CN108985328A - For differentiating the construction method and its system of the deep learning model of corneal ulceration - Google Patents
For differentiating the construction method and its system of the deep learning model of corneal ulceration Download PDFInfo
- Publication number
- CN108985328A CN108985328A CN201810586987.8A CN201810586987A CN108985328A CN 108985328 A CN108985328 A CN 108985328A CN 201810586987 A CN201810586987 A CN 201810586987A CN 108985328 A CN108985328 A CN 108985328A
- Authority
- CN
- China
- Prior art keywords
- deep learning
- learning model
- image
- pixel
- effective coverage
- 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
- 238000013136 deep learning model Methods 0.000 title claims abstract description 96
- 206010064996 Ulcerative keratitis Diseases 0.000 title claims abstract description 22
- 238000010276 construction Methods 0.000 title claims abstract description 22
- 238000012360 testing method Methods 0.000 claims abstract description 51
- 210000004087 cornea Anatomy 0.000 claims abstract description 41
- 238000000034 method Methods 0.000 claims description 27
- 230000008569 process Effects 0.000 claims description 24
- 238000012549 training Methods 0.000 claims description 19
- 238000012216 screening Methods 0.000 claims description 12
- 238000012545 processing Methods 0.000 claims description 9
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 7
- 230000008859 change Effects 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 238000000605 extraction Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 abstract description 6
- 230000004069 differentiation Effects 0.000 abstract description 5
- 230000000694 effects Effects 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 5
- 238000013527 convolutional neural network Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000000975 dye Substances 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- GNBHRKFJIUUOQI-UHFFFAOYSA-N fluorescein Chemical compound O1C(=O)C2=CC=CC=C2C21C1=CC=C(O)C=C1OC1=CC(O)=CC=C21 GNBHRKFJIUUOQI-UHFFFAOYSA-N 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000452 restraining effect Effects 0.000 description 1
- 208000024891 symptom Diseases 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Eye Examination Apparatus (AREA)
Abstract
The invention discloses the construction methods and its system for differentiating the deep learning model of corneal ulceration, cornea test image is handled, and different deep learning models is obtained using AlexNet and VGGNet, and after this, new deep learning model is obtained by all kinds of parameters of continuous percentage regulation learning model, after each adjustment, it adjusted is compared based on the higher deep learning model of discrimination with next time, therefore it can be optimized respectively for the parameters of deep learning model, generally by these optimizations, discrimination can be greatly promoted.Therefore, The present invention gives the construction method of the deep learning model after optimization, this deep learning model can replace human eye differentiation and discrimination with higher, and medical worker and researcher is facilitated to use.
Description
Technical field
The present invention relates to deep learning fields, in particular for the construction method of the deep learning model of differentiation corneal ulceration
And its system.
Background technique
In conventional methods where, doctor dyes the color distinction to be formed mainly by means of corneal fluorescein to judge in cornea
Healthy area and affected area artificially judge that this corneal ulceration symptom is in the serious phase then according to the feature of affected area
Or the transitional period, but the judgement of human eye there are error and is easy to be influenced by subjective factor, therefore can be to the objective of the state of an illness
Analysis has an impact, to influence the correct treatment to patient.
Summary of the invention
To solve the above-mentioned problems, the object of the present invention is to provide the structures for differentiating the deep learning model of corneal ulceration
Construction method and its system, the deep learning model after giving optimization can replace human eye differentiation and discrimination with higher,
Medical worker and researcher is facilitated to use.
In order to make up for the deficiencies of the prior art, the technical solution adopted by the present invention is that:
For differentiating the construction method of the deep learning model of corneal ulceration, comprising the following steps:
Cornea test image is handled, cornea processing image is obtained;
Two deep learning models are obtained based on AlexNet and VGGNet training and utilize this two deep learning models point
Not diagonal film process image is identified, discrimination is higher as the first deep learning model;
The convolution number of plies for adjusting the first deep learning model, by distinguishing before adjusting with the first deep learning model adjusted
Diagonal film process image is identified, discrimination is higher as the second deep learning model;
For adjusting the convolution nucleus number of the second deep learning model, by adjusting preceding and the second deep learning model adjusted
Diagonal film process image is identified respectively, and discrimination is higher as third deep learning model;
For adjusting the optimizer model of third deep learning model, by adjusting preceding and third deep learning mould adjusted
Diagonal film process image is identified type respectively, and discrimination is higher as final discrimination model.
Further, cornea test image is handled, comprising:
Cornea test image is cut, to extract effective coverage;The effective coverage is picture in cornea test image
The region that plain value is not zero;
Brightness and the random adjustment of contrast are carried out to the image after cutting, to realize that sample expands.
Further, cornea test image is cut, to extract effective coverage, comprising:
Setting up pixel value threshold value is X, progressively scans pixel in cornea test image;
The ordinate of pixel of first pixel value obtained greater than X, the vertical seat of uppermost point A as in effective coverage
Mark;
The ordinate of pixel of the last one pixel value obtained greater than X, bottom point C's is vertical as in effective coverage
Coordinate;
The abscissa of pixel of first pixel value obtained greater than X, the horizontal seat of leftmost point B as in effective coverage
Mark;
The abscissa of pixel of the last one pixel value obtained greater than X, the cross of rightmost point D as in effective coverage
Coordinate;
Along the transverse direction of A, C, the longitudinal direction of B, D are cut, to extract effective coverage.
Further, the random adjustment of brightness and contrast is carried out to the image after cutting, comprising: be based on formula g (x)=a*
F (x)+b adjusts brightness and contrast;Wherein, f (x) indicates that the image pixel before adjustment, g (x) indicate image slices adjusted
Element, a are the gain for controlling picture contrast, and b is the biasing for controlling brightness of image, and a, b are random in each adjustment
Value.
Preferably, in the optimizer model of adjustment third deep learning model, the optimizer model includes Adam excellent
Change device and RMSProp optimizer.
Preferably, the X value is 20.
Preferably, the value range of a is [0.5,1.5], and the value range of b is [- 50,50].
For differentiating the deep learning model construction system of corneal ulceration, comprising:
Preprocessing module obtains cornea processing image for handling cornea test image;
First screening module, for obtaining two deep learning models based on AlexNet and VGGNet training and utilizing this
Diagonal film process image is identified two deep learning models respectively, and discrimination is higher as the first deep learning mould
Type;
Second screening module, for adjusting the convolution number of plies of the first deep learning model, by before adjusting with adjusted the
Diagonal film process image is identified one deep learning model respectively, and discrimination is higher as the second deep learning model;
Third screening module, for adjusting the convolution nucleus number of the second deep learning model, by before adjusting with adjusted the
Diagonal film process image is identified two deep learning models respectively, and discrimination is higher as third deep learning model;
4th screening module, for adjusting the optimizer model of third deep learning model, by before adjusting with it is adjusted
Diagonal film process image is identified third deep learning model respectively, and discrimination is higher as final discrimination model.
Further, the preprocessing module includes:
Module is cut, for cutting to cornea test image, to extract effective coverage;The effective coverage is cornea
The region that pixel value is not zero in test image;
Enlargement module, for carrying out brightness and the random adjustment of contrast to the image after cutting, to realize sample
Expand.
Further, the cutting module includes:
Scan module is X for setting up pixel value threshold value, progressively scans pixel in cornea test image;
The ordinate of pixel of first pixel value obtained greater than X, the vertical seat of uppermost point A as in effective coverage
Mark;
The ordinate of pixel of the last one pixel value obtained greater than X, bottom point C's is vertical as in effective coverage
Coordinate;
The abscissa of pixel of first pixel value obtained greater than X, the horizontal seat of leftmost point B as in effective coverage
Mark;
The abscissa of pixel of the last one pixel value obtained greater than X, the cross of rightmost point D as in effective coverage
Coordinate;
Extraction module, for the transverse direction along A, C, the longitudinal direction of B, D are cut, to extract effective coverage.
The beneficial effects of the present invention are: handling cornea test image, to facilitate model to identify, is conducive to improve and know
Not rate;Also, different deep learning models is obtained based on AlexNet and VGGNet, and after this, by constantly adjusting
All kinds of parameters of deep learning model obtain new deep learning model, after each adjustment, with the higher depth of discrimination
Degree learning model based on this time it is adjusted be compared, therefore can for deep learning model parameters divide
It does not optimize, generally by these optimizations, discrimination can be greatly promoted.Therefore, The present invention gives the depths after optimization
The construction method of learning model is spent, this deep learning model can replace human eye differentiation and discrimination with higher, facilitate doctor
Business personnel and researcher use.
Detailed description of the invention
Present pre-ferred embodiments are provided, with reference to the accompanying drawing with the embodiment that the present invention will be described in detail.
Fig. 1 is step flow chart of the invention;
Fig. 2 is the schematic diagram that the present invention cuts cornea test image step.
Specific embodiment
Referring to Fig.1, for differentiating the construction method of the deep learning model of corneal ulceration, comprising the following steps:
Cornea test image is handled, cornea processing image is obtained;
Two deep learning models are obtained based on AlexNet and VGGNet training and utilize this two deep learning models point
Not diagonal film process image is identified, discrimination is higher as the first deep learning model;
The convolution number of plies for adjusting the first deep learning model, by distinguishing before adjusting with the first deep learning model adjusted
Diagonal film process image is identified, discrimination is higher as the second deep learning model;
For adjusting the convolution nucleus number of the second deep learning model, by adjusting preceding and the second deep learning model adjusted
Diagonal film process image is identified respectively, and discrimination is higher as third deep learning model;
For adjusting the optimizer model of third deep learning model, by adjusting preceding and third deep learning mould adjusted
Diagonal film process image is identified type respectively, and discrimination is higher as final discrimination model.
Further, cornea test image is handled, comprising:
Cornea test image is cut, to extract effective coverage;The effective coverage is picture in cornea test image
The region that plain value is not zero;
Brightness and the random adjustment of contrast are carried out to the image after cutting, to realize that sample expands.
Further, cornea test image is cut, to extract effective coverage, comprising:
Setting up pixel value threshold value is X, progressively scans pixel in cornea test image;
The ordinate of pixel of first pixel value obtained greater than X, the vertical seat of uppermost point A as in effective coverage
Mark;
The ordinate of pixel of the last one pixel value obtained greater than X, bottom point C's is vertical as in effective coverage
Coordinate;
The abscissa of pixel of first pixel value obtained greater than X, the horizontal seat of leftmost point B as in effective coverage
Mark;
The abscissa of pixel of the last one pixel value obtained greater than X, the cross of rightmost point D as in effective coverage
Coordinate;
Along the transverse direction of A, C, the longitudinal direction of B, D are cut, to extract effective coverage.
Further, the random adjustment of brightness and contrast is carried out to the image after cutting, comprising: be based on formula g (x)=a*
F (x)+b adjusts brightness and contrast;Wherein, f (x) indicates that the image pixel before adjustment, g (x) indicate image slices adjusted
Element, a are the gain for controlling picture contrast, and b is the biasing for controlling brightness of image, and a, b are random in each adjustment
Value.
Preferably, in the optimizer model of adjustment third deep learning model, the optimizer model includes Adam excellent
Change device and RMSProp optimizer.
Preferably, the X value is 20.
Preferably, the value range of a is [0.5,1.5], and the value range of b is [- 50,50].
For differentiating the deep learning model construction system of corneal ulceration, comprising:
Preprocessing module obtains cornea processing image for handling cornea test image;
First screening module, for obtaining two deep learning models based on AlexNet and VGGNet training and utilizing this
Diagonal film process image is identified two deep learning models respectively, and discrimination is higher as the first deep learning mould
Type;
Second screening module, for adjusting the convolution number of plies of the first deep learning model, by before adjusting with adjusted the
Diagonal film process image is identified one deep learning model respectively, and discrimination is higher as the second deep learning model;
Third screening module, for adjusting the convolution nucleus number of the second deep learning model, by before adjusting with adjusted the
Diagonal film process image is identified two deep learning models respectively, and discrimination is higher as third deep learning model;
4th screening module, for adjusting the optimizer model of third deep learning model, by before adjusting with it is adjusted
Diagonal film process image is identified third deep learning model respectively, and discrimination is higher as final discrimination model.
Further, the preprocessing module includes:
Module is cut, for cutting to cornea test image, to extract effective coverage;The effective coverage is cornea
The region that pixel value is not zero in test image;
Enlargement module, for carrying out brightness and the random adjustment of contrast to the image after cutting, to realize sample
Expand.
Further, the cutting module includes:
Scan module is X for setting up pixel value threshold value, progressively scans pixel in cornea test image;
The ordinate of pixel of first pixel value obtained greater than X, the vertical seat of uppermost point A as in effective coverage
Mark;
The ordinate of pixel of the last one pixel value obtained greater than X, bottom point C's is vertical as in effective coverage
Coordinate;
The abscissa of pixel of first pixel value obtained greater than X, the horizontal seat of leftmost point B as in effective coverage
Mark;
The abscissa of pixel of the last one pixel value obtained greater than X, the cross of rightmost point D as in effective coverage
Coordinate;
Extraction module, for the transverse direction along A, C, the longitudinal direction of B, D are cut, to extract effective coverage.
Further, the enlargement module carries out the random adjustment of brightness and contrast to the image after cutting, comprising: base
Brightness and contrast is adjusted in formula g (x)=a*f (x)+b;Wherein, f (x) indicates that the image pixel before adjustment, g (x) indicate to adjust
Image pixel after whole, a are the gain for controlling picture contrast, and b is the biasing for controlling brightness of image, and a, b are every
Random value in secondary adjustment.
Further, the optimizer model of third deep learning model includes Adam optimizer and RMSProp optimizer.
Further, the X value is 20.
Further, the value range of a is [0.5,1.5], and the value range of b is [- 50,50].
Specifically, referring to Fig. 2, often surrounding has the inactive area (area Ji Tu2Zhong S1 in the corneal ulceration image of acquisition
Domain), in order to eliminate inactive area to the interference of image recognition and reduce data volume, accelerates arithmetic speed, need right in image
Region that recognition effect works (i.e. effective coverage, Fig. 2 in the region S2) extracts.Since effective coverage and inactive area have
There is apparent boundary, and pixel value representated by inactive area is 0, so need to only mark four points in effective coverage (in Fig. 2
A, B, C, D point), since the pixel value magnitude range in each channel of the corneal ulceration image of acquisition is usually [0,255], base
In practical experience, it is 20 that pixel value threshold value, which is arranged,.
For the training of two kinds of convolutional neural networks structures of AlexNet and VGGNet, a kind of details of preferred training method
It is:
After extended, selecting training dataset is 6000 subtended angle film color pictures, including serious phase corneal ulceration Figure 30 00 opens, mistake
Phase corneal ulceration Figure 30 00 is crossed to open;
Having a size of 2592 × 1728, jpeg format all zooms in and out each picture of input, is scaled to original image
512 × 512 size;
It is 0.0001 that learning rate, which is arranged, train epochs 8000, and batch processing amount is 32;
Using the training method of ReLU activation primitive, GPU, verified using 5 retransposings to determine when to terminate training.
Based on above-mentioned condition, the two deep learning models that will acquire respectively test test picture, obtain
Test result is as shown in table 1:
1. two kinds of network models of table
Network type | AlexNet | VGGNet |
Test discrimination/% | 82.32 | 84.64 |
It can be seen from the results above that the learning model ratio for the automatic identification of corneal ulceration, based on VGGNet
The differentiation effect of learning model based on AlexNet is more preferable, so this programme determines to select convolutional neural networks
VGGNet, and further modification and perfect is carried out on its basis.
In the network structure of VGGNet, have 5 sections of convolution, have 2-3 convolutional layer in each section, modify to it with it is excellent
Change.During actual verification, the implicit number of nodes of the full articulamentum of traditional VGGNet is 4096, is found in training process, damage
Lose function loss decline quickly, good to training set recognition effect but undesirable in terms of the effect of test set, node mistake
More, model will appear over-fitting, it is therefore necessary to reduce implicit number of nodes.
It is debugged by many experiments, finally determines that the implicit number of nodes of full articulamentum is 120, with original classical VGGNet net
Network structure is compared, and port number reduces, and prevents the over-fitting of model, while decreasing parameter amount and calculation amount, and imply
The change of number of nodes will effectively improve discrimination, but continue to reduce implicit number of nodes then to recognition effect shadow in certain limit
Ring very little.It is specific relatively more as shown in table 2:
The change of 2. number of nodes of table is compared
Moreover, under same hardware condition and same test discrimination, the time-consuming of the VGG network structure in this programme
It substantially reduces, the iteration time of every wheel batch (batch processing number) is about 1.15s when training, and one picture time-consuming of test is about
1.05s;And as a comparison, traditional VGG network structure, the iteration time of every wheel batch is about 8.05s when training, test one
Picture time-consuming is about 6.05s;So from the aspect of comprehensive practical and efficiency, full articulamentum that this programme finally determines
Implicit number of nodes is 120.
Specific network parameter is as shown in table 3:
The setting of 3. network parameter of table
During training, the size of each batch is 32, and every training by a batch is just to the power of model
Value coefficient optimizes, and calculates the accuracy rate accuracy on corresponding loss function loss and corresponding cross validation collection.
Recognize by experiment, after the training of about 7500 steps, very little, system are basic for the amplitude of variation of loss and accuracy
It tends towards stability, model reaches convergence state.
In the case where same hardware condition and constant other parameters condition, the convolution number of plies of conventional model is 13, existing
The convolution number of plies for adjusting model carries out class test to test data set, and obtained test result is as shown in table 4:
The model discrimination of the different convolutional layers of table 4
It is found by table 4, possesses the discrimination highest of the learning model of 15 layers of convolutional layer, so the convolution number of plies of model is set
It is 15, and retains model at this time.
According to conclusion before, using the network structure with 15 layers of convolutional layer, and on this basis to the 4th section of network
The convolution nucleus number of convolutional layer is adjusted.Conventional model convolution nucleus number is 512-512-512, is now adjusted separately as 256-256-
256,356-356-356, to its 8000 step of iteration, trains corresponding three models, observes each leisure to carry out experiment comparison
The recognition effect of test set.Discrimination is as shown in table 5 below:
The model discrimination of the different convolution nucleus numbers of table 5
It is found by table 5, the discrimination highest of test set is reached using the network of 256-256-256 convolution nucleus number
88.20%, nearly 2 percentage points are higher by compared to other two models, is comprehensively considered, by the convolution of the 4th section of convolutional layer of network model
Nucleus number is set as 256-256-256, and retains model at this time.
Optimizer is a kind of majorized function for being minimized loss function loss, when data are input to convolutional neural networks
And after having calculated loss value, optimizer is needed to reduce the parameter in loss and more new model, to reach preferably training effect
Fruit.
Conventional model generally uses Adam optimizer, is now carried out using the model of RMSProp optimizer to test data set
Class test, obtained test result are as shown in table 6:
The model discrimination of 6 Different Optimization device of table
Find, using the model of RMSProp optimizer, the case where not restraining will occur, cause model to test set by table 6
Discrimination it is undesirable, so final model selection Adam optimizer, and save optimal models at this time.
In conclusion the final mask that uses of this programme is is 15, the 4th sections of convolutional layers based on VGGNet, the convolution number of plies
Convolution nucleus number is 256-256-256 and the convolutional neural networks using Adam optimizer, is up to for the discrimination of test set
88.20%.
Presently preferred embodiments of the present invention and basic principle is discussed in detail in the above content, but the invention is not limited to
Above embodiment, those skilled in the art should be recognized that also have on the premise of without prejudice to spirit of the invention it is various
Equivalent variations and replacement, these equivalent variations and replacement all fall within the protetion scope of the claimed invention.
Claims (10)
1. the construction method of the deep learning model for differentiating corneal ulceration, which comprises the following steps:
Cornea test image is handled, cornea processing image is obtained;
Two deep learning models are obtained based on AlexNet and VGGNet training and utilize this two deep learning models right respectively
Cornea processing image is identified, discrimination is higher as the first deep learning model;
The convolution number of plies for adjusting the first deep learning model, it is diagonal by distinguishing before adjusting with the first deep learning model adjusted
Film process image is identified, discrimination is higher as the second deep learning model;
For adjusting the convolution nucleus number of the second deep learning model, by distinguishing before adjusting with the second deep learning model adjusted
Diagonal film process image is identified, discrimination is higher as third deep learning model;
For adjusting the optimizer model of third deep learning model, by dividing before adjusting with third deep learning model adjusted
Not diagonal film process image is identified, discrimination is higher as final discrimination model.
2. according to claim 1 for differentiating the construction method of the deep learning model of corneal ulceration, which is characterized in that
Cornea test image is handled, comprising:
Cornea test image is cut, to extract effective coverage;The effective coverage is pixel value in cornea test image
The region being not zero;
Brightness and the random adjustment of contrast are carried out to the image after cutting, to realize that sample expands.
3. according to claim 2 for differentiating the construction method of the deep learning model of corneal ulceration, which is characterized in that
Cornea test image is cut, to extract effective coverage, comprising:
Setting up pixel value threshold value is X, progressively scans pixel in cornea test image;
The ordinate of pixel of first pixel value obtained greater than X, the ordinate of uppermost point A as in effective coverage;
The ordinate of pixel of the last one pixel value obtained greater than X, the vertical seat of bottom point C as in effective coverage
Mark;
The abscissa of pixel of first pixel value obtained greater than X, the abscissa of leftmost point B as in effective coverage;
The abscissa of pixel of the last one pixel value obtained greater than X, the horizontal seat of rightmost point D as in effective coverage
Mark;
Along the transverse direction of A, C, the longitudinal direction of B, D are cut, to extract effective coverage.
4. according to claim 2 for differentiating the construction method of the deep learning model of corneal ulceration, which is characterized in that
The random adjustment of brightness and contrast is carried out to the image after cutting, comprising: brightness is adjusted based on formula g (x)=a*f (x)+b
And contrast;Wherein, f (x) indicates that the image pixel before adjustment, g (x) indicate that image pixel adjusted, a are for control figure
The gain of image contrast, b are the biasing for controlling brightness of image, a, b random value in each adjustment.
5. according to claim 1 for differentiating the construction method of the deep learning model of corneal ulceration, which is characterized in that
In the optimizer model of adjustment third deep learning model, the optimizer model includes Adam optimizer and RMSProp excellent
Change device.
6. according to claim 3 for differentiating the construction method of the deep learning model of corneal ulceration, which is characterized in that
The X value is 20.
7. according to claim 4 for differentiating the construction method of the deep learning model of corneal ulceration, which is characterized in that
The value range of a is [0.5,1.5], and the value range of b is [- 50,50].
8. the deep learning model construction system of application construction method as claimed in claim 1 to 7, comprising:
Preprocessing module obtains cornea processing image for handling cornea test image;
First screening module, for obtaining two deep learning models based on AlexNet and VGGNet training and utilizing this two
Diagonal film process image is identified deep learning model respectively, and discrimination is higher as the first deep learning model;
Second screening module, for adjusting the convolution number of plies of the first deep learning model, by deep with adjusted first before adjusting
Diagonal film process image is identified degree learning model respectively, and discrimination is higher as the second deep learning model;
Third screening module, for adjusting the convolution nucleus number of the second deep learning model, by deep with adjusted second before adjusting
Diagonal film process image is identified degree learning model respectively, and discrimination is higher as third deep learning model;
4th screening module, for adjusting the optimizer model of third deep learning model, by adjusting preceding and third adjusted
Diagonal film process image is identified deep learning model respectively, and discrimination is higher as final discrimination model.
9. deep learning model construction system according to claim 8, which is characterized in that the preprocessing module includes:
Module is cut, for cutting to cornea test image, to extract effective coverage;The effective coverage is cornea test
The region that pixel value is not zero in image;
Enlargement module, for carrying out brightness and the random adjustment of contrast to the image after cutting, to realize that sample expands.
10. deep learning model construction system according to claim 9, which is characterized in that the cutting module includes:
Scan module is X for setting up pixel value threshold value, progressively scans pixel in cornea test image;
The ordinate of pixel of first pixel value obtained greater than X, the ordinate of uppermost point A as in effective coverage;
The ordinate of pixel of the last one pixel value obtained greater than X, the vertical seat of bottom point C as in effective coverage
Mark;
The abscissa of pixel of first pixel value obtained greater than X, the abscissa of leftmost point B as in effective coverage;
The abscissa of pixel of the last one pixel value obtained greater than X, the horizontal seat of rightmost point D as in effective coverage
Mark;
Extraction module, for the transverse direction along A, C, the longitudinal direction of B, D are cut, to extract effective coverage.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810586987.8A CN108985328A (en) | 2018-06-08 | 2018-06-08 | For differentiating the construction method and its system of the deep learning model of corneal ulceration |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810586987.8A CN108985328A (en) | 2018-06-08 | 2018-06-08 | For differentiating the construction method and its system of the deep learning model of corneal ulceration |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108985328A true CN108985328A (en) | 2018-12-11 |
Family
ID=64541055
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810586987.8A Pending CN108985328A (en) | 2018-06-08 | 2018-06-08 | For differentiating the construction method and its system of the deep learning model of corneal ulceration |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108985328A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110246158A (en) * | 2019-07-19 | 2019-09-17 | 上海交通大学医学院附属第九人民医院 | Eye illness detection device, method, electric terminal and storage medium |
CN110517219A (en) * | 2019-04-01 | 2019-11-29 | 刘泉 | A kind of corneal topography method of discrimination and system based on deep learning |
CN111134613A (en) * | 2019-11-21 | 2020-05-12 | 明灏科技(北京)有限公司 | Image recognition-based orthokeratology lens fitting method and system |
CN117764987A (en) * | 2024-02-22 | 2024-03-26 | 美迪信(天津)有限责任公司 | Cornea damage degree evaluation method, cornea damage degree evaluation device and storage medium |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106203354A (en) * | 2016-07-14 | 2016-12-07 | 南京信息工程大学 | Scene recognition method based on interacting depth structure |
CN106682578A (en) * | 2016-11-21 | 2017-05-17 | 北京交通大学 | Human face recognition method based on blink detection |
US20170140253A1 (en) * | 2015-11-12 | 2017-05-18 | Xerox Corporation | Multi-layer fusion in a convolutional neural network for image classification |
CN106709511A (en) * | 2016-12-08 | 2017-05-24 | 华中师范大学 | Urban rail transit panoramic monitoring video fault detection method based on depth learning |
CN107122597A (en) * | 2017-04-12 | 2017-09-01 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | A kind of corneal damage intelligent diagnosis system |
CN107480725A (en) * | 2017-08-23 | 2017-12-15 | 京东方科技集团股份有限公司 | Image-recognizing method, device and computer equipment based on deep learning |
-
2018
- 2018-06-08 CN CN201810586987.8A patent/CN108985328A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170140253A1 (en) * | 2015-11-12 | 2017-05-18 | Xerox Corporation | Multi-layer fusion in a convolutional neural network for image classification |
CN106203354A (en) * | 2016-07-14 | 2016-12-07 | 南京信息工程大学 | Scene recognition method based on interacting depth structure |
CN106682578A (en) * | 2016-11-21 | 2017-05-17 | 北京交通大学 | Human face recognition method based on blink detection |
CN106709511A (en) * | 2016-12-08 | 2017-05-24 | 华中师范大学 | Urban rail transit panoramic monitoring video fault detection method based on depth learning |
CN107122597A (en) * | 2017-04-12 | 2017-09-01 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | A kind of corneal damage intelligent diagnosis system |
CN107480725A (en) * | 2017-08-23 | 2017-12-15 | 京东方科技集团股份有限公司 | Image-recognizing method, device and computer equipment based on deep learning |
Non-Patent Citations (2)
Title |
---|
国家新课程教学策略研究组: "《青少年百科多媒体世界》", 31 July 2004 * |
郑宏达 等: "《人工智能视频分析 智能终端的崛起》", 31 August 2017 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110517219A (en) * | 2019-04-01 | 2019-11-29 | 刘泉 | A kind of corneal topography method of discrimination and system based on deep learning |
CN110246158A (en) * | 2019-07-19 | 2019-09-17 | 上海交通大学医学院附属第九人民医院 | Eye illness detection device, method, electric terminal and storage medium |
CN110246158B (en) * | 2019-07-19 | 2021-10-22 | 上海交通大学医学院附属第九人民医院 | Eye disease detection device, method, electronic terminal, and storage medium |
CN111134613A (en) * | 2019-11-21 | 2020-05-12 | 明灏科技(北京)有限公司 | Image recognition-based orthokeratology lens fitting method and system |
CN111134613B (en) * | 2019-11-21 | 2022-04-05 | 明灏科技(北京)有限公司 | Image recognition-based orthokeratology lens fitting method and system |
CN117764987A (en) * | 2024-02-22 | 2024-03-26 | 美迪信(天津)有限责任公司 | Cornea damage degree evaluation method, cornea damage degree evaluation device and storage medium |
CN117764987B (en) * | 2024-02-22 | 2024-04-26 | 美迪信(天津)有限责任公司 | Cornea damage degree evaluation method, cornea damage degree evaluation device and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108985328A (en) | For differentiating the construction method and its system of the deep learning model of corneal ulceration | |
CN109035255B (en) | Method for segmenting aorta with interlayer in CT image based on convolutional neural network | |
CN107506770A (en) | Diabetic retinopathy eye-ground photography standard picture generation method | |
CN106408564B (en) | A kind of method for processing fundus images based on deep learning, apparatus and system | |
CN109598733A (en) | Retinal fundus images dividing method based on the full convolutional neural networks of depth | |
CN102727200B (en) | Spinal vertebral and intervertebral disc dividing method, device, magnetic resonance imaging system | |
CN108171703B (en) | Method for automatically extracting trachea tree from chest CT image | |
CN106815596A (en) | A kind of Image Classifier method for building up and device | |
CN109087302A (en) | A kind of eye fundus image blood vessel segmentation method and apparatus | |
CN108053417A (en) | A kind of lung segmenting device of the 3DU-Net networks based on mixing coarse segmentation feature | |
CN109410204A (en) | A kind of processing of cortex cataract image and Enhancement Method based on CAM | |
CN107358262A (en) | The sorting technique and sorter of a kind of high-definition picture | |
CN109840913A (en) | The method and system of lump segmentation in a kind of mammography X | |
CN104346799B (en) | The extracting method of spinal cord in a kind of CT images | |
CN109166095A (en) | A kind of ophthalmoscopic image cup disk dividing method based on generation confrontation mechanism | |
CN103942780B (en) | Based on the thalamus and its minor structure dividing method that improve fuzzy connectedness algorithm | |
CN108053398A (en) | A kind of melanoma automatic testing method of semi-supervised feature learning | |
CN108492300A (en) | Tubular structure enhances the pulmonary vascular tree dividing method combined with energy function | |
CN108764041A (en) | The face identification method of facial image is blocked for lower part | |
CN109671094A (en) | A kind of eye fundus image blood vessel segmentation method based on frequency domain classification | |
CN102542556A (en) | Method for automatically extracting ultrasonic breast tumor image | |
CN109886946A (en) | Early senile maculopathy weakening supervision classification method based on deep learning | |
CN110443778A (en) | A method of detection industrial goods random defect | |
CN112087970A (en) | Information processing apparatus, information processing method, and computer program | |
CN101231745A (en) | Automatic partitioning method for optimizing image initial partitioning boundary |
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: 20181211 |