CN106570515A - Method and system for treating medical images - Google Patents
Method and system for treating medical images Download PDFInfo
- Publication number
- CN106570515A CN106570515A CN201610392890.4A CN201610392890A CN106570515A CN 106570515 A CN106570515 A CN 106570515A CN 201610392890 A CN201610392890 A CN 201610392890A CN 106570515 A CN106570515 A CN 106570515A
- Authority
- CN
- China
- Prior art keywords
- data
- medical images
- strengthen
- pathological changes
- medical image
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/40—Image enhancement or restoration by the use of histogram techniques
Abstract
The invention provides an accurate and reliable method and an accurate and reliable system for treating medical images and aims to solve a problem in the prior art. The method comprises steps that A, multiple original sample medical images after lesion point calibration are acquired; B, data pre-treatment on the multiple original sample medical images is carried out to acquire multiple training sample medical images; C, depth neural training for the multiple training sample medical images is carried out to acquire a lesion point identification model; and D, test medical images are inputted to the lesion point identification model to acquire a lesion point identification result.
Description
Technical field
The present invention relates to technical field of computer vision, a kind of particularly method and system for processing medical image.
Background technology
Pathological changes or potential pathological changes (such as enteral breath are identified computer automation is allowed from medical image using algorithm
Meat) it is a problem that for many years people are solved in trial.
Traditional Computer Automatic Recognition algorithm is so work：Original image input (pixel value) is converted into into people
The feature (human-engineered features) of manual definition, such as SIFT, HOG features etc..Then after convert these
Feature be put into a good shallow-layer detector of training in advance in detected, the process of detection is substantially it is to be understood that in original
The detection window of beginning picture one pre-set size of slip, if the detection fraction in some position calculation out is higher than
Some pre-set threshold value, then it is assumed that this position has our pathological changes interested or potential pathological changes.
But, prior art is applied
Positives), this is primarily due to the hidden layer of model that this kind of method uses generally only comprising an extraction feature, extracts
Feature be often not enough to portray and distinguish pathological changes point and normal region.
The content of the invention
In view of this, the present invention provides a kind of method and system for accurately and reliably processing medical image, existing to solve
Problem in technology.
First aspect present invention proposes a kind of method for processing medical image, including：Step A：Obtain multiple through demarcating
The original sample medical image of pathological changes point；Step B：Data prediction is carried out to multiple original sample medical images, is obtained
Multiple training sample medical images；Step C：Depth network training is carried out to multiple training sample medical images, disease is obtained
Height identification model；Step D：Test medical image is input into into the pathological changes point identification model, pathological changes point recognition result is obtained.
Optionally, step B includes：Multiple original sample medical images are carried out with black surround operation, breviary is removed
Graphic operation, picture size unitize in operation, the centralization of RGB numerical value and equalization operation, data enhancement operations at least one
Operation is planted, multiple training sample medical images are obtained.
Optionally, the data enhancement operations are at least one of following option：Data based on Arbitrary Rotation increase
By force, the data based on histogram equalization strengthen, the data based on white balance strengthen, the data based on mirror image operation strengthen, are based on
The data of random shearing are strengthened, are strengthened based on the data of the different illumination variations of simulation.
Optionally, step C includes：Tied using the network including convolutional layer, warp lamination, pond layer and full articulamentum
Structure carries out depth network training to multiple training sample medical images, obtains pathological changes point identification model.
Second aspect present invention proposes a kind of system for processing medical image, including：Acquisition module, for obtaining multiple Jing
Cross the original sample medical image for demarcating pathological changes point；Pretreatment module, for carrying out to multiple original sample medical images
Data prediction, obtains multiple training sample medical images；MBM, for entering to multiple training sample medical images
Row depth network training, obtains pathological changes point identification model；Identification module, for test medical image is input into the pathological changes point knowledge
Other model, obtains pathological changes point recognition result.
Optionally, the pretreatment module is additionally operable to：Multiple original sample medical images are carried out black surround operation,
Breviary graphic operation, picture size is gone to unitize in operation, the centralization of RGB numerical value and equalization operation, data enhancement operations
At least one operation, obtains multiple training sample medical images.
Optionally, in the pretreatment module, the data enhancement operations are at least one of following option：Based on any
Angle rotation data strengthen, based on histogram equalization data strengthen, based on white balance data strengthen, based on mirror image operation
Data strengthen, the data based on random shearing strengthen, strengthened based on the data of the different illumination variations of simulation.
Optionally, the MBM is additionally operable to：Using including convolutional layer, warp lamination, pond layer and full articulamentum
Network structure carries out depth network training to multiple training sample medical images, obtains pathological changes point identification model.
Technical scheme processes medical picture to identify pathological changes based on neutral net and depth learning technology
Point, at least has the advantages that：(1) accuracy of algorithm is greatly improved, is reduced and is failed to report.(2) depth net has been used
Network, overcomes shallow-layer network characterization and extracts insufficient problem.
Description of the drawings
Accompanying drawing does not constitute inappropriate limitation of the present invention for more fully understanding the present invention.Wherein：
Fig. 1 is the schematic diagram of the key step of the method for the process medical image according to embodiment of the present invention；
Fig. 2 be according to embodiment of the present invention process medical image method in deep learning network schematic diagram；
Fig. 3 is the schematic diagram of the main modular of the system of the process medical image according to embodiment of the present invention.
Specific embodiment
The exemplary embodiment of the present invention is explained below in conjunction with accompanying drawing, including embodiment of the present invention
They should be thought only exemplary to help understanding by various details.Therefore, those of ordinary skill in the art should recognize
Know, various changes and modifications can be made to embodiment described herein, without departing from scope and spirit of the present invention.
Equally, for clarity and conciseness, the description to known function and structure is eliminated in description below.
Shortcoming of the prior art has been elaborated in background technology.The deep learning adopted in technical scheme
The characteristics of neutral net used in scheme has extraction object high-level characteristic.As high-level characteristic information is low-level image feature information
Linear processes conversion, therefore deep-neural-network can more extract compared to shallow-layer network and can portray object to be described
Substitutive characteristics, such that it is able to lift scheme effect.
Fig. 1 is the schematic diagram of the key step of the method for the process medical image according to embodiment of the present invention.Such as Fig. 1 institutes
Show, the method for the process medical image of the embodiment mainly includes steps A to step D.
Step A：Obtain multiple original sample medical images through demarcating pathological changes point.
In one particular embodiment of the present invention, nearly 10,000 by professional (internist) to thousand of patients
Intestinal endoscopy picture is labeled, and annotation tool is using similar application softwares such as photoshop.Annotation process will lesion locations
Different colours are coated as figure layer.Marked content selectively can also include in addition to polyp and doubtful polyp locations
Under mucosa, lipoma, pleat, the content such as stool.The size difference for being generally found polyp is very big.If by the polyp for differing in size
Mix training pattern and can not possibly reach reasonable effect, therefore different classes of grade is classified as according to polyp size.
Step B：Data prediction is carried out to multiple original sample medical images, multiple training sample medical images are obtained.
Alternatively, multiple original sample medical images are carried out with black surround operation, breviary graphic operation, picture size is gone
At least one operation in unitized operation, the centralization of RGB numerical value and equalization operation, data enhancement operations, obtains multiple
Training sample medical image.
Specifically, following pretreatment is carried out to the original sample medical image that previous step is obtained：Remove image first
Black surround and the thumbnail for removing the upper left corner.Then image RGB data is carried out the unification of centralization, equalization and picture size
After the pretreatment such as change, the data form for being easy to neutral net to read, such as h5 or LMDB etc. are stored into.Wherein, centralization makes
With the scope of parameter between 100-150.Equalization is using parameter between 100-150.Instruction after size is unitized
White silk sample medical image magnitude range is 256 × 256 pixels between 512 × 512 pixels.
The enhanced concrete mode of data including but not limited to rotation at any angle, histogram equalization, white balance, mirror image behaviour
Work, random shearing, the different illumination variations of simulation etc..It is right with the operation of the different light conditions of simulation wherein to rotate at any angle
Lift scheme effect has bigger meaning.
Step C：Depth network training is carried out to multiple training sample medical images, pathological changes point identification model is obtained.
Alternatively, adopting includes convolutional layer (convolutional layer), warp lamination (deconvolutional
Layer), the network structure of pond layer (pooling layer) and full articulamentum (fully connected layer) is to multiple
Training sample medical image carries out depth network training, obtains pathological changes point identification model.Fig. 2 is according to embodiment of the present invention
Process the schematic diagram of the deep learning network in the method for medical image.
Optionally, using depth network training instrument Caffe (http：//caffe.berkeleyvision.org/) enter
Row model training.Using in addition to using network structure file, also need when this instrument define solver files.solver
File gives the method i.e. back-propagation algorithm of parameter of optimizing model (training).Key parameter is included but is not limited to：Base
Plinth learning rate (base learning rate) scope 0.0001 to 0.01；Study momentum (momentum) scope 0.9 to 0.99；
Weight penalty coefficient (weight_decay) scope 0.0001 to 0.001；The scope of epoch is set to 100 to 300.Need
Bright, " epoch " refers to all pictures gone around in once training set, and the meaning is and goes around 100 to 300 training here
All pictures in set.
Step D：Test medical image is input into into pathological changes point identification model, pathological changes point recognition result is obtained.
In a particular embodiment of the present invention, test set has nearly thousand sheets test medical image, wherein comprising three/
One positive sample (polyp) picture.The size of these test medical images and the size one of training sample medical image above
Cause.
Through Model Identification, it is found that rate of failing to report is very low.Therefore the method for the process medical image of the present invention can be helped
Doctor does the pre-sifted work of medical image, can so help doctor to save a large amount of manual times, raise labour productivity, effectively
Distribution medical resource in short supply.
Fig. 3 is the schematic diagram of the main modular of the system of the process medical image according to embodiment of the present invention.Such as Fig. 3 institutes
Show, the system of the process medical image of the embodiment, including：Acquisition module 100, pretreatment module 200, MBM 300
With identification module 400.
Acquisition module 100 is used to obtain multiple original sample medical images through demarcating pathological changes point.Pretreatment module 200
For data prediction being carried out to multiple original sample medical images, obtain multiple training sample medical images.Modeling mould
Block 300 obtains pathological changes point identification model for carrying out depth network training to multiple training sample medical images.Identification mould
Block 400 is used to for test medical image to be input into the pathological changes point identification model, obtains pathological changes point recognition result.
Alternatively, pretreatment module 200 is additionally operable to：Multiple original sample medical images are carried out black surround operation,
Breviary graphic operation, picture size is gone to unitize in operation, the centralization of RGB numerical value and equalization operation, data enhancement operations
At least one operation, obtains multiple training sample medical images.
Alternatively, in the pretreatment module 200, the data enhancement operations are at least one of following option：It is based on
The data of Arbitrary Rotation strengthen, the data based on histogram equalization strengthen, the data based on white balance strengthen, based on mirror image
The data of operation strengthen, the data based on random shearing strengthen, the data enhancing based on the different illumination variations of simulation.
Alternatively, MBM 300 is additionally operable to：Using the net including convolutional layer, warp lamination, pond layer and full articulamentum
Network structure carries out depth network training to multiple training sample medical images, obtains pathological changes point identification model.
Technical scheme processes medical picture to identify pathological changes based on neutral net and depth learning technology
Point, at least has the advantages that：(1) accuracy of algorithm is greatly improved, is reduced and is failed to report.(2) depth net has been used
Network, overcomes shallow-layer network characterization and extracts insufficient problem.
Above-mentioned specific embodiment, does not constitute limiting the scope of the invention.Those skilled in the art should be bright
It is white, depending on design requirement and other factors, various modifications, combination, sub-portfolio and replacement can occur.It is any
Modification, equivalent and improvement for being made within the spirit and principles in the present invention etc., should be included in the scope of the present invention
Within.
Claims (8)
1. it is a kind of process medical image method, it is characterised in that include：
Step A：Obtain multiple original sample medical images through demarcating pathological changes point；
Step B：Data prediction is carried out to multiple original sample medical images, multiple training sample medical images are obtained；
Step C：Depth network training is carried out to multiple training sample medical images, pathological changes point identification model is obtained；
Step D：Test medical image is input into into the pathological changes point identification model, pathological changes point recognition result is obtained.
2. it is according to claim 1 process medical image method, it is characterised in that step B includes：
Multiple original sample medical images are carried out black surround operation, go the unitized operation of breviary graphic operation, picture size,
At least one operation in the centralization of RGB numerical value and equalization operation, data enhancement operations, obtains multiple training sample medical science
Image.
3. it is according to claim 2 process medical image method, it is characterised in that the data enhancement operations are as follows
At least one of option：Based on Arbitrary Rotation data strengthen, based on histogram equalization data strengthen, based on white balance
Data strengthen, the data based on mirror image operation strengthen, the data based on random shearing strengthen, based on the different illumination variations of simulation
Data strengthen.
4. it is according to claim 1 process medical image method, it is characterised in that step C includes：
Using the network structure including convolutional layer, warp lamination, pond layer and full articulamentum to multiple training sample medical science
Image carries out depth network training, obtains pathological changes point identification model.
5. it is a kind of process medical image system, it is characterised in that include：
Acquisition module, for obtaining multiple original sample medical images through demarcating pathological changes point；
Pretreatment module, for carrying out data prediction to multiple original sample medical images, obtains multiple training samples
Medical image；
MBM, for carrying out depth network training to multiple training sample medical images, obtains pathological changes point identification mould
Type；
Identification module, for test medical image is input into the pathological changes point identification model, obtains pathological changes point recognition result.
6. it is according to claim 5 process medical image system, it is characterised in that the pretreatment module is additionally operable to：
Multiple original sample medical images are carried out with black surround operation, the unitized operation of breviary graphic operation, picture size, RGB is removed
At least one operation in numerical value centralization and equalization operation, data enhancement operations, obtains multiple training sample medical science figures
Picture.
7. it is according to claim 5 process medical image system, it is characterised in that it is in the pretreatment module, described
Data enhancement operations are at least one of following option：Based on Arbitrary Rotation data strengthen, based on histogram equalization
Data strengthen, the data based on white balance strengthen, the data based on mirror image operation strengthen, the data based on random shearing strengthen,
Strengthened based on the data of the different illumination variations of simulation.
8. it is according to claim 5 process medical image system, it is characterised in that the MBM is additionally operable to：
Using the network structure including convolutional layer, warp lamination, pond layer and full articulamentum to multiple training sample medical science
Image carries out depth network training, obtains pathological changes point identification model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610392890.4A CN106570515A (en) | 2016-05-26 | 2016-05-26 | Method and system for treating medical images |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610392890.4A CN106570515A (en) | 2016-05-26 | 2016-05-26 | Method and system for treating medical images |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106570515A true CN106570515A (en) | 2017-04-19 |
Family
ID=58532136
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610392890.4A Pending CN106570515A (en) | 2016-05-26 | 2016-05-26 | Method and system for treating medical images |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106570515A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107369151A (en) * | 2017-06-07 | 2017-11-21 | 万香波 | System and method are supported in GISTs pathological diagnosis based on big data deep learning |
| CN108010581A (en) * | 2017-12-01 | 2018-05-08 | 中南大学 | A kind of seborrheic keratosis and flat wart disease sorting technique |
| CN108009598A (en) * | 2017-12-27 | 2018-05-08 | 北京诸葛找房信息技术有限公司 | Floor plan recognition methods based on deep learning |
| CN108427988A (en) * | 2018-03-14 | 2018-08-21 | 重庆金山医疗器械有限公司 | A kind of alimentary canal anatomical position identification device |
| CN111127426A (en) * | 2019-12-23 | 2020-05-08 | 山东大学齐鲁医院 | Gastric mucosa cleanliness evaluation method and system based on deep learning |
| WO2020148797A1 (en) * | 2019-01-15 | 2020-07-23 | オリンパス株式会社 | Image processing system, learned model, and image processing method |
| CN113903433A (en) * | 2021-10-14 | 2022-01-07 | 数坤(北京)网络科技股份有限公司 | Image processing method and device and electronic equipment |
| CN114445406A (en) * | 2022-04-07 | 2022-05-06 | 武汉大学 | Enteroscopy image analysis method and device and medical image processing equipment |
-
2016
- 2016-05-26 CN CN201610392890.4A patent/CN106570515A/en active Pending
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107369151A (en) * | 2017-06-07 | 2017-11-21 | 万香波 | System and method are supported in GISTs pathological diagnosis based on big data deep learning |
| CN108010581A (en) * | 2017-12-01 | 2018-05-08 | 中南大学 | A kind of seborrheic keratosis and flat wart disease sorting technique |
| CN108010581B (en) * | 2017-12-01 | 2021-12-10 | 中南大学 | Method for classifying seborrheic keratosis and verruca plana diseases |
| CN108009598A (en) * | 2017-12-27 | 2018-05-08 | 北京诸葛找房信息技术有限公司 | Floor plan recognition methods based on deep learning |
| CN108427988A (en) * | 2018-03-14 | 2018-08-21 | 重庆金山医疗器械有限公司 | A kind of alimentary canal anatomical position identification device |
| WO2020148797A1 (en) * | 2019-01-15 | 2020-07-23 | オリンパス株式会社 | Image processing system, learned model, and image processing method |
| JPWO2020148797A1 (en) * | 2019-01-15 | 2021-09-30 | オリンパス株式会社 | Image processing system, trained model and image processing method |
| JP7104810B2 (en) | 2019-01-15 | 2022-07-21 | オリンパス株式会社 | Image processing system, trained model and image processing method |
| US11721086B2 (en) | 2019-01-15 | 2023-08-08 | Olympus Corporation | Image processing system and image processing method |
| CN111127426A (en) * | 2019-12-23 | 2020-05-08 | 山东大学齐鲁医院 | Gastric mucosa cleanliness evaluation method and system based on deep learning |
| CN113903433A (en) * | 2021-10-14 | 2022-01-07 | 数坤(北京)网络科技股份有限公司 | Image processing method and device and electronic equipment |
| CN114445406A (en) * | 2022-04-07 | 2022-05-06 | 武汉大学 | Enteroscopy image analysis method and device and medical image processing equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106570515A (en) | Method and system for treating medical images | |
| CN107316077B (en) | Automatic adipose cell counting method based on image segmentation and edge detection | |
| CN107977671B (en) | Tongue picture classification method based on multitask convolutional neural network | |
| JP7198577B2 (en) | Image analysis method, device, program, and method for manufacturing trained deep learning algorithm | |
| CN109791693B (en) | Digital pathology system and related workflow for providing visualized whole-slice image analysis | |
| CN110050281B (en) | Annotating objects in a learning image | |
| WO2018049084A1 (en) | Methods and systems for human imperceptible computerized color transfer | |
| CN103366361B (en) | Region growing methods and there is the region growing methods of mark function | |
| CN103020585B (en) | A kind of immuning tissue's positive cell and negative cells recognition methods | |
| CN108596102B (en) | RGB-D-based indoor scene object segmentation classifier construction method | |
| CN106530290A (en) | Medical image analysis method and device | |
| CN108805018A (en) | Road signs detection recognition method, electronic equipment, storage medium and system | |
| CN110189383B (en) | Traditional Chinese medicine tongue color and fur color quantitative analysis method based on machine learning | |
| KR102208694B1 (en) | Apparatus and method for analyzing mark in facility floor plan | |
| CN111508016B (en) | Vitiligo region chromaticity value and area calculation method based on image processing | |
| CN105138975B (en) | A kind of area of skin color of human body dividing method based on degree of depth conviction network | |
| Manning et al. | Image analysis and machine learning-based malaria assessment system | |
| Kanjalkar et al. | Detection and classification of plant leaf diseases using ANN | |
| CN111160194B (en) | Static gesture image recognition method based on multi-feature fusion | |
| CN113793301A (en) | Training method of fundus image analysis model based on dense convolution network model | |
| CN113223041A (en) | Method, system and storage medium for automatically extracting target area in image | |
| Mohan et al. | An intelligent recognition system for identification of wood species | |
| Zhang et al. | Pulse coupled neural network edge-based algorithm for image text locating | |
| CN110533049A (en) | The method and apparatus for extracting seal image | |
| Kunwar | Malaria detection using image processing and machine learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| DD01 | Delivery of document by public notice | ||
| DD01 | Delivery of document by public notice |
Addressee: Ding Peng Document name: Notification to Make Rectification |
|
| DD01 | Delivery of document by public notice | ||
| DD01 | Delivery of document by public notice |
Addressee: Ding Peng Document name: Notification of Passing Examination on Formalities |
|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170419 |