CN109740626A - The detection method of cancerous area in breast cancer pathological section based on deep learning - Google Patents
The detection method of cancerous area in breast cancer pathological section based on deep learning Download PDFInfo
- Publication number
- CN109740626A CN109740626A CN201811407164.0A CN201811407164A CN109740626A CN 109740626 A CN109740626 A CN 109740626A CN 201811407164 A CN201811407164 A CN 201811407164A CN 109740626 A CN109740626 A CN 109740626A
- Authority
- CN
- China
- Prior art keywords
- data
- breast cancer
- sample
- training
- pathological section
- 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
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The detection method of cancerous area in the invention discloses a kind of breast cancer pathological section based on deep learning;Present invention combination breast cancer pathological section does not have the characteristics of fixed-direction, systematically enhances technology using reasonable data, and the data including geometric transformations such as random cropping, rotation, left and right overturnings enhance technology.The data enhancing technology of the colour switchings such as random brightness, sharpening is also used simultaneously.Real-time perfoming data enhance when training, are capable of increasing the diversity of data set, expand training sample set, effectively improve the generalization ability of classifier.Finally with real-time oversampler method, each categorical data to solve the problems, such as data set is unbalanced;The present invention feature big for each classification breast cancer area size difference in data set slice, systematically used the equiprobable real-time oversampler method of classification and a variety of data enhancement methods reasonable on medical condition image, solve the problems, such as data nonbalance and it is benign, carcinoma in situ categorical data amount is seldom.
Description
Technical field
The invention belongs to artificial intelligence field, a kind of be related in the breast cancer pathological section based on deep learning cancer
Method for detecting area.
Background technique
It is on breast duct that breast cancer (breast cancer), which is a kind of cancer adenocarcinoma developed from breast tissue,
Chrotoplast is abnormal hyperplasia, the malignant tumour occurred more than self-repairing capability.It is a kind of serious harm women body and mind
The common malignant disease of health, not only jeopardizes patient vitals, moreover it is possible to cause the damage of woman organ, it has also become 50 years old or more woman
One of the main reason for female's malignant tumour is lethal.It clinically shows as micro-calcification clusters and breast lump, and early stage is asymptomatic,
With incidence of occult, more after it is poor the features such as.
Sentinel lymph node biopsy (sentinel lymph node biopsy, SLNB) is a kind of safe, accurate hand
Art mode, gradually substituting axillary Pulmonary resections become the standard art formula that early-stage breast cancer is treated, are that assessment cancer cell is
The no goldstandard along lymphatic vessel transfer diffusion and lymph node staging.The wound of the technology is small, reduces the edema of the upper extremity of postoperative patient
Disease incidence reduces complication probability of happening, hence it is evident that improves the postoperative quality of life of patient.The tissue collected in biopsy procedure is logical
Common hematoxylin and eosin (H&E) dyeing, are then analyzed by expert.Virologist by sentinel lymph node biopsy come
The microstructure and element for assessing tissue, are classified as normal tissue, non-malignant (benign) and malignant change, and carry out prognosis
Assessment.In the process, the relevant range of entire glass slide tissue scanning is assessed.Dyeing enhancing nucleus (purple) and cell
Matter (pink) and other interested structures.When analyzing stained tissue, virologist analyzes entire institutional framework,
And nucleus tissue, density and variability.For example, the tissue with infiltrating carcinoma shows the deformation of structure and higher
Cuclear density and variability, and in the normal tissue, structure is maintained and nucleus tissue is good.
According to American Cancer Society (American Cancer Society, ACS) studies have shown that the breast cancer of early stage is in cancer
In the non-diffused situation of cell, 5 annual survival rates are up to 98%.Early detection and diagnosis are to reduce breast cancer incidence and death
The most effective approach of rate.Early prevention, early discovery, early diagnosis, early treatment, are the key that breast cancer preventions.In breast cancer diagnosis
In, common method has the diagnosis of palpation formula, histodiagnosis, cytodiagnosis, imaging diagnosis etc..These diagnostic method ratios
It is more complex, it is typically necessary is completed by manual hand manipulation in actual use, and these artificial methods for extracting feature are easily led
The loss of Partial Feature information is caused, so that diagnosis recognition performance is not satisfactory, all Shortcomings in accuracy rate, omission factor.
To solve shortcoming and defect in the prior art, the present invention proposes a kind of breast cancer pathology based on deep learning
The detection method of cancerous area in slice combines deep learning with pathological section image recognition, assists diagnosis mammary gland
Cancer.
Cancerous area detection method of the present invention includes the following steps:
Step 1: data prediction, extracts the tissue part in data set slice using gray threshold method, after being used for
It is continuous to cut reasonable object detection area;The data set is breast cancer pathological section;
Step 2: original breast cancer pathological section is divided into training sample, verifying sample and test sample, reuse
Grid clipping method, cuts the picture sample of same size from the tissue regions of slice, and reduces the size of picture for former ruler
Very little 1/3;
Step 3: training sample and verifying sample are individually placed in the data file of classification, in the training process, needle
To the positive and negative imbalanced training sets problem in data acquisition system, solved using the real-time oversampler method that equiprobability samples, for number
According to the few problem of some classification sample sizes in set, solved using random cropping, rotation, overturning and color Enhancement Method
Certainly;
It is trained Step 4: the sample that sampling comes out is sequentially placed into convolutional neural networks, convolutional neural networks are adopted
With the Inception-Resnet-v2 network for using Inception and Resnet principle design structure;
Step 5: the top-down transmission of error is finely adjusted each layer of parameter by the data training of tape label,
Visualization real-time display is carried out by the training result of network and in the operation accuracy rate that verifying collection closes, is closed according to verifying collection
Exact value and penalty values curve carry out network fine tuning;
Step 6: after setting runs multiple algebra, until the accuracy rate of verifying collection reaches highest;Save all training process
In parameter and model, test sample is transported in network after training process, carries out predicting to classify without label image,
Prediction test image simultaneously calculates the corresponding classification possibility of representative;Finally output be normal tissue, benign, carcinoma in situ and infiltrating carcinoma,
The image for forecast sample result being synthesized a slice size is exactly last slice cancerous area testing result.
Beneficial effects of the present invention:
1) automation breast cancer pathological section cancerous area detection system proposed by the present invention can assist pathologist to examine
Ablactation gland cancer, mitigates surgeon stress, reduces human error, has given full play to the advantage of the self-teaching of deep learning, utilized depth
Layer network extracts enhanced advanced features, the experimental results showed that the feature that we learn has higher distinction;
2) differentiation of the characteristics of system is herein in connection with breast cancer diagnosis, carcinoma in situ and infiltrating carcinoma needs more about tissue
Information, the method and Inception structure of scaling have been used according to the feature, the information of abundant integrated tissue improves point
The accuracy of class;
3) feature big for each classification breast cancer area size difference in data set slice, systematically uses
The equiprobable real-time oversampler method of classification and a variety of data enhancement methods reasonable on medical condition image, solve number
According to the problem uneven and benign, carcinoma in situ categorical data amount is seldom.
Detailed description of the invention
Fig. 1 is the basic flow chart of the breast cancer cancerous area detection method the present invention is based on deep learning.
Fig. 2 is partial depth learning network structure chart;
Fig. 3 (A) is original pathological section;
Fig. 3 (B) is the cancerous area of doctor's mark;
Fig. 3 (C) is using cancerous area annotation results of the invention.
Specific embodiment
The present invention completes 4 classification task using an Inception-Resnet depth convolutional neural networks.The net
Network first carries out detection convolution sum pond method with small size convolution kernel to reduce calculation amount.Followed by a convolution group and more
The series connection of a residual error convolution group for realizing multiple dimensioned feature extraction, while can reduce optimization hardly possible using residual error method
Degree accelerates training speed, and most important effect is that network can be allowed to obtain enough gain effects from depth.Network is most
After be global average pond, not only can also play the effect of regularization for reducing calculation amount, improve model
Accuracy rate in actual use.
As shown in Figure 1, the specific embodiment of whole process of the present invention illustrated below is as follows:
1, it is analyzed using International image and identification meeting (ICIAR) database 2018 to breast cancer pathological section cancer
Region detection challenges disclosed 10 training set breast cancer pathology sectioning images and 20 slices as test set marked
Image, this data are marked by several virologists collected by the medical research leading in the world.Training method is band mark
The data of label carry out the study for having supervision;
2, tissue regions therein are extracted by gray threshold method to pathological section image, first the color of image
Space is transformed into HSV space from RGB, and fixed threshold value is arranged, extracts tissue regions;
3, the sample needed is gone out by grid clipping on the tissue regions of breast cancer pathological section, while is directed to sample
Sample image is reduced 3 times, is contained foot by the characteristics of image is small, cannot accommodate enough organizational informations using Zoom method
Enough information carries out the sort operation in next network model;
4, in input layer, for the benign and small feature of carcinoma in situ sample, combine breast cancer pathological section not solid first
The characteristics of determining direction systematically enhances the geometry such as technology, including random cropping, rotation, left and right overturning using reasonable data and becomes
The data enhancing technology changed.The data enhancing technology of the colour switchings such as random brightness, sharpening is also used simultaneously.Trained
When real-time perfoming data enhance, be capable of increasing the diversity of data set, expand training sample set, effectively improve classifier
Generalization ability.The equiprobable real-time oversampler method of classification is finally used, it is unbalanced come each categorical data for solving data set
Problem;
5, the adjustment of weight parameter and offset, momentum in analog physical, before accumulation are carried out using momentum optimization method
Momentum substitute real gradient.It when decline initial stage, is updated using last parameter, descent direction is consistent, is able to carry out very
Good acceleration when declining the middle and later periods, when local minimum vibrates back and forth, so that the amplitude of update increases, jumps out trap,
It when gradient change direction, can decline in related direction accelerating gradient, inhibit oscillation, to accelerate to restrain;
6, by the weight parameter matrix and offset in trained each layer, each layer being accordingly assigned in network, then
The network has the function of the feature extraction of breast cancer and identification.Prediction classification finally is carried out to the sample of test set, by result
It is corresponding to merge, shown in final result process such as Fig. 3 (A)~(C).
As shown in Fig. 2, the neural network that the present invention constructs in implementation process, mainly by convolution group, residual error module and point
Three parts of class device are constituted, and have used random inactivation come the phenomenon that avoiding over-fitting in a model.But during model training according to
So there are over-fittings.In order to solve this problem, we have used data enhancing on image data set, such as random cropping,
Flip horizontal, Random-Rotation, random brightness transformation and random sharpening etc..
In order to guarantee have in samples pictures enough in the case where not increasing samples pictures so as to cause calculation amount increase
Organizational information for classifying, the method use Zoom method, cut out 3 times of picture bigger than input sample picture, then
It reduces 3 times of sizes as input sample picture and constructs data set.
Inception-Resnet network used herein in breast cancer is instructed on ImageNet
Practice, which contains about 1,000,000 natural images and 1000 label/classification.Since our task is 4 classification,
Therefore network structure is had adjusted, the classification layer output layer neuron number of script is changed to 4, to meet this task 4
The requirement of classification.
The results show this method has higher specificity and sensitivity, illustrates data enhancing and scaling side
Method and the network have certain promotion effect for model identification.
Claims (1)
1. the detection method of cancerous area in the breast cancer pathological section based on deep learning, which is characterized in that this method is specific
The following steps are included:
Step 1: data prediction, extracts the tissue part in data set slice using gray threshold method, is used for subsequent cutting
Reasonable object detection area;The data set is breast cancer pathological section;
Step 2: original breast cancer pathological section is divided into training sample, verifying sample and test sample, grid sanction is reused
Shear method, cuts the picture sample of same size from the tissue regions of slice, and the size for reducing picture is the 1/ of full size
3;
Step 3: training sample and verifying sample are individually placed in the data file of classification, in the training process, for data
Positive and negative imbalanced training sets problem in set is solved using the real-time oversampler method that equiprobability samples, is closed for data sets
In the few problem of some classification sample sizes, solved using random cropping, rotation, overturning and color Enhancement Method;
It is trained Step 4: the sample that sampling comes out is sequentially placed into convolutional neural networks, convolutional neural networks, which use, to be made
With the Inception-Resnet-v2 network of Inception and Resnet principle design structure;
Step 5: by the data training of tape label, the top-down transmission of error is finely adjusted each layer of parameter, by net
The training result of network and visualization real-time display is carried out in the operation accuracy rate closed of verifying collection, it is accurate to close according to verifying collection
Value and penalty values curve carry out network fine tuning;
Step 6: after setting runs multiple algebra, until the accuracy rate of verifying collection reaches highest;Save the ginseng in all training process
Test sample is transported in network after training process, carries out predicting to classify without label image by several and model, prediction test
Image simultaneously calculates the corresponding classification possibility of representative;Finally output is normal tissue, benign, carcinoma in situ and infiltrating carcinoma, pre- test sample
The image that this result synthesizes a slice size is exactly last slice cancerous area testing result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811407164.0A CN109740626A (en) | 2018-11-23 | 2018-11-23 | The detection method of cancerous area in breast cancer pathological section based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811407164.0A CN109740626A (en) | 2018-11-23 | 2018-11-23 | The detection method of cancerous area in breast cancer pathological section based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109740626A true CN109740626A (en) | 2019-05-10 |
Family
ID=66358119
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811407164.0A Pending CN109740626A (en) | 2018-11-23 | 2018-11-23 | The detection method of cancerous area in breast cancer pathological section based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109740626A (en) |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110276802A (en) * | 2019-06-26 | 2019-09-24 | 湖南大学 | Illness tissue localization method, device and equipment in medical image |
CN110288542A (en) * | 2019-06-18 | 2019-09-27 | 福州数据技术研究院有限公司 | A kind of liver's pathological image sample Enhancement Method based on stochastic transformation |
CN110443268A (en) * | 2019-05-30 | 2019-11-12 | 杭州电子科技大学 | A kind of benign pernicious classification method of liver cancer CT image based on deep learning |
CN110633758A (en) * | 2019-09-20 | 2019-12-31 | 四川长虹电器股份有限公司 | Method for detecting and locating cancer region aiming at small sample or sample unbalance |
CN110738637A (en) * | 2019-09-19 | 2020-01-31 | 华中科技大学 | Automatic classification method and system for breast cancer pathological sections |
CN111159150A (en) * | 2019-12-19 | 2020-05-15 | 北京文安智能技术股份有限公司 | Data expansion method and device |
CN111209916A (en) * | 2019-12-31 | 2020-05-29 | 中国科学技术大学 | Focus identification method and system and identification equipment |
CN111680553A (en) * | 2020-04-29 | 2020-09-18 | 北京联合大学 | Pathological image identification method and system based on depth separable convolution |
CN111784593A (en) * | 2020-06-04 | 2020-10-16 | 广东省智能制造研究所 | Lung nodule CT image data enhancement method and system for deep learning |
CN111968147A (en) * | 2020-08-06 | 2020-11-20 | 电子科技大学 | Breast cancer pathological image comprehensive analysis system based on key point detection |
CN112102247A (en) * | 2020-08-18 | 2020-12-18 | 广州金域医学检验中心有限公司 | Pathological section quality evaluation method based on machine learning and related equipment |
CN112710996A (en) * | 2020-12-08 | 2021-04-27 | 中国人民解放军海军航空大学 | Data set expansion method and system for radar micro-motion target identification |
CN112950614A (en) * | 2021-03-22 | 2021-06-11 | 山西三友和智慧信息技术股份有限公司 | Breast cancer detection method based on multi-scale cavity convolution |
CN114897292A (en) * | 2022-03-28 | 2022-08-12 | 南京林业大学 | Fault diagnosis method and system based on multi-scale residual convolution variational network |
CN111784593B (en) * | 2020-06-04 | 2024-07-26 | 广东省智能制造研究所 | Deep learning-oriented lung nodule CT image data enhancement method and system |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107028593A (en) * | 2017-04-14 | 2017-08-11 | 成都知识视觉科技有限公司 | A kind of aided detection method of breast ductal carcinoma in situ |
US20180075597A1 (en) * | 2016-09-09 | 2018-03-15 | Siemens Healthcare Gmbh | Tissue Characterization Based on Machine Learning in Medical Imaging |
CN108230311A (en) * | 2018-01-03 | 2018-06-29 | 四川大学 | A kind of breast cancer detection method and device |
CN108717554A (en) * | 2018-05-22 | 2018-10-30 | 复旦大学附属肿瘤医院 | A kind of thyroid tumors histopathologic slide image classification method and its device |
-
2018
- 2018-11-23 CN CN201811407164.0A patent/CN109740626A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180075597A1 (en) * | 2016-09-09 | 2018-03-15 | Siemens Healthcare Gmbh | Tissue Characterization Based on Machine Learning in Medical Imaging |
CN107028593A (en) * | 2017-04-14 | 2017-08-11 | 成都知识视觉科技有限公司 | A kind of aided detection method of breast ductal carcinoma in situ |
CN108230311A (en) * | 2018-01-03 | 2018-06-29 | 四川大学 | A kind of breast cancer detection method and device |
CN108717554A (en) * | 2018-05-22 | 2018-10-30 | 复旦大学附属肿瘤医院 | A kind of thyroid tumors histopathologic slide image classification method and its device |
Cited By (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110443268A (en) * | 2019-05-30 | 2019-11-12 | 杭州电子科技大学 | A kind of benign pernicious classification method of liver cancer CT image based on deep learning |
CN110443268B (en) * | 2019-05-30 | 2022-02-08 | 杭州电子科技大学 | Liver CT image benign and malignant classification method based on deep learning |
CN110288542A (en) * | 2019-06-18 | 2019-09-27 | 福州数据技术研究院有限公司 | A kind of liver's pathological image sample Enhancement Method based on stochastic transformation |
CN110276802A (en) * | 2019-06-26 | 2019-09-24 | 湖南大学 | Illness tissue localization method, device and equipment in medical image |
CN110738637A (en) * | 2019-09-19 | 2020-01-31 | 华中科技大学 | Automatic classification method and system for breast cancer pathological sections |
CN110738637B (en) * | 2019-09-19 | 2022-10-14 | 华中科技大学 | Automatic classification system for breast cancer pathological sections |
CN110633758A (en) * | 2019-09-20 | 2019-12-31 | 四川长虹电器股份有限公司 | Method for detecting and locating cancer region aiming at small sample or sample unbalance |
CN111159150A (en) * | 2019-12-19 | 2020-05-15 | 北京文安智能技术股份有限公司 | Data expansion method and device |
CN111209916A (en) * | 2019-12-31 | 2020-05-29 | 中国科学技术大学 | Focus identification method and system and identification equipment |
CN111209916B (en) * | 2019-12-31 | 2024-01-23 | 中国科学技术大学 | Focus identification method and system and identification equipment |
CN111680553A (en) * | 2020-04-29 | 2020-09-18 | 北京联合大学 | Pathological image identification method and system based on depth separable convolution |
CN111784593A (en) * | 2020-06-04 | 2020-10-16 | 广东省智能制造研究所 | Lung nodule CT image data enhancement method and system for deep learning |
CN111784593B (en) * | 2020-06-04 | 2024-07-26 | 广东省智能制造研究所 | Deep learning-oriented lung nodule CT image data enhancement method and system |
CN111968147A (en) * | 2020-08-06 | 2020-11-20 | 电子科技大学 | Breast cancer pathological image comprehensive analysis system based on key point detection |
CN111968147B (en) * | 2020-08-06 | 2022-03-15 | 电子科技大学 | Breast cancer pathological image comprehensive analysis system based on key point detection |
CN112102247A (en) * | 2020-08-18 | 2020-12-18 | 广州金域医学检验中心有限公司 | Pathological section quality evaluation method based on machine learning and related equipment |
CN112102247B (en) * | 2020-08-18 | 2024-05-14 | 广州金域医学检验中心有限公司 | Machine learning-based pathological section quality evaluation method and related equipment |
CN112710996A (en) * | 2020-12-08 | 2021-04-27 | 中国人民解放军海军航空大学 | Data set expansion method and system for radar micro-motion target identification |
CN112950614B (en) * | 2021-03-22 | 2023-01-10 | 山西三友和智慧信息技术股份有限公司 | Breast cancer pathological image classification method based on multi-scale cavity convolution |
CN112950614A (en) * | 2021-03-22 | 2021-06-11 | 山西三友和智慧信息技术股份有限公司 | Breast cancer detection method based on multi-scale cavity convolution |
CN114897292A (en) * | 2022-03-28 | 2022-08-12 | 南京林业大学 | Fault diagnosis method and system based on multi-scale residual convolution variational network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109740626A (en) | The detection method of cancerous area in breast cancer pathological section based on deep learning | |
Wang et al. | Using deep convolutional neural networks for multi-classification of thyroid tumor by histopathology: a large-scale pilot study | |
CN109492711B (en) | Malignant melanoma and non-malignant melanoma classification system based on deep learning | |
CN108717554A (en) | A kind of thyroid tumors histopathologic slide image classification method and its device | |
CN109903284A (en) | A kind of other method and system of HER2 immunohistochemistry image automatic judging | |
Pan et al. | Mitosis detection techniques in H&E stained breast cancer pathological images: A comprehensive review | |
AU2012228063A1 (en) | Histology analysis | |
CN110391015A (en) | A method of tumor immunity is quantified based on image group | |
US20170084021A1 (en) | Grading of Glandular Tissue Cancer by Detailed Image Analysis of Stained Tissue Slices | |
CN112263217B (en) | Improved convolutional neural network-based non-melanoma skin cancer pathological image lesion area detection method | |
CN112990214A (en) | Medical image feature recognition prediction model | |
CN113420793A (en) | Improved convolutional neural network ResNeSt 50-based gastric ring cell carcinoma classification method | |
CN113011257A (en) | Breast cancer immunohistochemical artificial intelligence interpretation method | |
CN117036343B (en) | FFOCT image analysis method and device for identifying axillary lymph node metastasis | |
CN113762395A (en) | Pancreatic bile duct type ampulla carcinoma classification model generation method and image classification method | |
Li et al. | Computer-aided detection breast cancer in whole slide image | |
CN113838020B (en) | Lesion area quantification method based on molybdenum target image | |
Mathina Kani et al. | Classification of skin lesion images using modified Inception V3 model with transfer learning and augmentation techniques | |
CN114648509A (en) | Thyroid cancer detection system based on multi-classification task | |
Zheng et al. | WPNet: Wide Pyramid Network for Recognition of HER2 Expression Levels in Breast Cancer Evaluation | |
Sakamoto et al. | Deep-learning-based identification of odontogenic keratocysts in hematoxylin-and eosin-stained jaw cyst specimens | |
CN112819042A (en) | Method, system and medium for processing esophageal squamous dysplasia image | |
Bhattacharjee et al. | Invasive Region Segmentation using Pre-trained UNet and Prognosis Analysis of Breast Cancer based on Tumor-Stroma Ratio | |
Phan et al. | Deep learning for detection and classification of nuclear protein in breast cancer tissue | |
Subramanya | Deep Learning Models to Characterize Smooth Muscle Fibers in Hematoxylin and Eosin Stained Histopathological Images of the Urinary Bladder |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190510 |