear sample satisfaction |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109800805A (en) * | 2019-01-14 | 2019-05-24 | 上海联影智能医疗科技有限公司 | Image processing system and computer equipment based on artificial intelligence |
CN112334990A (en) * | 2019-06-04 | 2021-02-05 | 艾多特公司 | Automatic cervical cancer diagnosis system |
US12087445B2 (en) | 2019-06-04 | 2024-09-10 | Aidot Inc. | Automatic cervical cancer diagnosis system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107563381B (en) | Multi-feature fusion target detection method based on full convolution network | |
CN106296699A (en) | Cerebral tumor dividing method based on deep neural network and multi-modal MRI image | |
CN109035255A (en) | A kind of sandwich aorta segmentation method in the CT image based on convolutional neural networks | |
CN107730507A (en) | A kind of lesion region automatic division method based on deep learning | |
CN110458847A (en) | Automatic coronary artery segmentation and center line extraction method based on CTA image | |
CN107610123A (en) | A kind of image aesthetic quality evaluation method based on depth convolutional neural networks | |
CN106780460A (en) | A kind of Lung neoplasm automatic checkout system for chest CT image | |
CN108334909A (en) | Cervical carcinoma TCT digital slices data analysing methods based on ResNet | |
CN107492095A (en) | Medical image pulmonary nodule detection method based on deep learning | |
CN112070772A (en) | Blood leukocyte image segmentation method based on UNet + + and ResNet | |
CN108647731A (en) | Cervical carcinoma identification model training method based on Active Learning | |
CN108389190A (en) | A kind of Lung neoplasm automatic testing method based on deep learning method | |
CN110047082A (en) | Pancreatic Neuroendocrine Tumors automatic division method and system based on deep learning | |
CN109325942A (en) | Eye fundus image Structural Techniques based on full convolutional neural networks | |
CN110070540A (en) | Image generating method, device, computer equipment and storage medium | |
CN109102498A (en) | A kind of method of cluster type nucleus segmentation in cervical smear image | |
CN107871316A (en) | A kind of X-ray hand bone interest region extraction method based on deep neural network | |
CN110321844A (en) | A kind of quick iris detection method based on convolutional neural networks | |
CN107578402A (en) | A kind of detecting system and its application method of the pulmonary nodule based on deep learning | |
CN108986115A (en) | Medical image cutting method, device and intelligent terminal | |
CN108629369A (en) | A kind of Visible Urine Sediment Components automatic identifying method based on Trimmed SSD | |
CN109886946A (en) | Early senile maculopathy weakening supervision classification method based on deep learning | |
CN108765409A (en) | A kind of screening technique of the candidate nodule based on CT images | |
CN109215035A (en) | A kind of brain MRI hippocampus three-dimensional dividing method based on deep learning | |
CN110377659A (en) | A kind of intelligence chart recommender system and method |
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 |
Application publication date: 20181012 |
|
RJ01 | Rejection of invention patent application after publication |