CN108960281A - A kind of melanoma classification method based on nonrandom obfuscated data enhancement method - Google Patents
A kind of melanoma classification method based on nonrandom obfuscated data enhancement method Download PDFInfo
- Publication number
- CN108960281A CN108960281A CN201810505394.4A CN201810505394A CN108960281A CN 108960281 A CN108960281 A CN 108960281A CN 201810505394 A CN201810505394 A CN 201810505394A CN 108960281 A CN108960281 A CN 108960281A
- Authority
- CN
- China
- Prior art keywords
- training
- nonrandom
- data
- melanoma
- enhancing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
Abstract
A kind of melanoma classification method based on nonrandom obfuscated data enhancement method, comprising the following steps: the skin lens image in step 1. training set original first carries out data enhancing, and enhancement method is nonrandom cover；Initial data is carried out mixing to be then placed in network being trained by step 2. with the different data for covering position respectively, 4 relatively good models of effect are finally selected according to performance of the model on verifying collection, then then the training sample enhancing sample and initial data of this 4 models put together as final melanoma disaggregated model is trained to obtain melanoma disaggregated model；Step 3. training pattern；The feature extracted in step 3 is put into softmax classifier by step 4. classifies.The present invention provides a kind of melanoma classification method of high reliablity, can effectively classify to patient skin microscopy image, achievees the effect that assist last diagnostic.
Description
Technical field
The invention belongs to computer vision fields, construct one for skin based on the convolutional neural networks in deep learning
The disaggregated model of skin cancer melanoma.
Background technique
Melanoma is a kind of malignant cutaneous cancer as caused by melanocyte.Lead to death because of cutaneum carcinoma all
In case, melanoma accounts for about 75%.In cutaneum carcinoma disease, melanoma is most fatal cutaneum carcinoma lesion.With calculating
The rapid development of machine technology and artificial intelligence, depth learning technology obtain extensively in numerous areas, the especially fields such as computer vision
General use becomes artificial intelligence field research field the most active.Wherein image classification as one piece in visual range extremely
Influential project direction, the favor by vast deep learning researchers.This is carried out based on depth learning technology deep
Enter research, will have very important theoretical value and preferable application development prospect.For skin disease and maligna element
The research of tumor is beneficial to doctors' more preferably diagnosis conditions of patients.
Summary of the invention
In order to overcome the lower deficiency of reliability of existing melanoma classification method, the present invention provides a kind of high reliablity
Melanoma classification method, can effectively classify to patient skin microscopy image, reach auxiliary last diagnostic
Effect.
The technical solution adopted by the present invention to solve the technical problems is:
A kind of melanoma classification method based on nonrandom obfuscated data enhancement method, comprising the following steps:
The nonrandom obfuscated data enhancing of step 1.: the skin lens image in training set original first carries out data enhancing, increases
Strong mode is nonrandom cover；
Step 2. selects the model to do very well: using the data enhancement method of nonrandom cover, first distinguishing initial data
It carries out mixing to be then placed in network being trained with the different data for covering position, finally the table according to model on verifying collection
4 relatively good models of effect are now selected, then the enhancing sample of this 4 models and initial data are put together as most
Then the training sample of whole melanoma disaggregated model is trained to obtain melanoma disaggregated model；
Step 3. training pattern: training data is put into the network of 50 layers of depth convolution residual error and carries out feature extraction；
In training process, network parameter is initialized using the model of 50 layers of pre-training on ImageNet data set of residual error,
It is finely adjusted network parameter on this basis；
The feature extracted in step 3 is put into softmax classifier by step 4. classifies.
Further, in the step 1, nonrandom obfuscated data enhance the step of it is as follows:
1.1 successively take picture img from training set, and obtain the wide img_w and high img_h of img；
The wide w=img_w/3, high h=img_h/3 in region are covered in 1.2 initialization；
1.3 determine that the top left co-ordinate for covering region is respectively (x1, y1)=(0,0), (x2, y2)=(w, 0), (x3, y3)
=(2w, 0), (x4, y4)=(0, h), (x5, y5)=(w, h), (x6, y6)=(2w, h), (x7, y7)=(0,2h), (x8, y8)=
(w, 2h), (x9, y9)=(2w, 2h)；
1.4 determine that covering region is Mi=(xi, yi, xi+ w, yi+ h), wherein i=1,2,3 ... 9；
1.5 will cover region MiPixel value be set to 0；
1.6 save enhanced picture.
Further, in the step 2, the process of 50 layer network of convolution residual error training is as follows:
2.1: every original image I can generate 9 new pictures after nonrandom cover enhancing, respectively correspond covered area
Domain is the picture of M1, M2, M3, M4, M5, M6, M7, M8, M9, then by original image cover regions different with this nine kinds respectively
Picture is combined into 9 training sets, for example the I_M1 of training set 1=original image I+ enhancing trains 9 models；
2.2: being verified for training 9 models come in step 2.1 using the data of verifying collection, assess mould
4 models to behave oneself best are selected in performance of the type on AC, AP, AUC, SE, SP；
2.3: by original non-reinforced sample and four models using to enhancing sample be grouped together, composition one
A new training set；
2.4: new training set is put into 50 layers of convolution residual error of network and is trained, the last layer in training process
Output be 2, while using when the initialization of network residual error parameter in ImageNet pre-training as initiation parameter,
Then the input by the output of the last layer as softmax classifier.
Beneficial effects of the present invention are mainly manifested in: effectively being classified to patient skin microscopy image, reached auxiliary
Help the effect of last diagnostic.
Detailed description of the invention
Fig. 1 is the nonrandom schematic diagram for covering enhancement method in the present invention.
The schematic network structure that Fig. 2 is 50 layers of convolution residual error used in the present invention.
Fig. 3 is the schematic diagram of entire disaggregated model in the present invention.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.
Referring to Fig.1~Fig. 3, a kind of melanoma classification method based on nonrandom obfuscated data enhancement method, including with
Lower step:
The nonrandom obfuscated data enhancing of step 1.: the skin lens image in training set original first carries out data enhancing, increases
Strong mode is nonrandom cover；
Step 2. selects the model to do very well: using the data enhancement method of nonrandom cover, first distinguishing initial data
It carries out mixing to be then placed in network being trained with the different data for covering position, finally the table according to model on verifying collection
4 relatively good models of effect are now selected, then the enhancing sample of this 4 models and initial data are put together as most
Then the training sample of whole melanoma disaggregated model is trained to obtain melanoma disaggregated model；
Step 3. training pattern: training data is put into the network of 50 layers of depth convolution residual error and carries out feature extraction；
In training process, network parameter is initialized using the model of 50 layers of pre-training on ImageNet data set of residual error,
It is finely adjusted network parameter on this basis；
The feature extracted in step 3 is put into softmax classifier by step 4. classifies.
Further, in the step 1, nonrandom obfuscated data enhance the step of it is as follows:
1.1 successively take picture img from training set, and obtain the wide img_w and high img_h of img；
The wide w=img_w/3, high h=img_h/3 in region are covered in 1.2 initialization；
1.3 determine that the top left co-ordinate for covering region is respectively (x1, y1)=(0,0), (x2, y2)=(w, 0), (x3, y3)
=(2w, 0), (x4, y4)=(0, h), (x5, y5)=(w, h), (x6, y6)=(2w, h), (x7, y7)=(0,2h), (x8, y8)=
(w, 2h), (x9, y9)=(2w, 2h)；
1.4 determine that covering region is Mi=(xi, yi, xi+ w, yi+ h), wherein i=1,2,3 ... 9；
1.5 will cover region MiPixel value be set to 0；
1.7 save enhanced picture.
Further, in the step 2, the process of 50 layer network of convolution residual error training is as follows:
2.1: every original image I can generate 9 new pictures after nonrandom cover enhancing, respectively correspond covered area
Domain is the picture of M1, M2, M3, M4, M5, M6, M7, M8, M9, then by original image cover regions different with this nine kinds respectively
Picture is combined into 9 training sets, for example the I_M1 of training set 1=original image I+ enhancing trains 9 models；
2.2: being verified for training 9 models come in step 2.1 using the data of verifying collection, assess mould
4 models to behave oneself best are selected in performance of the type on AC, AP, AUC, SE, SP；
2.3: by original non-reinforced sample and four models using to enhancing sample be grouped together, composition one
A new training set；
2.4: new training set is put into 50 layers of convolution residual error of network and is trained, the last layer in training process
Output be 2, while using when the initialization of network residual error parameter in ImageNet pre-training as initiation parameter,
Then the input by the output of the last layer as softmax classifier.
In the present embodiment, for, there are two class skin lens image samples, one kind is melanoma, another in original training set
Class is mixed image, that is, non-black melanoma of seborrheic keratosis and mole.
The melanoma classification method based on nonrandom obfuscated data enhancement method of the present embodiment, comprising the following steps:
Step 1 enhances nonrandom obfuscated data, we successively take out sample from original training set, takes one every time
Enhanced, original image backup saves.
1.1: sample ISIC_0000074.jpg is got from training set, first obtains the wide img_w=1022 of this image,
High img_w=767；
1.2: initialization determines the wide w=1022/3=340 (rounding), h=767/3=255 for covering region；
1.3: determination covers the top left co-ordinate in region respectively (0,0), (340,0), (680,0), (0,255), (340,
255), (680,255), (0,510), (340,510), (680,510)；
1.4: determining that cover region be respectively M1 be (0,0,340,255) M2 is (340,0,680,255), M3 for (680,
0,1020,255), M4 is (0,255,340,510), and M5 is (340,255,680,510), and M6 is (680,255,1020,510),
M7 is (0,510,340,765), and M8 is (340,510,680,765), and M9 is (680,510,1020,765)；
1.5: by cover region M1, M2, M3, M4, M5, M6, M7, M8, M9 pixel value this be 0；
1.6: save cover after sample be respectively designated as Mask1_0000074.jpg, Mask2_0000074.jpg,
Mask3_0000074.jpg、Mask4_0000074.jpg、Mask5_0000074.jpg、Mask6_0000074.jpg、
Mask7_0000074.jpg,Mask8_0000074.jpg,Mask9_0000074.jpg；
Last enhanced sample is as shown in Figure 1.
All sample standard deviations in training set are carried out random cover and expanded by step 2, then by original sample respectively and
The data that difference covers positions carry out mixing composition training set and are then placed in 50 layer network of residual error and are trained, when training still
Use pre-training model as parameter initialization, solicit orders picture for example by ISIC0000074.jpg respectively and Mask1_
0000074.jpg、Mask2_0000074.jpg、Mask3_0000074.jpg、Mask4_0000074.jpg、Mask5_
0000074.jpg、Mask6_0000074.jpg、Mask7_0000074.jpg、Mask8_0000074.jpg、Mask9_
0000074.jpg separately constitutes training set 1-9, and then training set 1-9 is put into 50 layer network of residual error respectively and is trained
Optimization, finally selects the training set to do very well according to the performance of each training set, its training data is put together and forms mould
The final training set of type.
The data of determining final training set are put into the network of 50 layers of depth convolution residual error and carry out feature by step 3.
It extracts.In training process, network parameter is carried out just using the model of 50 layers of pre-training on ImageNet data set of residual error
Beginningization is finely adjusted network parameter on this basis.
The feature extracted in 3 is put into softmax classifier by step 4. classifies.
50 layer network structure of depth convolution residual error is as shown in Fig. 2, whole model flow is as shown in Figure 3.Model it is final
Result referring to table 1:
Table 1.
Claims (3)
1. a kind of melanoma classification method based on nonrandom obfuscated data enhancement method, which is characterized in that the method packet
Include following steps:
The nonrandom obfuscated data enhancing of step 1.: the skin lens image in training set original first carries out data enhancing, enhancing side
Formula is nonrandom cover；
Step 2. selects the model that does very well: using the data enhancement method of nonrandom cover, first respectively and not by initial data
It carries out mixing to be then placed in network being trained with the data for covering position, finally be selected according to performance of the model on verifying collection
4 relatively good models of effect are selected out, then the enhancing sample of this 4 models and initial data are put together as final
Then the training sample of melanoma disaggregated model is trained to obtain melanoma disaggregated model；
Step 3. training pattern: training data is put into the network of 50 layers of depth convolution residual error and carries out feature extraction；Training
In the process, network parameter is initialized using the model of 50 layers of pre-training on ImageNet data set of residual error, herein
On the basis of be finely adjusted network parameter；
The feature extracted in step 3 is put into softmax classifier by step 4. classifies.
2. a kind of melanoma classification method based on nonrandom obfuscated data enhancement method as described in claim 1, special
Sign is, in the step 1, nonrandom obfuscated data enhance the step of it is as follows:
1.1 successively take picture img from training set, and obtain the wide img_w and high img_h of img；
The wide w=img_w/3, high h=img_h/3 in region are covered in 1.2 initialization；
1.3 determine that the top left co-ordinate for covering region is respectively (x1, y1)=(0,0), (x2, y2)=(w, 0), (x3, y3)=
(2w, 0), (x4, y4)=(0, h), (x5, y5)=(w, h), (x6, y6)=(2w, h), (x7, y7)=(0,2h), (x8, y8)=
(w, 2h), (x9, y9)=(2w, 2h)；
1.4 determine that covering region is Mi=(xi, yi, xi+ w, yi+ h), wherein i=1,2,3 ... 9；
1.5 will cover region MiPixel value be set to 0；
1.6 save enhanced picture.
3. a kind of melanoma classification method based on nonrandom obfuscated data enhancement method as claimed in claim 1 or 2,
It is characterized in that, in the step 2, the process of 50 layer network of convolution residual error training is as follows:
2.1: every original image I can generate 9 new pictures after nonrandom cover enhancing, respectively correspond cover region and be
The picture of M1, M2, M3, M4, M5, M6, M7, M8, M9, then by original image respectively it is different with this nine kinds cover regions pictures
9 training sets are combined into, for example the I_M1 of training set 1=original image I+ enhancing trains 9 models；
2.2: being verified for training 9 models come in step 2.1 using the data of verifying collection, assessment models exist
4 models to behave oneself best are selected in performance on AC, AP, AUC, SE, SP；
2.3: by original non-reinforced sample and four models using to enhancing sample be grouped together, form one it is new
Training set；
2.4: new training set being put into 50 layers of convolution residual error of network and be trained, the last layer is defeated in training process
Out it is 2, while uses in the parameter of ImageNet pre-training as initiation parameter, then when the initialization of network residual error
Input by the output of the last layer as softmax classifier.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810505394.4A CN108960281B (en) | 2018-05-24 | 2018-05-24 | Melanoma classification model establishing method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810505394.4A CN108960281B (en) | 2018-05-24 | 2018-05-24 | Melanoma classification model establishing method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108960281A true CN108960281A (en) | 2018-12-07 |
| CN108960281B CN108960281B (en) | 2020-05-05 |
Family
ID=64499599
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810505394.4A Active CN108960281B (en) | 2018-05-24 | 2018-05-24 | Melanoma classification model establishing method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108960281B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110379414A (en) * | 2019-07-22 | 2019-10-25 | 出门问问(苏州)信息科技有限公司 | Acoustic model enhances training method, device, readable storage medium storing program for executing and calculates equipment |
| CN111640125A (en) * | 2020-05-29 | 2020-09-08 | 广西大学 | Mask R-CNN-based aerial photograph building detection and segmentation method and device |
| CN113139932A (en) * | 2021-03-23 | 2021-07-20 | 广东省科学院智能制造研究所 | Deep learning defect image identification method and system based on ensemble learning |
| CN113139578A (en) * | 2021-03-23 | 2021-07-20 | 广东省科学院智能制造研究所 | Deep learning image classification method and system based on optimal training set |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101604376A (en) * | 2008-10-11 | 2009-12-16 | 大连大学 | Face identification method based on the HMM-SVM mixture model |
| US7650024B2 (en) * | 2005-06-07 | 2010-01-19 | George Mason Intellectual Properties, Inc. | Dissipative functional microarrays for classification |
| US20120087575A1 (en) * | 2007-06-19 | 2012-04-12 | Microsoft Corporation | Recognizing hand poses and/or object classes |
| US8315465B1 (en) * | 2009-01-12 | 2012-11-20 | Google Inc. | Effective feature classification in images |
| US20150235109A1 (en) * | 2010-03-19 | 2015-08-20 | Canon Kabushiki Kaisha | Learning method and apparatus for pattern recognition |
| CN107463966A (en) * | 2017-08-17 | 2017-12-12 | 电子科技大学 | Radar range profile's target identification method based on dual-depth neutral net |
| CN107506774A (en) * | 2017-10-09 | 2017-12-22 | 深圳市唯特视科技有限公司 | A kind of segmentation layered perception neural networks method based on local attention mask |
| CN107766894A (en) * | 2017-11-03 | 2018-03-06 | 吉林大学 | Remote sensing images spatial term method based on notice mechanism and deep learning |
| CN107909059A (en) * | 2017-11-30 | 2018-04-13 | 中南大学 | It is a kind of towards cooperateing with complicated City scenarios the traffic mark board of bionical vision to detect and recognition methods |
| CN107944379A (en) * | 2017-11-20 | 2018-04-20 | 中国科学院自动化研究所 | White of the eye image super-resolution rebuilding and image enchancing method based on deep learning |
-
2018
- 2018-05-24 CN CN201810505394.4A patent/CN108960281B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7650024B2 (en) * | 2005-06-07 | 2010-01-19 | George Mason Intellectual Properties, Inc. | Dissipative functional microarrays for classification |
| US20120087575A1 (en) * | 2007-06-19 | 2012-04-12 | Microsoft Corporation | Recognizing hand poses and/or object classes |
| CN101604376A (en) * | 2008-10-11 | 2009-12-16 | 大连大学 | Face identification method based on the HMM-SVM mixture model |
| US8315465B1 (en) * | 2009-01-12 | 2012-11-20 | Google Inc. | Effective feature classification in images |
| US20150235109A1 (en) * | 2010-03-19 | 2015-08-20 | Canon Kabushiki Kaisha | Learning method and apparatus for pattern recognition |
| CN107463966A (en) * | 2017-08-17 | 2017-12-12 | 电子科技大学 | Radar range profile's target identification method based on dual-depth neutral net |
| CN107506774A (en) * | 2017-10-09 | 2017-12-22 | 深圳市唯特视科技有限公司 | A kind of segmentation layered perception neural networks method based on local attention mask |
| CN107766894A (en) * | 2017-11-03 | 2018-03-06 | 吉林大学 | Remote sensing images spatial term method based on notice mechanism and deep learning |
| CN107944379A (en) * | 2017-11-20 | 2018-04-20 | 中国科学院自动化研究所 | White of the eye image super-resolution rebuilding and image enchancing method based on deep learning |
| CN107909059A (en) * | 2017-11-30 | 2018-04-13 | 中南大学 | It is a kind of towards cooperateing with complicated City scenarios the traffic mark board of bionical vision to detect and recognition methods |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110379414A (en) * | 2019-07-22 | 2019-10-25 | 出门问问(苏州)信息科技有限公司 | Acoustic model enhances training method, device, readable storage medium storing program for executing and calculates equipment |
| CN110379414B (en) * | 2019-07-22 | 2021-12-03 | 出门问问(苏州)信息科技有限公司 | Acoustic model enhancement training method and device, readable storage medium and computing equipment |
| CN111640125A (en) * | 2020-05-29 | 2020-09-08 | 广西大学 | Mask R-CNN-based aerial photograph building detection and segmentation method and device |
| CN111640125B (en) * | 2020-05-29 | 2022-11-18 | 广西大学 | Aerial photography graph building detection and segmentation method and device based on Mask R-CNN |
| CN113139932A (en) * | 2021-03-23 | 2021-07-20 | 广东省科学院智能制造研究所 | Deep learning defect image identification method and system based on ensemble learning |
| CN113139578A (en) * | 2021-03-23 | 2021-07-20 | 广东省科学院智能制造研究所 | Deep learning image classification method and system based on optimal training set |
| CN113139932B (en) * | 2021-03-23 | 2022-12-20 | 广东省科学院智能制造研究所 | Deep learning defect image identification method and system based on ensemble learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108960281B (en) | 2020-05-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108960281A (en) | A kind of melanoma classification method based on nonrandom obfuscated data enhancement method | |
| CN104182772B (en) | A kind of gesture identification method based on deep learning | |
| CN104834922B (en) | Gesture identification method based on hybrid neural networks | |
| CN106599854A (en) | Method for automatically recognizing face expressions based on multi-characteristic fusion | |
| WO2017084204A1 (en) | Method and system for tracking human body skeleton point in two-dimensional video stream | |
| Wang et al. | Background-driven salient object detection | |
| CN105205449B (en) | Sign Language Recognition Method based on deep learning | |
| CN112395442B (en) | Automatic identification and content filtering method for popular pictures on mobile internet | |
| CN106204449A (en) | A kind of single image super resolution ratio reconstruction method based on symmetrical degree of depth network | |
| CN103854016B (en) | Jointly there is human body behavior classifying identification method and the system of feature based on directivity | |
| CN102855468A (en) | Single sample face recognition method in photo recognition | |
| CN109766822B (en) | Gesture recognition method and system based on neural network | |
| CN101615245A (en) | Expression recognition method based on AVR and enhancing LBP | |
| CN108053398A (en) | A kind of melanoma automatic testing method of semi-supervised feature learning | |
| CN109902585A (en) | A kind of three modality fusion recognition methods of finger based on graph model | |
| CN107808376A (en) | A kind of detection method of raising one's hand based on deep learning | |
| CN112906550B (en) | Static gesture recognition method based on watershed transformation | |
| WO2022127494A1 (en) | Pose recognition model training method and apparatus, pose recognition method, and terminal device | |
| CN103679136A (en) | Hand back vein identity recognition method based on combination of local macroscopic features and microscopic features | |
| CN109829924A (en) | A kind of image quality evaluating method based on body feature analysis | |
| CN107748798A (en) | A kind of hand-drawing image search method based on multilayer visual expression and depth network | |
| CN104036468A (en) | Super-resolution reconstruction method for single-frame images on basis of pre-amplification non-negative neighbor embedding | |
| CN107229949A (en) | A kind of complex illumination hypograph feature extracting method | |
| Xiao et al. | Pedestrian object detection with fusion of visual attention mechanism and semantic computation | |
| CN106203448A (en) | A kind of scene classification method based on Nonlinear Scale Space Theory |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |