CN110517219A - A kind of corneal topography method of discrimination and system based on deep learning - Google Patents
A kind of corneal topography method of discrimination and system based on deep learning Download PDFInfo
- Publication number
- CN110517219A CN110517219A CN201910256689.7A CN201910256689A CN110517219A CN 110517219 A CN110517219 A CN 110517219A CN 201910256689 A CN201910256689 A CN 201910256689A CN 110517219 A CN110517219 A CN 110517219A
- Authority
- CN
- China
- Prior art keywords
- corneal topography
- model
- data
- corneal
- discrimination
- 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
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B3/00—Apparatus for testing the eyes; Instruments for examining the eyes
- A61B3/10—Objective types, i.e. instruments for examining the eyes independent of the patients' perceptions or reactions
- A61B3/107—Objective types, i.e. instruments for examining the eyes independent of the patients' perceptions or reactions for determining the shape or measuring the curvature of the cornea
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
- G06T2207/10012—Stereo images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30041—Eye; Retina; Ophthalmic
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Theoretical Computer Science (AREA)
- Ophthalmology & Optometry (AREA)
- Quality & Reliability (AREA)
- General Physics & Mathematics (AREA)
- Radiology & Medical Imaging (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biomedical Technology (AREA)
- Heart & Thoracic Surgery (AREA)
- Molecular Biology (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Public Health (AREA)
- Veterinary Medicine (AREA)
- Eye Examination Apparatus (AREA)
Abstract
The invention discloses a kind of corneal topography method of discrimination and system based on deep learning, the corneal topography obtained in the prior art is pre-processed, obtain the manageable corneal topography characteristics data of corneal topography discrimination model, corneal topography characteristics data are inputted into corneal topography discrimination model, corneal surface shape result is obtained by corneal topography discrimination model.Corneal topography analyze by corneal topography discrimination model and determines its form as a result, the result that doctor can export according to the corneal topography discrimination model directly determines corneal surface shape, and predictablity rate is high.A kind of corneal topography method of discrimination and system based on deep learning provided by the invention, Morphological Identification is carried out to corneal topography using trained convolutional neural networks model, solves the problems, such as to lack the corneal surface shape discrimination technology for carrying out corneal topography deep learning processing analysis in the prior art.
Description
Technical field
The present invention relates to technical field of medical image processing more particularly to a kind of corneal topography based on deep learning to sentence
Other method and system.
Background technique
World Health Organization's current research data report shows, at present China's A nearsighted person's number up to 600,000,000,19~22 years old
The poor eyesight rate of university student is up to 80% or more, and the crowd for having a mind to row refractive surgery increases year by year.It is all to be intended to carry out angle
The patient of film refractive surgery, it is preoperative to be both needed to through corneal topography inspection and evaluation corneal surface shape, to exclude operation taboo.
Current corneal topography is obtained based on modern cornea topographic map detection technique, the system for acquiring corneal surface shape
Have its corresponding corneal surface shape analysis software, can screening go out to be not suitable for row cornea refractive surgery patient.But the number of this analysis
It is white database according to basis, directlying adopt current system will lead to yellow's cornea type identification inaccuracy.Mesh
Preceding corneal surface shape assessment is to be judged by refractive surgery doctor according to clinical experience, this to judge one to lack specific mark
Standard, the result that each doctor obtains are not quite similar, the less doctor of clinical experience be likely to be obtained incorrect judging result to
Increase operation risk, furthermore, assessment is carried out to corneal surface shape and needs the regular hour, the feelings originally lacked in face of medical resource
Condition allows doctor to judge that the corneal surface shape of patient undoubtedly increases the work load of doctor one by one, and doctor can not handle and more attaches most importance to
The work wanted, this is a kind of waste to medical resource.
The reason is that in the prior art, lack the corneal surface shape deep learning discrimination technology for being directed to corneal topography,
In real life, a kind of medical device that can judge automatically corneal surface shape is needed, the angle of patient is judged with unified safe criterion
Film form had both reduced the workload of doctor, also avoided the possible danger of erroneous judgement.
Summary of the invention
The present invention provides a kind of corneal topography method of discrimination and system based on deep learning, it is intended to solve existing skill
Lack the problem of corneal surface shape discrimination technology of deep learning processing analysis is carried out to corneal topography in art.
To achieve the above object, the present invention provides a kind of corneal topography method of discrimination based on deep learning, comprising:
It obtains corneal topography to be discriminated and carries out image preprocessing and obtain corneal topography characteristics data；
Corneal topography characteristics data are inputted into corneal topography discrimination model, obtain the output of corneal topography discrimination model
Corneal surface shape result.
To achieve the above object, the present invention provides a kind of corneal topography judgement system based on deep learning, comprising:
Data capture unit to be discriminated, for obtaining corneal topography to be discriminated and carrying out image preprocessing with obtaining cornea
Shape characteristic；
Corneal surface shape judgement unit obtains angle for corneal topography characteristics data to be inputted corneal topography discrimination model
The corneal surface shape result of film topographic map discrimination model output.
Compared with prior art, a kind of corneal topography method of discrimination based on deep learning disclosed by the invention and it is
System, the corneal topography obtained in the prior art is pre-processed to obtain the manageable cornea of corneal topography discrimination model
Corneal topography characteristics data are inputted corneal topography discrimination model, pass through corneal topography discrimination model by features of terrain data
Obtain corneal surface shape result.Corneal topography analyze by corneal topography discrimination model and determines its form as a result, doctor
The result that life can be exported according to the corneal topography discrimination model directly determines corneal surface shape, does not need according to corneal topography
Figure judged, reduces the work of doctor, also, since corneal topography discrimination model of the present invention is by largely
Determine the convolutional neural networks model that the corneal topography training of form result obtains, predictablity rate is high, avoids reality
In because of the misjudgment that doctor's clinical experience is insufficient and is likely to occur the case where.It is provided by the invention a kind of based on deep learning
Corneal topography method of discrimination and system, using trained reliable convolutional neural networks model to corneal topography carry out shape
State differentiates solve the problems, such as to lack the corneal surface shape discrimination technology for carrying out corneal topography processing analysis in the prior art.
Detailed description of the invention
Fig. 1 is a kind of flow chart of the corneal topography method of discrimination based on deep learning of the present invention；
Fig. 2 is that invention obtains the schematic diagram of corneal topography characteristics data to corneal topography progress image preprocessing；
Fig. 3 is a kind of system block diagram 1 of the corneal topography judgement system based on deep learning of the present invention；
Fig. 4 is a kind of system block diagram 2 of the corneal topography judgement system based on deep learning of the present invention；
Fig. 5 is a kind of system block diagram 3 of the corneal topography judgement system based on deep learning of the present invention.
Specific embodiment
As shown in Figure 1, a kind of corneal topography method of discrimination based on deep learning, comprising: with obtaining cornea to be discriminated
Shape figure simultaneously carries out image preprocessing and obtains corneal topography characteristics data；Corneal topography characteristics data input corneal topography is differentiated
Model obtains the corneal surface shape result of corneal topography discrimination model output.
A kind of corneal topography method of discrimination based on deep learning disclosed by the invention, the angle that will be obtained in the prior art
Film topographic map is pre-processed to obtain the manageable corneal topography characteristics data of corneal topography discrimination model, by corneal topography
Characteristic inputs corneal topography discrimination model, obtains corneal surface shape result by corneal topography discrimination model.Pass through angle
Film topographic map discrimination model, which analyze to corneal topography, determines its form as a result, doctor can sentence according to the corneal topography
The result of other model output directly determines corneal surface shape, does not need to be judged according to corneal topography, reduces the work of doctor
Make, also, since corneal topography discrimination model of the present invention is the corneal topography by largely having determined that form result
The convolutional neural networks model that figure training obtains, predictablity rate is high, avoids in practice because doctor's clinical experience is insufficient
And be likely to occur misjudgment the case where.A kind of corneal topography method of discrimination based on deep learning provided by the invention,
Morphological Identification is carried out to corneal topography using trained reliable convolutional neural networks model, solves and lacks in the prior art
The problem of corneal surface shape discrimination technology of deep learning processing analysis is carried out to corneal topography.In the present embodiment, the angle
Film form result includes normal cornea, suspicious corneal surface shape exception, early stage keratoconus, keratoconus, myopia cornea refractive
After operation this five kinds as a result, suspicious corneal surface shape patients with abnormal Post-operative complications risk is high, and for early stage keratoconus,
Keratoconus, it is not recommended that carry out cornea refractive surgery.Corneal surface shape can be directly obtained according to corneal topography to be discriminated as a result,
The corneal surface shape for determining patient instructs refractive surgery decision.
Further, it is described obtain corneal topography to be discriminated and carry out image preprocessing obtain corneal topography characteristics number
According to, comprising: obtain corneal topography diagram data to be discriminated；The invalid data in corneal topography diagram data is deleted, effective cornea is obtained
Topographic map data merges effective corneal topography data, obtains corneal topography characteristics data.
In a preferred embodiment of the invention, the corneal topography got is cut, is only retained effective
Data delete invalid data, and the corneal topography characteristics data avoided include a large amount of hashes, influence convolutional neural networks
Identification judgement, while reducing data processing amount and model calculation efficiency can be improved.
Modern cornea topographic map (corneal topography) detection is the projection based on cornea image, by Placido
Disk composition, principle are the ring of light that a concentric circles is first projected on cornea, are captured again by video camera by cornea later
The image for the ring that tear film layer is reflected analyzes these data finally by computer software.Concentric interannular
Apart from smaller, then diopter of cornea is higher at this, these data can be depicted as colored temperature figure by computer.But it is based on
The data of the detection system of Placido disk only anterior surface of cornea collected, and the data without posterior surface of cornea, and cornea
Rear surface is the more sensitive feature of kerectasis disease.Scheimpflug camera then can solve this problem, and be one
The camera of a rotation, can produce 3-D image, contain all information from anterior surface of cornea to crystal rear surface.Cornea is disconnected
Layer figure (corneal tomography) is exactly for describing based on image caused by Scheimpflug image instrument.But I
The corneal topography meaning clinically often said include cornea faultage image.What it is using Scheimpflug camera technique is
System can provide the landform and thickness of entire cornea front and rear surfaces, can generate anterior surface of cornea height by software processing
Figure, rear surface height map, corneal thickness figure etc., corneal curvature figure can also be exported from the data of height map.It uses at present
The instrument of Scheimpflug camera mainly has 4 kinds, respectively TMS-5 (Japanese Tomey company), Pentacam HR (Germany
OCULUS company), Sirius (Italian CSO company) and Galilei (Ziemer company, Switzerland).The more common cornea in China
Topographic map instrument is Pentacam and Sirius.Pentacam is the Scheimpflug camera and still camera of rotation
In conjunction with according to the obtained altitude information of its photo, the standard report page that can reconstruct it includes that anterior surface of cornea is axial bent
Rate figure, anterior surface of cornea height map, posterior corneal surface height figure, corneal thickness figure.The corneal topography system of Sirius is then
The combination of Placido disk and a Scheimpflug system singly rotated, the standard report page include that anterior surface of cornea is tangential
Curvature chart, anterior surface of cornea height map, posterior corneal surface height figure, corneal thickness figure, but these data have its own system
Algorithm can not be compared to each other with respective graphical in Pentacam.Several corneal topography instruments of document report have it is very high can
Repeatability, when measuring corneal astigmatism, the result of Pentacam is more reliable.Since Pentacam dioptric tetrad figure preferably reflects
Cornea true form, so select the Pentacam dioptric tetrad figure as the corneal topography diagram data in the present embodiment.
The original image of Pentacam dioptric tetrad figure includes two parts, and left side is patient data and anterior ocular segment parameter, the right side
Side is that tetrad dioptric constitutes figure (anterior surface of cornea axial direction curvature chart, anterior surface of cornea height map, posterior corneal surface height figure, angle
Film thickness figure, position are followed successively by upper left, upper right, bottom right, lower-left), which, which presents, screens refractive surgery trouble to clinician
The most direct four width figure of person.Sharp regions warm tones (red and orange) display, flat in anterior surface of cornea axial direction curvature chart
Region is shown with cool tone (green and blue).Form can behave as symmetrical knot, asymmetric knot, circle, ellipse, not advise
Then etc..Cornea front and rear surfaces height map is that anterior corneal surface detected is described corneal height by matching with reference surface
Details.Point higher than reference surface is considered as increasing, and is indicated with positive value, and the point lower than reference surface is considered as recess,
It is indicated with negative value, equally shows corneal surface shape with cold and warm tone.Since corneal height is not by axial direction, direction, corneal curvature
It influences, so anomaly of cornea can be detected more accurately.The reference surface default setting diameter of cornea front and rear surfaces height map is
8.0mm can be adjusted as needed.Corneal thickness figure is the graphical representation of entire corneal thickness distribution.Measurement result is equally shown
On the topographic map of coloud coding, it is red and it is orange represent relatively thin region, blue and green represent the thicker region of cornea.
In the prior art, refractive surgery doctor need to be according to comprehensive analysis such as four diagram shapes and color characteristics, to corneal surface shape
It judges.There are also the system such as BAD-D expansions that auxiliary doctor's row corneal surface shape differentiates in Pentacam eye anterior segment analysis system
Analysis, keratoconus ABCD stage division；Sirius three-dimension disclocation corneal topography system is then calculated according to SVM (support vector machines)
Corneal surface shape is divided into normal (Normal), doubtful keratoconus (Suspect keratoconus), keratoconus by method
(Abnormal or treated), myopia postoperative straightening after (Keratoconus compatible), exception or treatment
(Myopic Post-OP).Analysis is made in the differentiation that domestic refractive surgery doctor can combine each system, makes preoperative tentative diagnosis.
However, the discrimination standard of two systems foundation is mainly according to white database, the judgement for Chinese myopia population is also needed
According to operative doctor experience and need the interpreting blueprints process of certain time.
So in the present embodiment, select Pentacam dioptric tetrad figure as the corneal topography diagram data in the present embodiment,
And original Pentacam dioptric tetrad figure is only retained to the most important tetrad dioptric composition figure of cornea Morphological Identification (before cornea
Surface axial direction curvature chart, anterior surface of cornea height map, posterior corneal surface height figure, corneal thickness figure, position are followed successively by upper left, the right side
Upper, bottom right, lower-left).
As shown in Fig. 2, in the present embodiment, for Pentacam dioptric tetrad figure pre-treatment step specifically, by dioptric
Useless region cuts (i.e. deletion invalid data), the splicing of right side effective coverage on the left of tetrad figure (i.e. original cornea topographic map data)
(i.e. merging valid data), obtain pretreatment image (i.e. corneal topography characteristics data)；Pixel value [0,255] is zoomed to [0,
1], data normalization is carried out.Original cornea topographic map data includes four image sections: front surface axial direction curvature chart, front surface
Height map, rear surface height map, corneal thickness figure.It is final to differentiate that result is according to four image section forms, temperature, BAD-D
Expansion analysis in conjunction with big data feature and expertise knowledge integration as a result, analyze.Since the present embodiment is to original corneal topography
Effective coverage is extracted in figure, the more accurate accurately status of the model that training is obtained and the angle for judging corneal topography
Film form.
Further, the acquisition step of the corneal topography discrimination model includes: and obtains cornea to differentiate sample data set,
The cornea differentiates that sample data set includes training set, verifying collection, test set；By training set and verifying collection input convolutional Neural net
Network model, computation model output valve differentiate that sample data concentrates the error between corresponding target value and model prediction quasi- with cornea
True rate；Model Weight is updated according to residual error, training set and verifying collection are inputted into updated convolutional neural networks model, calculate mould
Type output valve differentiates that sample data concentrates error and model prediction accuracy rate between corresponding target value with cornea；It will update secondary
Number reaches the convolutional neural networks model of predetermined value as corneal topography discrimination model.
In a preferred embodiment of the invention, corneal topography discrimination model is to differentiate sample data according to a large amount of corneas
The convolutional neural networks model that training obtains.Since convolutional neural networks model has the characteristics that classification is accurate, the present embodiment is logical
It crosses and largely has determined that the corneal topography training convolutional neural networks model of corneal surface shape result obtains, it is ensured that the cornea
The discriminant accuracy of shape figure discrimination model.
In the present embodiment, the training process of convolutional neural networks model are as follows:
S1, network carry out the initialization of weight；
S2, input training set data obtain output valve by the propagation forward of convolutional layer, pond layer, full articulamentum；
S3 finds out error and Model Monitoring index accuracy rate between the output valve of network and target value；
S4 collects verifying, repeats S2, S3；
S5 updates the weight of network according to residual error；
S6 repeats S2, S3, S4, when the number of iterations reaches predetermined value, terminates training.
In the present embodiment, convolutional neural networks model selection Resnet model, loss function formula is:
Weight { y in formulaj=t } indicate t class weight,Expression prevents the penalty term of over-fitting,
InFor the weight of penalty term, x indicates that the picture of input, y indicate corresponding classification, and m, n, kw and k represent batch.
Classical convolutional neural networks CNN model have LeNet, AlexNet, VGGNet, Xception, Inception,
ResNet, ZFNet, DenseNet model etc..In general the deeper effect of network is better, but in the biggish neural network of depth
In, deep learning can cause gradient to disappear and gradient explosion due to network depth.The corresponding solution of tradition is the first of data
Beginningization (normlizedinitializatiton) and regularization (batch normlization), though the solution solves
The problem of gradient of having determined, it but will cause the degeneration of network performance, i.e. depth down, accuracy rate but has dropped.One in CNN history
A milestone event is the appearance of ResNet model.ResNet model can well solve degenerate problem and gradient problem, from
And deeper CNN model can be trained, to realize higher accuracy.Therefore, the Resnet of preferable performance is selected here
Model.
Further, the acquisition cornea differentiates that sample data set includes: to obtain corneal topography diagram data to be discriminated and right
Answer category result；The invalid data in corneal topography diagram data is deleted, effective cornea topographic map data is obtained, merges effective cornea
Topographic map data, extraction obtain corneal topography characteristics data；Angle is obtained according to corneal topography characteristics data and corresponding category result
Film differentiates the sample data that sample data is concentrated.
In a preferred embodiment of the invention, the corneal topography got is cut, is only retained effective
Data delete invalid data, and the corneal topography characteristics data avoided include a large amount of hashes, influence convolutional neural networks
Identification judgement, while reducing data processing amount and model calculation efficiency can be improved.
In the present embodiment, the pre-treatment step for Pentacam dioptric tetrad figure is specifically, (i.e. former by dioptric tetrad figure
Beginning corneal topography diagram data) left side useless region cutting (i.e. deletion invalid data), the splicing of right side effective coverage (merges effective
Data), obtain pretreatment image (i.e. corneal topography characteristics data)；Image after processing is divided into training set, verifying collection and test
Collection；Pixel value [0,255] is zoomed to [0,1], data normalization is carried out.Original cornea topographic map data includes four image portions
Point: front surface axial direction curvature chart, front surface height map, rear surface height map, corneal thickness figure.It is final to differentiate that result is according to four
A image section form, temperature are analyzed in conjunction with big data feature and expertise knowledge integration.Since the present embodiment is to rudimentary horn
Effective coverage is extracted in film topographic map, is allowed and is trained the more accurate accurately status of obtained model and judge corneal topography
The corneal surface shape of figure.
Further, the acquisition step of the corneal topography discrimination model further include: update test set input model
Number reaches the convolutional neural networks model of predetermined value, computation model predictablity rate, when model prediction accuracy rate is not less than pre-
If threshold value, then updated convolutional neural networks model is corneal topography discrimination model；When model prediction accuracy rate is less than
Preset threshold expands training set sample size, again training convolutional neural networks model, until model prediction accuracy rate is not less than
Preset threshold.
In a preferred embodiment of the invention, convolutional neural networks Model checking corneal surface shape is verified by test set
Accuracy rate, the convolutional neural networks model for reaching accuracy rate are used as corneal topography discrimination model.It is accurate for not, meeting
The convolutional neural networks model that rate requires then repeats previous embodiment to convolutional neural networks model using more training samples
In training step, to guarantee obtained corneal topography discrimination model differentiation accuracy rate with higher.
As shown in figure 3, a kind of corneal topography method of discrimination based on deep learning, comprising:
Data capture unit to be discriminated, for obtaining corneal topography to be discriminated and carrying out image preprocessing with obtaining cornea
Shape characteristic；
Corneal surface shape judgement unit obtains angle for corneal topography characteristics data to be inputted corneal topography discrimination model
The corneal surface shape result of film topographic map discrimination model output.
A kind of corneal topography judgement system based on deep learning disclosed by the invention, the angle that will be obtained in the prior art
Film topographic map is pre-processed to obtain the manageable corneal topography characteristics data of corneal topography discrimination model, by corneal topography
Characteristic inputs corneal topography discrimination model, obtains corneal surface shape result by corneal topography discrimination model.Pass through angle
Film topographic map discrimination model, which analyze to corneal topography, determines its form as a result, doctor can sentence according to the corneal topography
The result of other model output directly determines corneal surface shape, does not need to be judged according to corneal topography, reduces the work of doctor
Make, also, since corneal topography discrimination model of the present invention is the corneal topography by largely having determined that form result
The convolutional neural networks model that figure training obtains, predictablity rate is high, avoids in practice because doctor's clinical experience is insufficient
And be likely to occur misjudgment the case where.A kind of corneal topography judgement system based on deep learning provided by the invention,
Morphological Identification is carried out to corneal topography using trained reliable convolutional neural networks model, solves and lacks in the prior art
The problem of corneal surface shape discrimination technology of deep learning processing analysis is carried out to corneal topography.In the present embodiment, the angle
Film form result includes normal cornea, suspicious corneal surface shape exception, early stage keratoconus, keratoconus, myopia cornea refractive
After operation this five kinds as a result, corneal surface shape can be directly obtained according to corneal topography to be discriminated as a result, determine patient cornea
Form.
As shown in figure 4, the data capture unit to be discriminated includes:
Corneal topography data acquisition module to be discriminated, for obtaining corneal topography diagram data to be discriminated；
Corneal topography characteristics data acquisition module to be discriminated is obtained for deleting the invalid data in corneal topography diagram data
To effective cornea topographic map data, merges effective corneal topography data, obtain corneal topography characteristics data.
In a preferred embodiment of the invention, the corneal topography got is cut, is only retained effective
Data delete invalid data, and the corneal topography characteristics data avoided include a large amount of hashes, influence convolutional neural networks
Identification judgement, while reducing data processing amount and model calculation efficiency can be improved.In the present embodiment, Pentacam dioptric is selected
Tetrad figure only retains as the corneal topography diagram data in the present embodiment, and to original Pentacam dioptric tetrad figure to cornea
The most important tetrad dioptric of Morphological Identification constitutes figure (table after anterior surface of cornea axial direction curvature chart, anterior surface of cornea height map, cornea
Face height map, corneal thickness figure, position are followed successively by upper left, upper right, bottom right, lower-left).By dioptric tetrad figure (i.e. original cornea
Shape diagram data) left side useless region cutting (i.e. deletion invalid data), the splicing of right side effective coverage (merges valid data), obtains
To pretreatment image (i.e. corneal topography characteristics data)；Pixel value [0,255] is zoomed to [0,1], data normalization is carried out.By
In the present embodiment to effective coverage is extracted in original corneal topography, allow the obtained model of training more it is accurate accurately
Status and the corneal surface shape for judging corneal topography.
As shown in figure 4, the corneal surface shape judgement unit includes:
Sample data set obtains module, differentiates that sample data set, the cornea differentiate sample data set for obtaining cornea
Including training set, verifying collection, test set；
Model training module, for training set and verifying collection to be inputted convolutional neural networks model, computation model output valve
Differentiate that sample data concentrates error and model prediction accuracy rate between corresponding target value with cornea；
Training set and verifying collection are inputted updated volume for updating Model Weight according to residual error by model modification module
Product neural network model, computation model output valve differentiate that sample data concentrates error and mould between corresponding target value with cornea
Type predictablity rate；
Model generation module reaches predetermined value for Model Weight update times, according to updated convolutional neural networks
Model obtains corneal topography discrimination model.
In a preferred embodiment of the invention, corneal topography discrimination model is to differentiate sample data according to a large amount of corneas
The convolutional neural networks model that training obtains.Since convolutional neural networks model has the characteristics that classification is accurate, the present embodiment is logical
It crosses and largely has determined that the corneal topography training convolutional neural networks model of corneal surface shape result obtains, it is ensured that the cornea
The discriminant accuracy of shape figure discrimination model.Classical convolutional neural networks CNN model have LeNet, AlexNet, VGGNet,
Xception, Inception, ResNet, ZFNet, DenseNet model etc..In general the deeper effect of network is better, but
In the biggish neural network of depth, deep learning can cause gradient to disappear and gradient explosion due to network depth.Tradition is corresponding
Solution is the initialization (normlizedinitializatiton) and regularization (batch normlization) of data,
Though but the solution solves the problems, such as gradient, will cause the degeneration of network performance, i.e. depth down, accuracy rate is but
It has dropped.A milestone event in CNN history is the appearance of ResNet model.ResNet model can well solve degeneration
Problem and gradient problem, so as to train deeper CNN model, to realize higher accuracy.Therefore, the present embodiment
Select the Resnet model of preferable performance.
As shown in figure 5, the sample data set acquisition module includes:
Sample corneal topography obtains module, for obtaining corneal topography diagram data to be discriminated and corresponding category result；
Sample characteristics data acquisition module obtains effective cornea for deleting the invalid data in corneal topography diagram data
Topographic map data, merges effective corneal topography data, and extraction obtains corneal topography characteristics data；
Sample data generation module differentiates sample for obtaining cornea according to corneal topography characteristics data and corresponding category result
The sample data that notebook data is concentrated.
In a preferred embodiment of the invention, the corneal topography got is cut, is only retained effective
Data delete invalid data, and the corneal topography characteristics data avoided include a large amount of hashes, influence convolutional neural networks
Identification judgement, while reducing data processing amount and model calculation efficiency can be improved.
As shown in figure 5, the corneal surface shape judgement unit further include:
Model measurement module, for test set input model update times to be reached to the convolutional neural networks mould of predetermined value
Type, computation model predictablity rate, when model prediction accuracy rate is not less than preset threshold, then updated convolutional neural networks mould
Type is corneal topography discrimination model；When model prediction accuracy rate be less than preset threshold, expand training set sample size, again
Training convolutional neural networks model, until model prediction accuracy rate is not less than preset threshold.
In a preferred embodiment of the invention, convolutional neural networks Model checking cornea type is verified by test set
Accuracy rate, the convolutional neural networks model for reaching accuracy rate are used as corneal topography discrimination model.It is accurate for not, meeting
The convolutional neural networks model that rate requires then repeats previous embodiment to convolutional neural networks model using more training samples
In training step, to guarantee obtained corneal topography discrimination model differentiation accuracy rate with higher.
The above is a preferred embodiment of the present invention, it is noted that for those skilled in the art
For, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also considered as
Protection scope of the present invention.
Claims (10)
1. a kind of corneal topography method of discrimination based on deep learning characterized by comprising
It obtains corneal topography to be discriminated and carries out image preprocessing and obtain corneal topography characteristics data；
Corneal topography characteristics data are inputted into corneal topography discrimination model, obtain the cornea of corneal topography discrimination model output
Form result.
2. a kind of corneal topography method of discrimination based on deep learning according to claim 1, which is characterized in that described
It obtains corneal topography to be discriminated and carries out image preprocessing and obtain corneal topography characteristics data, comprising:
Obtain corneal topography diagram data to be discriminated；
The invalid data in corneal topography diagram data is deleted, effective cornea topographic map data is obtained, merges effective corneal topography
Data obtain corneal topography characteristics data.
3. a kind of corneal topography method of discrimination based on deep learning according to claim 1 or 2, which is characterized in that
The acquisition step of the corneal topography discrimination model includes:
It obtains cornea and differentiates that sample data set, the cornea differentiate that sample data set includes training set, verifying collection, test set；
By training set and verifying collection input convolutional neural networks model, computation model output valve and cornea differentiate that sample data is concentrated
Error and model prediction accuracy rate between corresponding target value；
Model Weight is updated according to residual error, training set and verifying collection are inputted into updated convolutional neural networks model, calculate mould
Type output valve differentiates that sample data concentrates error and model prediction accuracy rate between corresponding target value with cornea；
Update times are reached into the convolutional neural networks model of predetermined value as corneal topography discrimination model.
4. a kind of corneal topography method of discrimination based on deep learning according to claim 3, which is characterized in that described
It obtains cornea and differentiates that sample data set includes:
Obtain corneal topography diagram data to be discriminated and corresponding category result；
The invalid data in corneal topography diagram data is deleted, effective cornea topographic map data is obtained, merges effective corneal topography
Data, extraction obtain corneal topography characteristics data；
The sample data that cornea differentiates sample data concentration is obtained according to corneal topography characteristics data and corresponding category result.
5. a kind of corneal topography method of discrimination based on deep learning according to claim 3, which is characterized in that described
The acquisition step of corneal topography discrimination model further include:
Test set input model update times reach to the convolutional neural networks model of predetermined value, computation model predictablity rate,
When model prediction accuracy rate is not less than preset threshold, then updated convolutional neural networks model is that corneal topography differentiates mould
Type；When model prediction accuracy rate is less than preset threshold, expand training set sample size, training convolutional neural networks model again,
Until model prediction accuracy rate is not less than preset threshold.
6. a kind of corneal topography judgement system based on deep learning characterized by comprising
Data capture unit to be discriminated obtains corneal topography spy for obtaining corneal topography to be discriminated and carrying out image preprocessing
Levy data；
Corneal surface shape judgement unit, for corneal topography characteristics data to be inputted corneal topography discrimination model, with obtaining cornea
The corneal surface shape result of shape figure discrimination model output.
7. a kind of corneal topography method of discrimination based on deep learning according to claim 6, which is characterized in that described
Data capture unit to be discriminated includes:
Corneal topography data acquisition module to be discriminated, for obtaining corneal topography diagram data to be discriminated；
Corneal topography characteristics data acquisition module to be discriminated is had for deleting the invalid data in corneal topography diagram data
Corneal topography diagram data is imitated, merges effective corneal topography data, obtains corneal topography characteristics data.
8. a kind of corneal topography method of discrimination based on deep learning according to claim 6 or 7, which is characterized in that
The corneal surface shape judgement unit includes:
Sample data set obtains module, differentiates that sample data set, the cornea differentiate that sample data set includes for obtaining cornea
Training set, verifying collection, test set；
Model training module, for training set and verifying collection to be inputted convolutional neural networks model, computation model output valve and angle
Film differentiates that sample data concentrates error and model prediction accuracy rate between corresponding target value；
Training set and verifying collection are inputted updated convolution mind for updating Model Weight according to residual error by model modification module
Through network model, computation model output valve differentiates that sample data concentrates the error between corresponding target value and model pre- with cornea
Survey accuracy rate；
Model generation module reaches predetermined value for Model Weight update times, according to updated convolutional neural networks model
Obtain corneal topography discrimination model.
9. a kind of corneal topography method of discrimination based on deep learning according to claim 8, which is characterized in that described
Sample data set obtains module
Sample corneal topography obtains module, for obtaining corneal topography diagram data to be discriminated and corresponding category result；
Sample characteristics data acquisition module obtains effective corneal topography for deleting the invalid data in corneal topography diagram data
Diagram data, merges effective corneal topography data, and extraction obtains corneal topography characteristics data；
Sample data generation module differentiates sample number for obtaining cornea according to corneal topography characteristics data and corresponding category result
According to the sample data of concentration.
10. a kind of corneal topography method of discrimination based on deep learning according to claim 8, which is characterized in that institute
State corneal surface shape judgement unit further include:
Model measurement module is counted for test set input model update times to be reached to the convolutional neural networks model of predetermined value
Model prediction accuracy rate is calculated, when model prediction accuracy rate is not less than preset threshold, then updated convolutional neural networks model is
For corneal topography discrimination model；When model prediction accuracy rate is less than preset threshold, expansion training set sample size is trained again
Convolutional neural networks model, until model prediction accuracy rate is not less than preset threshold.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910256689.7A CN110517219B (en) | 2019-04-01 | 2019-04-01 | Corneal topography distinguishing method and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910256689.7A CN110517219B (en) | 2019-04-01 | 2019-04-01 | Corneal topography distinguishing method and system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110517219A true CN110517219A (en) | 2019-11-29 |
| CN110517219B CN110517219B (en) | 2023-04-07 |
Family
ID=68622273
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910256689.7A Active CN110517219B (en) | 2019-04-01 | 2019-04-01 | Corneal topography distinguishing method and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110517219B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160431A (en) * | 2019-12-19 | 2020-05-15 | 浙江大学 | Method and device for identifying keratoconus based on multi-dimensional feature fusion |
| CN111288973A (en) * | 2020-01-23 | 2020-06-16 | 中山大学 | Method and device for obtaining flow rate of sea surface, computer equipment and storage medium |
| CN111414972A (en) * | 2020-03-30 | 2020-07-14 | 王雁 | XGboost-based automatic eye refractive correction multi-source data classification method |
| CN111553402A (en) * | 2020-04-22 | 2020-08-18 | 首都医科大学附属北京同仁医院 | Intelligent orthokeratology lens selecting system and method based on big data and deep learning |
| CN112036448A (en) * | 2020-08-11 | 2020-12-04 | 上海鹰瞳医疗科技有限公司 | Keratoconus identification method and device |
| CN113822928A (en) * | 2021-11-22 | 2021-12-21 | 杭州目乐医疗科技股份有限公司 | Corneal topography reconstruction method and device |
| WO2023272876A1 (en) * | 2021-06-28 | 2023-01-05 | 上海美沃精密仪器股份有限公司 | Binocular keratoconus diagnosis method based on multi-modal data |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102961118A (en) * | 2012-11-23 | 2013-03-13 | 浙江工业大学 | Method for measuring diopter and drawing corneal topography diagram based on Placido plate |
| CN107847126A (en) * | 2015-07-24 | 2018-03-27 | 卡尔蔡司医疗技术股份公司 | Method for determining cornea topographic map |
| CN108009525A (en) * | 2017-12-25 | 2018-05-08 | 北京航空航天大学 | A kind of specific objective recognition methods over the ground of the unmanned plane based on convolutional neural networks |
| CN108814544A (en) * | 2018-04-25 | 2018-11-16 | 天津市索维电子技术有限公司 | A kind of anterior ocular segment morphological analysis method |
| CN108985328A (en) * | 2018-06-08 | 2018-12-11 | 佛山市顺德区中山大学研究院 | For differentiating the construction method and its system of the deep learning model of corneal ulceration |
| CN109410268A (en) * | 2018-11-06 | 2019-03-01 | 温州雷蒙光电科技有限公司 | A kind of determination method and system in the concentric loop center of circle of corneal topography |
-
2019
- 2019-04-01 CN CN201910256689.7A patent/CN110517219B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102961118A (en) * | 2012-11-23 | 2013-03-13 | 浙江工业大学 | Method for measuring diopter and drawing corneal topography diagram based on Placido plate |
| CN107847126A (en) * | 2015-07-24 | 2018-03-27 | 卡尔蔡司医疗技术股份公司 | Method for determining cornea topographic map |
| CN108009525A (en) * | 2017-12-25 | 2018-05-08 | 北京航空航天大学 | A kind of specific objective recognition methods over the ground of the unmanned plane based on convolutional neural networks |
| CN108814544A (en) * | 2018-04-25 | 2018-11-16 | 天津市索维电子技术有限公司 | A kind of anterior ocular segment morphological analysis method |
| CN108985328A (en) * | 2018-06-08 | 2018-12-11 | 佛山市顺德区中山大学研究院 | For differentiating the construction method and its system of the deep learning model of corneal ulceration |
| CN109410268A (en) * | 2018-11-06 | 2019-03-01 | 温州雷蒙光电科技有限公司 | A kind of determination method and system in the concentric loop center of circle of corneal topography |
Non-Patent Citations (2)
| Title |
|---|
| 刘洋等: ""自动化机器学习在眼部疾病识别及分类中的初步应用"", 《中国数字医学》 * |
| 杨露菁等: "《智能图像处理及应用》", 31 March 2019 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160431A (en) * | 2019-12-19 | 2020-05-15 | 浙江大学 | Method and device for identifying keratoconus based on multi-dimensional feature fusion |
| CN111288973A (en) * | 2020-01-23 | 2020-06-16 | 中山大学 | Method and device for obtaining flow rate of sea surface, computer equipment and storage medium |
| CN111288973B (en) * | 2020-01-23 | 2021-09-24 | 中山大学 | Method and device for obtaining flow rate of sea surface, computer equipment and storage medium |
| CN111414972A (en) * | 2020-03-30 | 2020-07-14 | 王雁 | XGboost-based automatic eye refractive correction multi-source data classification method |
| CN111414972B (en) * | 2020-03-30 | 2023-09-05 | 王雁 | XGBoost-based eye refraction correction multisource data automatic classification method |
| CN111553402B (en) * | 2020-04-22 | 2023-07-18 | 首都医科大学附属北京同仁医院 | Intelligent cornea shaping lens selecting system and method based on big data and deep learning |
| CN111553402A (en) * | 2020-04-22 | 2020-08-18 | 首都医科大学附属北京同仁医院 | Intelligent orthokeratology lens selecting system and method based on big data and deep learning |
| CN112036448A (en) * | 2020-08-11 | 2020-12-04 | 上海鹰瞳医疗科技有限公司 | Keratoconus identification method and device |
| US11717151B2 (en) | 2021-06-28 | 2023-08-08 | Shanghai MediWorks Precision Instruments Co., Ltd. | Method for early diagnosis of keratoconus based on multi-modal data |
| WO2023272876A1 (en) * | 2021-06-28 | 2023-01-05 | 上海美沃精密仪器股份有限公司 | Binocular keratoconus diagnosis method based on multi-modal data |
| JP2023540651A (en) * | 2021-06-28 | 2023-09-26 | 上海美沃精密儀器股▲フン▼有限公司 | Binocular keratoconus testing method based on multimodal data |
| JP7454679B2 (en) | 2021-06-28 | 2024-03-22 | 上海美沃精密儀器股▲フン▼有限公司 | Binocular keratoconus evaluation system based on multimodal data |
| CN113822928A (en) * | 2021-11-22 | 2021-12-21 | 杭州目乐医疗科技股份有限公司 | Corneal topography reconstruction method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110517219B (en) | 2023-04-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110517219A (en) | A kind of corneal topography method of discrimination and system based on deep learning | |
| Hogarty et al. | Current state and future prospects of artificial intelligence in ophthalmology: a review | |
| Zhang et al. | A survey on computer aided diagnosis for ocular diseases | |
| Hu et al. | Unified diagnosis framework for automated nuclear cataract grading based on smartphone slit-lamp images | |
| Kauppi | Eye fundus image analysis for automatic detection of diabetic retinopathy | |
| Goldbaum et al. | Progression of patterns (POP): a machine classifier algorithm to identify glaucoma progression in visual fields | |
| US20230022921A1 (en) | System and method for analyzing corneal lesion using anterior ocular segment image, and computer-readable recording medium | |
| CN109036556A (en) | A method of keratoconus case is diagnosed based on machine learning | |
| CN113243887B (en) | Intelligent diagnosis and treatment instrument for macular degeneration of old people | |
| Yugha et al. | An automated glaucoma detection from fundus images based on deep learning network | |
| Adel et al. | Automatic classification of retinal eye diseases from optical coherence tomography using transfer learning | |
| Wang et al. | Accurate disease detection quantification of iris based retinal images using random implication image classifier technique | |
| CN114445666A (en) | Deep learning-based method and system for classifying left eye, right eye and visual field positions of fundus images | |
| Giancardo | Automated fundus images analysis techniques to screen retinal diseases in diabetic patients | |
| Mukherjee et al. | Predictive diagnosis of glaucoma based on analysis of focal notching along the neuro-retinal rim using machine learning | |
| Consejo et al. | Detection of subclinical keratoconus with a validated alternative method to corneal densitometry | |
| Suedumrong et al. | Application of deep convolutional neural networks vgg-16 and ***net for level diabetic retinopathy detection | |
| CN113273959B (en) | Portable diabetic retinopathy diagnosis and treatment instrument | |
| Wang et al. | Advances in artificial intelligence models and algorithms in the field of optometry | |
| Ratul et al. | Detection of Primary Open Angle Glaucoma Based on Deep CNN Using Fundus Images | |
| Tekouabou et al. | Machine learning aprroach for early detection of glaucoma from visual fields | |
| Haihong et al. | SAE-wAMD: A self-attention enhanced convolution neural network for fine-grained classification of wet age-related macular degeneration using OCT Images | |
| Otuna-Hernández et al. | Diagnosis and degree of evolution in a Keratoconus-type corneal Ectasia from image processing | |
| Saleh et al. | Artificial intelligence in corneal topography | |
| ÇİFTCİ | ARTIFICIAL INTELLIGENCE FOR CATARACT |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20210817 Address after: 510000 No. 54 martyrs South Road, Guangdong, Guangzhou Applicant after: ZHONGSHAN OPHTHALMIC CENTER, SUN YAT-SEN University Address before: 510000 Zhongshan ophthalmology center of Sun Yat sen University (Zhujiang New Town hospital area), No. 7, Jinsui Road, Guangzhou, Guangdong Province Applicant before: Liu Quan Applicant before: Xie Yi Applicant before: Lin Haotian Applicant before: Zhao Lanqin |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |