CN109726743A - A kind of retina OCT image classification method based on Three dimensional convolution neural network - Google Patents
A kind of retina OCT image classification method based on Three dimensional convolution neural network Download PDFInfo
- Publication number
- CN109726743A CN109726743A CN201811516615.4A CN201811516615A CN109726743A CN 109726743 A CN109726743 A CN 109726743A CN 201811516615 A CN201811516615 A CN 201811516615A CN 109726743 A CN109726743 A CN 109726743A
- Authority
- CN
- China
- Prior art keywords
- training
- layer
- convolutional layer
- model
- network
- 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
Abstract
The retina OCT image classification method based on Three dimensional convolution neural network that the invention discloses a kind of, comprising the following steps: acquired image is simultaneously divided into training set and test set by (a) acquisition image, is pre-processed to image；(b) VinceptionC3D network structure is built, VinceptionC3D network is the improvement based on C3D convolutional neural networks, the Vinception module of fusion multi-channel feature is added on the basis of C3D network, and batch standardized method is applied in original C3D network, (c) training and test of model: using the pre-training model of C3D as the pre-training model of VinceptionC3D, trained VinceptionC3D model is obtained using the network after the data training load pre-training model in training set, after model training, with test set test model.The present invention can carry out whole classification to three-dimensional retina OCT image, improve efficiency and lay the foundation with analysis for the segmentation of subsequent retina OCT image.
Description
Technical field
The invention belongs to retinal images classification methods, and in particular to a kind of retina based on Three dimensional convolution neural network
OCT image classification method.
Background technique
Retina is by the outwardly extending optic nerve distal tissues of brain, and its structure is complicated, fine, fragile and be metabolized prosperous
It contains.Its blood vessel belongs to terminal vascular system, and histanoxia caused by any pathologic destruction and angiemphraxis etc. can cause
Tissue necrosis loses the function of its impression and conduction light stimulus.
Spectral domain optical coherence tomography (OCT) is to detect light through eye subject refractive media up to retina, passes through acquisition
Thickness and range information that intraocular different tissues interface reflection provides simultaneously are reduced into image and data to work, it has also become
The indispensable clinical examination technology of ophthalmology, can provide three-dimensional retinal images.There are three types of adopt common retina OCT image
Integrated mode: macula lutea center, disk center, and the big visual field including the first two region.In order to realize it is high performance from
Whether normally dynamicization computer-aided diagnosis system is acquired region and identification to retina OCT image first and divides
Class be it is very necessary, the segmentation of subsequent retina OCT image and classification can be made more efficient.
Currently, existing retina OCT image sorting algorithm is classified based on two dimensional image, such as
The OCT image classification method based on HOG operator and support vector machines that Srinivasan et al. is proposed, can be used for age related
The classification of macular degeneration and diabetic macular edema lesion retina；Venhuizen et al. propose based on machine learning
Age-related macular degeneration lesion automatic grading system；The diabetic keratopathy view based on transfer learning that Chakraborty et al. is proposed
The swollen automatic classification with Local Electroretinogram of film water, but these algorithms cannot make full use of three-dimensional space feature to believe
Breath, single image are easily affected by noise, and lead to classification error.The existing Three dimensional convolution image applied to medical image
Sorter network has parallel supplement to be hoisted, between shortage characteristic pattern in classification accuracy, and class categories are relatively single, special
It is not normal/improper for macula lutea center of classifying simultaneously, the three kinds of scan patterns progress of disk center and the big visual field
Classification work have not been reported.
Convolutional neural networks (Convolutional Neural Networks, CNN) are a kind of comprising convolution or correlation
The feedforward neural network for calculating and having depth structure, is one of representative algorithm of deep learning.Three dimensional convolution neural network is
For carrying out video Activity recognition, medical domain often uses Three dimensional convolution neural network, existing Three dimensional convolution network knot
Structure as shown in figure 4, have 8 3D convolutional layers, 5 pond 3D layers, 2 full articulamentums, behind connect a softmax output layer,
The step-length of all 3D convolution kernels is all 1, and the size of convolution kernel is 3 × 3 × 3.All maximum pond layers are in addition to first pond
The core for changing layer is except 1 × 2 × 2, and the core of other pond layers is 2 × 2 × 2.
Summary of the invention
A kind of retina based on Three dimensional convolution neural network in order to overcome the deficiencies in the prior art, the present invention provides
OCT image classification method.
In order to solve the above-mentioned technical problem, the present invention adopts the following technical scheme:
A kind of retina OCT image classification method based on Three dimensional convolution neural network, comprising the following steps: (a) acquires image
And acquired image is divided into training set, verifying collection and test set, image is pre-processed；(b) it builds
VinceptionC3D network structure, VinceptionC3D network is the improvement based on C3D convolutional neural networks, in C3D network
On the basis of be added to the Vinception module of fusion multi-channel feature, and a batch standardized method has been applied to original
In C3D network, (c) training and test of model: using the pre-training model of C3D as the pre-training mould of VinceptionC3D
Type obtains trained VinceptionC3D model using the network after the data training load pre-training model in training set,
After model training, with test set test model.
Preferably, including to macula lutea center, disk center, the big visual field to the pretreatment of image in step (a)
Retina three-dimensional OCT image carries out down-sampling and subtracts average value processing, carries out Random Level overturning to data and carrys out EDS extended data set.
Preferably, in step (b) in VinceptionC3D network structure successively include 3D convolutional layer 1a, pond layer 1,
3D convolutional layer 2a, pond layer 2,3D convolutional layer 3a, 3D convolutional layer 3b, pond layer 3, Vinception module 1, pond layer 4,
Vinception module 2, pond layer 5, full articulamentum 6, full articulamentum 7 and softmax output layer.
Preferably, being activated behind each convolutional layer with Relu function, Relu activation front is also all added to
Normalization layer is criticized, for fixing the mean value and variance of every layer of input signal.
Preferably, Vinception module includes the convolutional layer c1, a convolutional layer c2 of two superpositions, a convolutional layer
C3 and concat layers, convolutional layer c1, convolutional layer c2, convolutional layer c3 are set side by side, convolutional layer c1, convolutional layer c2, convolutional layer c3
It is connect afterwards with concat layers, convolutional layer c4 is connected with after concat layers.
Preferably, two Vinception modules receive one layer of characteristic pattern transmitted, by Vinception module
Multiscale Fusion after, export the characteristic pattern of semantic relative abundance, then pond is followed by next layer.
Preferably, the convolution kernel of convolutional layer c1 is 3 × 3 × 3, the convolution kernel of convolutional layer c2 is 3 × 3 × 3, convolutional layer c3
Convolution kernel be 1 × 1 × 1, the convolution kernel of convolutional layer c4 is 1 × 1 × 1.
Preferably, the training and test of step (c) model method particularly includes: load is in video image first
The pre-training model trained on Sports-1M data set uses Adam optimization algorithm, setting on the basis of pre-training model
The initial learning rate of characteristic extraction part, the initial learning rate of output layer, minimum batch, highest the number of iterations, by the figure of training set
As being sequentially inputted to be trained in model, after the every iteration n times of training set, is verified with verifying collection, obtained after the completion of training
Final training pattern is simultaneously tested with test set；Cross validation is rolled over to assess training pattern using K on entire data set
Performance.
Preferably, the initial learning rate of characteristic extraction part is 1e-4 in step (c), the initial learning rate of output layer is
1e-3, minimum batch are set as 30, and the number of iterations is 6000 times, every iteration 10 times, primary with verifying collection verifying, when test
The performance of training pattern is assessed on entire data set using 3 folding cross validations.
The beneficial effects of the present invention are:
1, the big visual field, disk center, three kinds of macula lutea center retina OCT scan are directed to present invention firstly provides a kind of
Normal/improper three-dimensional automatic classification method --- the VinceptionC3D neural network of image, can be to three-dimensional retina OCT
Image carries out whole classification, improves efficiency and lays the foundation with analysis for the segmentation of subsequent retina OCT image.
2, the network architecture of a multiple dimensioned three-dimensional feature figure of fusion is proposed in the present invention, and it is named as
On the one hand the local sparsity structure of optimization can be carried out approximate, another aspect with approximate intensive ingredient by Vinception module
Each layer of convolution nucleus number can be increased, this makes the width for increasing network, but does not consume computing resource excessively again, simultaneously
The multiple dimensioned feature of this network integration, allows next layer network to possess the abstract characteristics of different scale.
3,1 × 1 × 1 Three dimensional convolution layer with across channel interaction and information integration is placed on by Vinception module
Concat layers below, not only can use the pre-training model Sports-1M trained on the video images, but also can merge more
The feature in channel reduces number of parameters, prevents over-fitting occur during training and test.
4, there are 4 3D convolutional layers in overall model structure of the invention, 5 pond 3D layers connect below in 4 3D convolution
It is 2 Vinception modules, the two Vinception modules receive one layer of characteristic pattern transmitted, by Vinception
After the Multiscale Fusion of layer, the characteristic pattern of semantic relative abundance is exported, then pond is followed by next layer, and pond is in order to obtain more
Wide receptive field, Vinception module are put behind, and can be merged multiple dimensioned three-dimensional feature figure high-rise in this way, be avoided
More image detail features are lost in high level.Vinception module is only applied in high level, and parameter can be prevented excessive.
5, in the present invention, activation primitive uses Relu function, behind each convolutional layer, before Relu active coating all
It is added to batch normalization layer, to fix the mean value and variance of every layer of input signal, so that training data and test data meet together
Distribution, overcomes that network convergence rate is slow and gradient disperse problem.
Detailed description of the invention
Fig. 1 is the schematic network structure of VinceptionC3D in the present invention；
Fig. 2 is the schematic network structure of Vinception module in the present invention；
Fig. 3 is the position view for the three-dimensional retinal images classified in the present invention；
Fig. 4 is the schematic network structure of Three dimensional convolution neural network in the prior art.
Specific embodiment
Present invention will be further described below with reference to the accompanying drawings and specific embodiments:
A kind of retina OCT image classification method based on Three dimensional convolution neural network, comprising the following steps: (a) acquires image
And acquired image is divided into training set, verifying collection and test set, image is pre-processed；(b) it builds
VinceptionC3D network structure, VinceptionC3D network is the improvement based on C3D convolutional neural networks, in C3D network
On the basis of be added to the Vinception module of fusion multi-channel feature, and a batch standardized method has been applied to original
In C3D network, (c) training and test of model: using the pre-training model of C3D as the pre-training mould of VinceptionC3D
Type obtains trained VinceptionC3D model using the network after the data training load pre-training model in training set,
After model training, with the excellent performance to evaluate the invention of test set test model.
It include to macula lutea center, disk center, big visual field retina three-dimensional to the pretreatment of image in step (a)
OCT image carries out down-sampling and subtracts average value processing, carries out Random Level overturning to data and carrys out EDS extended data set.
It successively include 3D convolutional layer 1a, pond layer 1,3D convolutional layer in VinceptionC3D network structure in step (b)
2a, pond layer 2,3D convolutional layer 3a, 3D convolutional layer 3b, pond layer 3, Vinception module 1, pond layer 4, Vinception mould
Block 2, pond layer 5, full articulamentum 6, full articulamentum 7 and softmax output layer.
It is activated behind each convolutional layer with Relu function, Relu activation front is also all added to batch standardization
Layer, for fixing the mean value and variance of every layer of input signal.
Vinception module include two superposition convolutional layer c1, a convolutional layer c2, a convolutional layer c3 and
Concat layers, convolutional layer c1, convolutional layer c2, convolutional layer c3 are set side by side, after convolutional layer c1, convolutional layer c2, convolutional layer c3 with
Concat layers connect, and are connected with convolutional layer c4 after concat layers.
Two Vinception modules receive one layer of characteristic pattern transmitted, and multiple dimensioned by Vinception module is melted
After conjunction, the characteristic pattern of semantic relative abundance is exported, then pond is followed by next layer.
The convolution kernel of convolutional layer c1 is 3 × 3 × 3, and the convolution kernel of convolutional layer c2 is 3 × 3 × 3, the convolution kernel of convolutional layer c3
It is 1 × 1 × 1, the convolution kernel of convolutional layer c4 is 1 × 1 × 1.
The training of step (c) model and test method particularly includes: load is in video image Sports-1M data set first
On the pre-training model trained, Adam optimization algorithm is used on the basis of pre-training model, at the beginning of characteristic extraction part is set
The image of training set is sequentially inputted to mould by beginning learning rate, the initial learning rate of output layer, minimum batch, highest the number of iterations
Be trained in type, after every iteration n times, with verifying collection verified, training after the completion of obtain final training pattern and with survey
Examination collection is tested；Cross validation is rolled over to assess the performance of training pattern using K on entire data set.
In step (c), the initial learning rate of characteristic extraction part is 1e-4, and the initial learning rate of output layer is 1e-3, minimum
Batch is set as 30, and the number of iterations is 6000 times, every iteration 10 times, primary with verifying collection verifying, in entire data set when test
The upper performance that training pattern is assessed using 3 folding cross validations.
Embodiment 1
(1) acquisition and pretreatment of image:
According to the needs of clinical ophthalmology, retina OCT image often has 3 kinds of scan patterns: macula lutea center, disk center with
And the big visual field, as shown in figure 3, (a) is retina entirety OCT image, the big view of visual representation on the color picture image of eyeground in figure
Wild region, disk area, the position of macula lutea central area correspond respectively to (b) in figure, (c), (d).It can will adopt
The three-dimensional retina OCT image collected is divided into 6 classes: improper big field-of-view image (ANW), normal big field-of-view image (NW), just
Normal macula lutea center image (ANM), normal macula image (NM), improper disk images (ANO), normal optic nerve cream
Head image (NO).
Used in this example from 873 of 671 subjects three-dimensional retina OCT images as data set, for training with
Assessment, wherein improper big field-of-view image (ANW) 24, normal big field-of-view image (NW) 37, improper macula lutea center image
(ANM) 427, normal macula center image (NM) 131, improper disk center image (ANO) 121, normally
Disk center image (NO) 133.Since the three-dimensional retina OCT image of acquisition has different resolution ratio, have
Different scan settings and resolution ratio.For uniform size size, and the calculating cost of learning tasks is reduced, all three-dimensional OCT
Image is all down sampled to 96 × 96 × 16 using the method for bilinear interpolation, then carries out subtracting average value processing to improve difference degree.
In order to equilibrium data and prevent over-fitting, pair unless 5 class little data of normal macula center image are carried out with 30%
The mode of probability level overturning realize data extending.
(2) VinceptionC3D network structure is built:
The model of the present embodiment is to carry out on Three dimensional convolution neural network C3D model improved, and Fig. 4 illustrates original C3D volumes
Product neural network structure.The network is the Three dimensional convolution neural network of a progress video Activity recognition, has 8 3D convolution
Layer, 5 pond 3D layers, 2 full articulamentums, behind connect a softmax output layer, the step-length of all 3D convolution kernels is all 1,
The size of convolution kernel is 3 × 3 × 3.All maximum pond layers in addition to the core of first pond layer be 1 × 2 × 2 other than, other
The core of pond layer is 2 × 2 × 2, in order to prematurely merge depth information, but also the length of available 16 frame so as to
It prepares for subsequent down-sampling.
The network architecture of a multiple dimensioned three-dimensional feature figure of fusion is proposed in the present embodiment, as shown in Figure 1, and ordering it
Entitled Vinception module, this module have used for reference the wind of inception module in two dimensional image sorter network GoogLeNet
Lattice.Inception module can the local sparsity structure of optimization with approximate intensive ingredient come approximate, on the other hand can increase
Add each layer of convolution nucleus number, this makes the width for increasing network, but does not consume computing resource excessively again, while this net
Network has merged multiple dimensioned feature, and next layer network is allowed to possess the abstract characteristics of different scale.The present embodiment is proposed
Vinception module effect it is as follows:
(1) Vinception module uses Three dimensional convolution, as shown in Fig. 2, its interlaminar action is respectively: two 3 × 3 × 3 volumes
Lamination superposition, is equivalent to one 5 × 5 × 5 convolutional layer, available wider array of receptive field, independent one 3 × 3 × 3 convolution
The available relatively small receptive field of layer, similarly 1 × 1 × 1 convolutional layer obtains smaller receptive field, different semantic features
Parallel Fusion can increase the understanding to Feature Semantics, so that the Feature Semantics finally exported are more abundant, add again after concat
One 1 × 1 × 1 convolutional layer is added, such benefit is can to increase non-linear layer, so that network expresses more samples, together
When reduce number of parameters.In this example, the step-length of above-mentioned each convolutional layer is 1 × 1 × 1.
It (2) is that Vinception module will have and interact across channel and information is whole with original inception module difference
1 × 1 × 1 Three dimensional convolution layer of cooperation be placed on concat layers below, both can use trained on the video images it is pre-
Training pattern Sports-1M, and the feature of multichannel can be merged, number of parameters is reduced, is prevented in training and the process tested
In there is over-fitting.
Fig. 1 illustrates the overall model structure of the present embodiment, which has 4 3D convolutional layers, the volume of this 4 3D convolutional layers
Product core is 3 × 3 × 3,5 pond 3D layers, and the core of first pond layer is 1 × 2 × 2, and the core of remaining pond layer is all 2 × 2
× 2, the step-length of first pond layer is 1 × 2 × 2, and the step-length of remaining pond layer is all 2 × 2 × 2；It is connect behind 4 3D convolution
Be 2 described Vinception modules above, the two Vinception modules receive one layer of characteristic pattern transmitted, pass through
After crossing Vinception layers of Multiscale Fusion, the characteristic pattern of semantic relative abundance is exported, then pond is followed by next layer, Chi Hua
It is that wider array of receptive field, Vinception module put behind and be in order to obtain, multiple dimensioned three can be merged high-rise in this way
Dimensional feature figure avoids and loses more image detail features in high level.Parameter is excessive in order to prevent simultaneously, only in high level
Using.
In the present embodiment, activation primitive uses Relu function, behind each convolutional layer, before Relu active coating all
It is added to batch normalization layer, to fix the mean value and variance of every layer of input signal, so that training data and test data meet together
Distribution, overcomes that network convergence rate is slow and gradient disperse problem.The class categories number that full articulamentum exports below is 6, net
The three-dimensional OCT retinal images size of network input is 96 × 96 × 16, and the content of arrow meaning is to pass through each Three dimensional convolution layer
Later and the 3-D image size of Chi Huahou.
(3) training and test of model:
The performance of the present embodiment is assessed using 3 folding cross validations on entire data set.Load is in video image first
The pre-training model trained on Sports-1M data set uses Adam optimization algorithm, feature on the basis of pre-training model
Extracting the initial learning rate in part is 1e-4, and the initial learning rate of output layer is 1e-3, and minimum batch is set as 30, at 3
Iteration 6000 times on Tesla K40m GPU, every iteration 10 times verifyings are primary, obtain final training pattern and carry out test set
Test.
(4) experimental result:
Initial three-dimensional convolutional neural networks are assessed respectively and compared to the present embodiment using 3 folding cross validations in entire data set
VinceptionC3D network proposed in C3D and the present embodiment.The corresponding confusion matrix of experimental result as shown in table 1, table 2,
TransferC3D indicates that initial three-dimensional convolutional neural networks C3D, VinceptionC3D indicate the present embodiment.It can be seen that original
Three dimensional convolution neural network C3D classification accuracy is 92.09%, and after improvement namely the classification accuracy of the present embodiment is reachable
94.04%。
Table 1, table 2 illustrate the confusion matrix of 6 classification (ANW, NW, ANM, NM, ANO, NO) of both models.
Classification accuracy of all categories in the present embodiment in addition to a kind of classification be equal to improve before accuracy rate, other classifications are better than
Classification accuracy before improvement.In our data set, ANO OCT image is mainly from glaucoma patient.In being with ONH
The retinopathy image and normal picture of the heart are much like, this can lead to wrong classification rate relatively high between ANO and NO, phase
Instead, other retinas are easier to be distinguished from each other.Obscuring for ANO and NO contributes to a biggish share to whole error rate.It is comprehensive
From the point of view of closing, the present embodiment can carry out the classification of three-dimensional retina OCT image in this task, can regard to individual three-dimensional in test
Nethike embrane OCT image makes good judgement, and obtains one better than the good performance before improving, and is more conducive to three-dimensional view
The classification of film OCT image.
The confusion matrix of 1 initial three-dimensional convolutional neural networks C3D of table
The confusion matrix of table 2VinceptionC3D
So far, a kind of to be directed to the big visual field, disk center, three kinds of macula lutea center retina OCT scan image is normal/non-just
Normal three-dimensional automatic classification method --- VinceptionC3D neural network has been carried out and is verified.It is in an experiment
Performance is better than original C3D Three dimensional convolution neural network, and the invention, which can make three-dimensional retina OCT image, more preferably to be judged, from
For another aspect, the network structure of the present embodiment is simultaneously uncomplicated, and number of parameters is few, and Vinception module is utilized and makes
Network structure more efficiently and robust, facilitates the classification and detection of three-dimensional retina OCT image, substantially increases three-dimensional view
The screening efficiency of film OCT image.The present embodiment combines the acquisition and pretreatment, VinceptionC3D network model of image
Build with training and test, make the research of subsequent retinal disease, such as layering segmentation, lesion segmentation, registration obtain very greatly
It helps.
The technical means disclosed in the embodiments of the present invention is not limited only to technological means disclosed in above embodiment, further includes
Technical solution consisting of any combination of the above technical features.It should be pointed out 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 (9)
1. a kind of retina OCT image classification method based on Three dimensional convolution neural network, which is characterized in that including following step
Suddenly :(a) acquired image is simultaneously divided into training set, verifying collection and test set by acquisition image, is pre-processed to image；(b)
VinceptionC3D network structure is built, VinceptionC3D network is the improvement based on C3D convolutional neural networks, in C3D
It is added to the Vinception module of fusion multi-channel feature on the basis of network, and batch standardized method has been applied to original
In beginning C3D network, (c) training and test of model: using the pre-training model of C3D as the pre-training of VinceptionC3D
Model obtains trained VinceptionC3D mould using the network after the data training load pre-training model in training set
Type, after model training, with test set test model.
2. the retina OCT image classification method based on Three dimensional convolution neural network, feature exist as described in claim 1
In, in step (a) to the pretreatment of image include to macula lutea center, disk center, the big visual field retina three-dimensional OCT figure
As carrying out down-sampling and subtracting average value processing, Random Level overturning is carried out to data and carrys out EDS extended data set.
3. the retina OCT image classification method based on Three dimensional convolution neural network, feature exist as described in claim 1
In successively including 3D convolutional layer 1a, pond layer 1,3D convolutional layer 2a, Chi Hua in VinceptionC3D network structure in step (b)
Layer 2,3D convolutional layer 3a, 3D convolutional layer 3b, pond layer 3, Vinception module 1, pond layer 4, Vinception module 2, pond
Change layer 5, full articulamentum 6, full articulamentum 7 and softmax output layer.
4. the retina OCT image classification method based on Three dimensional convolution neural network, feature exist as claimed in claim 3
In being activated behind each convolutional layer with Relu function, Relu activation front is also all added to batch normalization layer, is used to
The mean value and variance of fixed every layer of input signal.
5. the retina OCT image classification side based on Three dimensional convolution neural network as described in claims 1 to 4 any one
Method, which is characterized in that Vinception module includes the convolutional layer c1, a convolutional layer c2 of two superpositions, a convolutional layer c3
And concat layers, convolutional layer c1, convolutional layer c2, convolutional layer c3 are set side by side, after convolutional layer c1, convolutional layer c2, convolutional layer c3
It is connect with concat layers, convolutional layer c4 is connected with after concat layers.
6. the retina OCT image classification method based on Three dimensional convolution neural network, feature exist as claimed in claim 5
In, two Vinception modules receive one layer of characteristic pattern transmitted, after the Multiscale Fusion of Vinception module,
The characteristic pattern of semantic relative abundance is exported, then pond is followed by next layer.
7. the retina OCT image classification method based on Three dimensional convolution neural network, feature exist as claimed in claim 5
In the convolution kernel of convolutional layer c1 is 3 × 3 × 3, and the convolution kernel of convolutional layer c2 is 3 × 3 × 3, and the convolution kernel of convolutional layer c3 is 1 × 1
The convolution kernel of × 1, convolutional layer c4 are 1 × 1 × 1.
8. the retina OCT image classification method based on Three dimensional convolution neural network, feature exist as described in claim 1
In the training and test of, step (c) model method particularly includes: first load is instructed on video image Sports-1M data set
The pre-training model practiced, uses Adam optimization algorithm on the basis of pre-training model, and setting characteristic extraction part is initially learned
The image of training set is sequentially inputted in model by habit rate, the initial learning rate of output layer, minimum batch, highest the number of iterations
It is trained, after every iteration n times, is verified with verifying collection, obtain final training pattern after the completion of training and use test set
It is tested；Cross validation is rolled over to assess the performance of training pattern using K on entire data set.
9. the retina OCT image classification method based on Three dimensional convolution neural network, feature exist as claimed in claim 8
In in step (c), the initial learning rate of characteristic extraction part is 1e-4, and the initial learning rate of output layer is 1e-3, minimum batch
It is set as 30, the number of iterations is 6000 times, every iteration 10 times, primary with verifying collection verifying, and when test adopts on entire data set
The performance of training pattern is assessed with 3 folding cross validations.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811516615.4A CN109726743B (en) | 2018-12-12 | 2018-12-12 | Retina OCT image classification method based on three-dimensional convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811516615.4A CN109726743B (en) | 2018-12-12 | 2018-12-12 | Retina OCT image classification method based on three-dimensional convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109726743A true CN109726743A (en) | 2019-05-07 |
| CN109726743B CN109726743B (en) | 2023-03-24 |
Family
ID=66295279
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811516615.4A Active CN109726743B (en) | 2018-12-12 | 2018-12-12 | Retina OCT image classification method based on three-dimensional convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109726743B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111583291A (en) * | 2020-04-20 | 2020-08-25 | 中山大学 | Layer segmentation method and system for retina layer and effusion region based on deep learning |
| CN111598894A (en) * | 2020-04-17 | 2020-08-28 | 哈尔滨工业大学 | Retina blood vessel image segmentation system based on global information convolution neural network |
| CN111680741A (en) * | 2020-06-04 | 2020-09-18 | 南京理工大学 | Automatic debugging method of computer-aided interferometer based on deep learning |
| CN111696109A (en) * | 2020-05-25 | 2020-09-22 | 深圳大学 | High-precision layer segmentation method for retina OCT three-dimensional image |
| CN111796180A (en) * | 2020-06-23 | 2020-10-20 | 广西电网有限责任公司电力科学研究院 | Automatic identification method and device for mechanical fault of high-voltage switch |
| CN111862009A (en) * | 2020-07-02 | 2020-10-30 | 清华大学深圳国际研究生院 | Classification method of fundus OCT images and computer-readable storage medium |
| CN112132782A (en) * | 2020-08-21 | 2020-12-25 | 广东省人民医院 | Method and terminal for processing DME typing based on deep neural network |
| CN112257709A (en) * | 2020-10-23 | 2021-01-22 | 北京云杉世界信息技术有限公司 | Signboard photo auditing method and device, electronic equipment and readable storage medium |
| CN112990315A (en) * | 2021-03-17 | 2021-06-18 | 北京大学 | 3D shape image classification method of equal-variation 3D convolution network based on partial differential operator |
| CN117710373A (en) * | 2024-02-05 | 2024-03-15 | 中国科学院宁波材料技术与工程研究所 | Anti-interference OCTA retinal vessel extraction method and device |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107437092A (en) * | 2017-06-28 | 2017-12-05 | 苏州比格威医疗科技有限公司 | The sorting algorithm of retina OCT image based on Three dimensional convolution neutral net |
-
2018
- 2018-12-12 CN CN201811516615.4A patent/CN109726743B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107437092A (en) * | 2017-06-28 | 2017-12-05 | 苏州比格威医疗科技有限公司 | The sorting algorithm of retina OCT image based on Three dimensional convolution neutral net |
Non-Patent Citations (1)
| Title |
|---|
| JO˜AO CARREIRA等: "Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset", 《ARXIV》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111598894B (en) * | 2020-04-17 | 2021-02-09 | 哈尔滨工业大学 | Retina blood vessel image segmentation system based on global information convolution neural network |
| CN111598894A (en) * | 2020-04-17 | 2020-08-28 | 哈尔滨工业大学 | Retina blood vessel image segmentation system based on global information convolution neural network |
| CN111583291A (en) * | 2020-04-20 | 2020-08-25 | 中山大学 | Layer segmentation method and system for retina layer and effusion region based on deep learning |
| CN111583291B (en) * | 2020-04-20 | 2023-04-18 | 中山大学 | Layer segmentation method and system for retina layer and effusion region based on deep learning |
| CN111696109A (en) * | 2020-05-25 | 2020-09-22 | 深圳大学 | High-precision layer segmentation method for retina OCT three-dimensional image |
| CN111680741A (en) * | 2020-06-04 | 2020-09-18 | 南京理工大学 | Automatic debugging method of computer-aided interferometer based on deep learning |
| CN111680741B (en) * | 2020-06-04 | 2022-09-20 | 南京理工大学 | Automatic debugging method of computer-aided interferometer based on deep learning |
| CN111796180A (en) * | 2020-06-23 | 2020-10-20 | 广西电网有限责任公司电力科学研究院 | Automatic identification method and device for mechanical fault of high-voltage switch |
| CN111862009A (en) * | 2020-07-02 | 2020-10-30 | 清华大学深圳国际研究生院 | Classification method of fundus OCT images and computer-readable storage medium |
| CN111862009B (en) * | 2020-07-02 | 2024-01-16 | 清华大学深圳国际研究生院 | Classifying method of fundus OCT (optical coherence tomography) images and computer readable storage medium |
| CN112132782A (en) * | 2020-08-21 | 2020-12-25 | 广东省人民医院 | Method and terminal for processing DME typing based on deep neural network |
| CN112132782B (en) * | 2020-08-21 | 2023-09-05 | 广东省人民医院 | Method and terminal for processing DME (DME) typing based on deep neural network |
| CN112257709A (en) * | 2020-10-23 | 2021-01-22 | 北京云杉世界信息技术有限公司 | Signboard photo auditing method and device, electronic equipment and readable storage medium |
| CN112257709B (en) * | 2020-10-23 | 2024-05-07 | 北京云杉世界信息技术有限公司 | Signboard photo auditing method and device, electronic equipment and readable storage medium |
| CN112990315A (en) * | 2021-03-17 | 2021-06-18 | 北京大学 | 3D shape image classification method of equal-variation 3D convolution network based on partial differential operator |
| CN112990315B (en) * | 2021-03-17 | 2023-10-20 | 北京大学 | 3D shape image classification method of constant-variation 3D convolution network based on partial differential operator |
| CN117710373A (en) * | 2024-02-05 | 2024-03-15 | 中国科学院宁波材料技术与工程研究所 | Anti-interference OCTA retinal vessel extraction method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109726743B (en) | 2023-03-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109726743A (en) | A kind of retina OCT image classification method based on Three dimensional convolution neural network | |
| Rasti et al. | Macular OCT classification using a multi-scale convolutional neural network ensemble | |
| CN107423571B (en) | Diabetic retinopathy identifying system based on eye fundus image | |
| CN107358605B (en) | The deep neural network apparatus and system of diabetic retinopathy for identification | |
| CN107437092B (en) | The classification method of retina OCT image based on Three dimensional convolution neural network | |
| CN106682389B (en) | A kind of Eye disease for monitoring hypertension initiation is health management system arranged | |
| CN110197493A (en) | Eye fundus image blood vessel segmentation method | |
| CN107680683A (en) | A kind of AI eye healths appraisal procedure | |
| CN108537801A (en) | Based on the retinal angiomatous image partition method for generating confrontation network | |
| CN108765422A (en) | A kind of retinal images blood vessel automatic division method | |
| CN109376767A (en) | Retina OCT image classification method based on deep learning | |
| Sertkaya et al. | Diagnosis of eye retinal diseases based on convolutional neural networks using optical coherence images | |
| CN109509552A (en) | A kind of mental disease automatic distinguishing method of the multi-level features fusion based on function connects network | |
| CN106780439B (en) | A method of screening eye fundus image | |
| CN114548158B (en) | Data processing method for blood sugar prediction | |
| Paul et al. | Octx: Ensembled deep learning model to detect retinal disorders | |
| CN110013216A (en) | A kind of artificial intelligence cataract analysis system | |
| Jain et al. | Rider manta ray foraging optimization-based generative adversarial network and CNN feature for detecting glaucoma | |
| CN115409764A (en) | Multi-mode fundus blood vessel segmentation method and device based on domain self-adaptation | |
| CN110503636A (en) | Parameter regulation means, lesion prediction technique, parameter adjustment controls and electronic equipment | |
| CN111951950A (en) | Three-dimensional data medical classification system, method and device based on deep learning | |
| CN109009074A (en) | A kind of sudden cardiac death auxiliary prior-warning device based on deep learning | |
| CN115036019A (en) | Physical and mental health state evaluation method based on CNN-XGboost model and EEGECG | |
| Kazi et al. | Processing retinal images to discover diseases | |
| Taş et al. | Detection of retinal diseases from ophthalmological images based on convolutional neural network architecture. |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |