CN107203778A - PVR intensity grade detecting system and method - Google Patents

Abstract

The invention discloses a system and method for detecting the degree of retinopathy. The method includes: after receiving the retinopathy picture to be identified, using a predetermined recognition model to identify the received retinopathy picture, and outputting the recognition result ; Wherein, the predetermined recognition model is a convolutional neural network model obtained by training a preset number of sample pictures marked with different grades of retinopathy in advance; according to the mapping between the predetermined recognition results and grades of retinopathy relationship to determine the degree of retinopathy corresponding to the output recognition result. The present invention does not need to perform complex feature extraction calculations on eye images, is simpler, and can determine the corresponding grades of different retinopathy degrees according to the recognition results, and can effectively perform fine-grained recognition of the degree of retinopathy of patients.

Description

Retinopathy grade detection system and method

technical field

The invention relates to the field of computer technology, in particular to a system and method for detecting the degree of retinal disease.

Background technique

According to a survey of more than 93 million labor force populations in developed countries, diabetic retinopathy is a leading cause of eye blindness. Currently, the identification of diabetic retinopathy usually requires feature extraction on eye images (for example, ocular vascular structure , optic disc, retinal central groove and other feature extraction), the algorithm of feature extraction is complex and has poor performance.

Contents of the invention

The main purpose of the present invention is to provide a system and method for detecting the degree of retinopathy, aiming at finely identifying the degree of retinopathy of a patient simply and effectively.

In order to achieve the above object, the present invention provides a grade detection system for the degree of retinopathy, which includes:

The recognition module is configured to, after receiving the retinopathy picture to be recognized, use a predetermined recognition model to recognize the received retinopathy picture, and output a recognition result; A convolutional neural network model obtained by training a preset number of sample pictures marked with different levels of retinopathy;

The determination module is configured to determine the degree of retinopathy corresponding to the output recognition result according to the predetermined mapping relationship between the recognition result and the degree of retinopathy.

Preferably, the training process of the predetermined recognition model is as follows:

A. Set a corresponding preset number of sample pictures for each preset degree of retinopathy, and mark the corresponding degree of retinopathy for each sample picture;

B. Perform image preprocessing on each sample image to obtain a training image to be trained by the model;

C, divide all training pictures into the training set of the first proportion and the verification set of the second proportion;

D. Using the training set to train the predetermined recognition model;

E. Use the verification set to verify the accuracy of the trained recognition model. If the accuracy is greater than or equal to the preset accuracy, the training ends. Or, if the accuracy is less than the preset accuracy, then increase the degree of retinopathy corresponding to each grade. number of sample images and re-execute steps B, C, D, and E above.

Preferably, said step B includes:

Scaling the shorter side length of each sample picture to a first preset size to obtain a corresponding first picture, and randomly cropping a second picture of a second preset size on each first picture;

Adjusting each predetermined preset type parameter value of each second picture to a corresponding standard parameter value according to a standard parameter value corresponding to each predetermined preset type parameter, so as to obtain a corresponding third picture;

Perform a flip operation in a preset direction on each third picture, and perform a twist operation on each third picture according to a preset twist angle to obtain a fourth picture corresponding to each third picture, and use each fourth picture as a model to be trained training pictures.

Preferably, the deep convolutional neural network model includes an input layer and a plurality of network layers, and the network layers include a convolution layer, a pooling layer, a fully connected layer and a classifier layer, and the corresponding activation functions of each of the network layers f(x) is:

f(x)=max(α*x,0)

Wherein, α is a preset leakage rate, and x represents a numerical input of a neuron in the deep convolutional neural network model.

Preferably, the cross-entropy H(P, Q) corresponding to each of the network layers is:

H(P,Q)＝H(P)+D _KL (P||Q)

Among them, P and Q are two probability distributions, H(P) is the expectation of probability distribution P, H(P)=-Σ _x∈X P(x)logP(x), x is the sample space X of probability distribution P Any sample in , P(x) represents the probability that sample x is selected; the expression of D _KL (P||Q) is x is any sample in the common sample space X of probability distribution P and Q, P(x) represents the probability that sample x is selected on probability distribution P, and Q(x) represents the probability that sample x is selected on probability distribution Q.

Preferably, the scoring function of the predetermined recognition model for:

in, O _{i, j} represents the number of pictures that are actually predicted as i for the first time and j for the second time, O represents an N*N matrix, O _{i, j} represents the matrix elements in matrix O, and N represents the number of pictures participating in the prediction Number of pictures, prediction result i, j ∈ {0, 1, 2, 3, 4}, E _{i, j} represents the number of images that should appear for the first prediction as i and the second prediction as j, E is the expected prediction The N*N matrix of the result, E _i,j represent the matrix elements in the matrix E.

Preferably, the cross-entropy loss function L(x, :W) for:

Among them, x represents the input of the model, Represents the label corresponding to the input, W represents the preset model parameters, X represents the model input space, f(x:W) represents the output of the model after transforming the input x, ζ represents the normalization factor, ||W|| ² means to sum the matrix elements:

W _i+1 ＝W _i +ΔW _i+1

Among them, ΔW _i+1 represents the update increment of the weight matrix at time i+1, α is the potential energy item, β is the weight decay coefficient, γ is the learning rate of the model, and W _i represents the state value of the weight matrix at time i , _Di represents the i-th batch of input, Indicates the average gradient corresponding to the i-th batch of inputs.

Preferably, the predetermined identification model includes at least one fully connected layer, the initial value of each weight in the predetermined identification model is determined by random sampling from a preset weight range, and the connection weight of the fully connected layer The probability of being discarded is set to a first preset value, the weight decay coefficient in the cross-entropy loss function is set to a second preset value, and the potential energy item in the cross-quotient loss function is set to a third preset value.

In addition, in order to achieve the above object, the present invention also provides a method for detecting the degree of retinopathy, the method comprising the following steps:

After receiving the picture of retinopathy to be identified, use a predetermined recognition model to identify the received picture of retinopathy, and output the recognition result; The convolutional neural network model obtained by training the preset number of sample pictures of degree level;

According to the predetermined mapping relationship between the recognition result and the degree of retinopathy, the degree of retinopathy corresponding to the output recognition result is determined.

Preferably, the training process of the predetermined recognition model is as follows:

A. Set a corresponding preset number of sample pictures for each preset degree of retinopathy, and mark the corresponding degree of retinopathy for each sample picture;

B. Perform image preprocessing on each sample image to obtain a training image to be trained by the model;

C, divide all training pictures into the training set of the first proportion and the verification set of the second proportion;

D. Using the training set to train the predetermined recognition model;

E. Use the verification set to verify the accuracy of the trained recognition model. If the accuracy is greater than or equal to the preset accuracy, the training ends. Or, if the accuracy is less than the preset accuracy, then increase the degree of retinopathy corresponding to each grade. number of sample images and re-execute steps B, C, D, and E above.

The retinopathy degree level detection system and method proposed by the present invention recognize the received retinopathy pictures through the deep convolutional neural network model obtained by training based on a preset number of sample pictures marked with different retinopathy degree levels, and According to the identification result, the corresponding degree of retinal disease is determined. Since it only needs to recognize the received retinal lesion pictures according to the pre-trained deep convolutional neural network model, it does not need to perform complex feature extraction operations on eye images, which is simpler and can determine the corresponding different retinal lesions according to the recognition results The level of degree can effectively identify the degree of retinopathy of the patient in a fine-grained manner.

Description of drawings

Fig. 1 is a schematic flow chart of an embodiment of the detection method of the degree of retinopathy of the present invention;

FIG. 2 is a schematic diagram of the operating environment of a preferred embodiment of the retinopathy degree grade detection system 10 of the present invention;

FIG. 3 is a schematic diagram of functional modules of an embodiment of the system for detecting the degree of retinopathy according to the present invention.

The realization of the purpose of the present invention, functional characteristics and advantages will be further described in conjunction with the embodiments and with reference to the accompanying drawings.

detailed description

In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer and clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

The invention provides a detection method for grades of retinal lesions.

Referring to FIG. 1 , FIG. 1 is a schematic flowchart of an embodiment of the method for detecting the degree of retinopathy in the present invention.

In one embodiment, the detection method of the degree of retinopathy includes:

Step S10, after receiving the retinopathy picture to be identified, use a predetermined recognition model to identify the received retinopathy picture, and output the recognition result; A deep convolutional neural network model obtained by training a preset number of sample pictures of different degrees of retinopathy.

In this embodiment, the retinopathy degree level detection system receives the retinopathy level detection request sent by the user including the retinopathy picture to be identified, for example, receives the retinopathy level level detection request sent by the user through a terminal such as a mobile phone, a tablet computer, and a self-service terminal device. Level detection request, such as receiving the level detection request of the degree of retinopathy sent by the user on the client pre-installed in the mobile phone, tablet computer, self-service terminal equipment, etc. The retinopathy degree grade detection request sent from the browser system.

Retinopathy degree level detection system, after receiving the user's retinopathy degree level detection request, uses the pre-trained recognition model to identify the received retinopathy pictures to be identified, and recognizes the received retinopathy level to be identified The recognition result of the image in the recognition model. The identification model can be continuously trained, learned, verified, optimized, etc. by identifying a large number of preset sample pictures marked with different levels of retinal lesions in advance, so as to train it to accurately identify the corresponding levels of different levels of retinal lesions. Annotated model of . For example, the recognition model may adopt a deep convolutional neural network model (Convolutional Neural Network, CNN) model and the like.

Step S20 , according to the predetermined mapping relationship between the recognition result and the degree of retinopathy, determine the degree of retinopathy corresponding to the output recognition result.

After the pre-trained deep convolutional neural network model is used to identify the received retinopathy pictures and obtain the recognition results, the corresponding output recognition results can be determined according to the predetermined mapping relationship between the recognition results and the degree of retinopathy. Retinopathy degree grade, the determined retinopathy degree grade is the retinopathy degree grade corresponding to the received retinopathy picture. For example, in one embodiment, the recognition results include a first recognition result (for example, marked as "0"), a second recognition result (for example, marked as "1"), a third recognition result (for example, marked as is "2"), the fourth recognition result (for example, marked as "3") and the fifth recognition result (for example, marked as "4"), the grades of the degree of retinopathy include the first grade, the second grade, the first grade Level 3, Level 4 and Level 5. The mapping relationship between different recognition results and grades of retinopathy can be determined in advance, for example, the first recognition result corresponds to the first grade, the second recognition result corresponds to the second grade, the third recognition result corresponds to the third grade, and the fourth recognition result corresponds to The fourth level, the fifth recognition result corresponds to the fifth level. For example, specifically, the first level may correspond to normal and mild non-proliferative diabetic retinopathy, and the picture of retinopathy corresponding to the first level shows only individual hemangiomas, hard exudates, retinal hemorrhages and the like. The second grade can correspond to non-proliferative diabetic retinopathy without clinically significant macular edema. This second grade corresponds to retinopathy pictures showing microangiomas, hard exudates, retinal hemorrhages, looped or beaded veins. The third grade can correspond to non-proliferative diabetic retinopathy with clinically significant macular edema (C SME). Exudation, retinal hemorrhage. The fourth grade can correspond to proliferative retinopathy in the non-high-risk stage. The retinopathy picture corresponding to the fourth grade shows neovascularization in the outer area of the optic disc and proliferative changes in retinal microvascular formation in other areas. The fifth grade can correspond to the proliferative retinopathy in the high-risk stage, and the retinopathy picture corresponding to the fifth grade shows neovascularization in the optic disc region, vitreous body or preretinal hemorrhage.

In this way, after using the pre-trained recognition model to recognize the received retinopathy pictures and obtain the recognition results, the corresponding different grades of retinopathy can be determined according to the obtained different recognition results, so as to achieve a variety of refined Accurate identification of grades of retinopathy.

In this embodiment, a deep convolutional neural network model obtained by training based on a preset number of sample pictures marked with different grades of retinopathy is used to identify received retinopathy pictures, and determine the corresponding grade of retinopathy according to the recognition results . Since it only needs to recognize the received retinal lesion pictures according to the pre-trained deep convolutional neural network model, it does not need to perform complex feature extraction operations on eye images, which is simpler and can determine the corresponding different retinal lesions according to the recognition results The level of degree can effectively identify the degree of retinopathy of the patient in a fine-grained manner.

Further, in other embodiments, the training process of the predetermined recognition model is as follows:

A. Prepare corresponding presets for each preset grade of retinopathy (such as grade 1, grade 2, grade 3, grade 4 and grade 5, or mild, mild, moderate, severe, etc.) number of sample pictures, and mark the corresponding grade of retinal lesion for each sample picture;

B. Perform image preprocessing on each sample image to obtain a training image to be trained for the model. Model training is performed after image preprocessing such as scaling, cropping, flipping, and/or distorting operations are performed on each sample image to effectively improve the authenticity and accuracy of model training. For example, in one implementation manner, performing image preprocessing on each sample image may include:

Scale the length of the shorter side of each sample picture to a first preset size (for example, 640 pixels) to obtain a corresponding first picture, and randomly crop out a second preset size (for example, 256* 256 pixels) of the second image;

According to the standard parameter values corresponding to each predetermined preset type parameter (for example, color, brightness and/or contrast, etc.) (for example, the standard parameter value corresponding to color is a1, the standard parameter value corresponding to brightness is a2, and the standard parameter value corresponding to contrast The standard parameter value is a3), each predetermined preset type parameter value of each second picture is adjusted to the corresponding standard parameter value, and the corresponding third picture is obtained, so as to eliminate the external conditions when taking the sample picture as a medical picture The resulting pictures are unclear, improving the effectiveness of model training; for example, adjusting the brightness value of each second picture to the standard parameter value a2, and adjusting the contrast value of each second picture to the standard parameter value a3;

Flip each third picture in a preset direction (for example, horizontal and vertical directions), and perform a twist operation on each third picture according to a preset twist angle (for example, 30 degrees), to obtain the third picture corresponding to each third picture Four pictures, each fourth picture is a training picture of the corresponding sample picture. Among them, the function of flipping and twisting operations is to simulate various forms of pictures in actual business scenarios. Through these flipping and twisting operations, the scale of the data set can be increased, thereby improving the authenticity and practicability of model training.

C, divide all training pictures into the training set of the first proportion (for example, 50%), the verification set of the second proportion (for example, 25%);

D. Using the training set to train the predetermined recognition model;

E. Use the verification set to verify the accuracy of the trained recognition model. If the accuracy is greater than or equal to the preset accuracy, the training ends. Or, if the accuracy is less than the preset accuracy, then increase the degree of retinopathy corresponding to each grade. The number of sample pictures and re-execute the above steps B, C, D, E until the accuracy of the trained recognition model is greater than or equal to the preset accuracy.

Further, in other embodiments, the predetermined recognition model, that is, the deep convolutional neural network model includes an input layer and a plurality of network layers, and the network layers include a convolutional layer, a pooling layer, a fully connected layer, and a classification Optionally, the deep convolutional neural network model can also include a network layer (i.e. Dropout layer) with a mechanism for randomly discarding certain connection weights. The function of this network layer is to improve the recognition accuracy of the model.

In a specific embodiment, the deep convolutional neural network model consists of 1 input layer, 11 convolutional layers, 5 pooling layers, and 1 network layer with a mechanism of randomly discarding some connection weights (i.e. Dropout layer), a fully connected layer, and a classifier layer. The detailed structure of the deep convolutional neural network model is shown in Table 1 below:

Table 1

Among them: Layer Name represents the name of the network layer, Input represents the data input layer of the network, Conv represents the convolutional layer of the model, Conv1 represents the first convolutional layer of the model, MaxPool represents the maximum pooling layer of the model, and MaxPool1 represents the first A pooling layer based on the maximum value, Dropout means a network layer with a mechanism of randomly discarding some connection weights, Avgpool5 means the fifth pooling layer but uses the mean value for pooling, Fc means the fully connected layer in the model, Fc1 Indicates the first fully connected layer, Softmax indicates the Softmax classifier layer; Batch Size indicates the number of input images of the current layer; Kernel Size indicates the scale of the convolution kernel of the current layer (for example, Kernel Size can be equal to 3, indicating the scale of the convolution kernel is 3x 3); Stride Size represents the moving step of the convolution kernel, that is, the distance to move to the next convolution position after one convolution; Output Size represents the size of the network layer output feature map. It should be noted that the pooling methods of the pooling layer in this embodiment include but are not limited to Mean pooling (average sampling), Max pooling (maximum sampling), Overlapping (overlapping sampling), L2pooling (mean square sampling), Local Contrast Normalization (normalized sampling), Stochasticpooling (random sampling), Def-pooling (deformation constraint sampling), etc.

Further, in other embodiments, in order to improve the recognition accuracy of the model, each of the network layers (for example, a convolutional layer, a pooling layer, a network layer with a mechanism for randomly discarding certain connection weights, a fully connected layer, and a classifier) layer, etc.) corresponding to the activation function f(x) is:

f(x)=max(α*x,0)

Among them, α is the leakage rate, and x represents a numerical input of a neuron in the deep convolutional neural network model. In a preferred implementation of this embodiment, α is set to 0.5. After a comparative test of the same test data set, compared with other existing activation functions, the recognition accuracy of the deep convolutional neural network model is improved by about 3% through the activation function f(x) of this embodiment.

Further, in other embodiments, in order to improve the recognition accuracy of the model, each of the network layers (for example, a convolutional layer, a pooling layer, a network layer with a mechanism for randomly discarding certain connection weights, a fully connected layer, and a classifier) The corresponding cross entropy H(P,Q) is:

H(P,Q)＝H(P)+D _KL (P||Q)

Among them, P and Q are two probability distributions, H(P) is the expectation of probability distribution P, H(P)=-∑ _x∈X P(x)logP(x), x is the sample space X of probability distribution P Any sample in , P(x) represents the probability that sample x is selected; the expression of D _KL (P||Q) is x is any sample in the common sample space X of probability distribution P and Q, P(x) represents the probability that sample x is selected on probability distribution P, and Q(x) represents the probability that sample x is selected on probability distribution Q.

Further, in order to ensure the efficiency and accuracy of model training, the cross-entropy loss function L(x, :W) for:

Among them, x represents the input of the model, Represents the label corresponding to the input, W represents the preset model parameters, X represents the model input space, f(x:W) represents the output of the model after transforming the input x, ζ represents the normalization factor, ||W|| ² means to sum the matrix elements:

W _i+1 ＝W _i +ΔW _i+1

Among them, ΔW _i+1 represents the update increment of the weight matrix at time i+1, α is the potential energy item, β is the weight decay coefficient, γ is the learning rate of the model, and W _i represents the state value of the weight matrix at time i , _Di represents the i-th batch of input, Indicates the average gradient corresponding to the i-th batch of inputs.

In this embodiment, cross entropy can be used as a loss function in a neural network (machine learning). For example, P represents the distribution of real marks, and Q is the predicted mark distribution of the trained model. The cross entropy loss function can measure P and Q similarity to ensure the accuracy of model training. Moreover, cross entropy as a loss function can avoid the problem of reduced learning rate of the mean square error loss function during gradient descent, so it can ensure the efficiency of model training.

Further, in other embodiments, the deep convolutional neural network model includes at least one fully connected layer, and the initial value of each weight in the predetermined recognition model ranges from a preset weight range (for example, (0, 1) weight range) is determined by random sampling, the probability of the connection weight of the fully connected layer being discarded (Dropout) is set to a first preset value (for example, 0.5), and the weight decay coefficient in the cross quotient loss function It is set to a second preset value (for example, 0.0005), and the potential energy item in the cross quotient loss function is set to a third preset value (for example, 0.9).

Further, in other embodiments, the scoring function of the predetermined recognition model for:

in, O _{i, j} represents the number of pictures that are actually predicted as i for the first time and j for the second time, O represents an N*N matrix, O _{i, j} represents the matrix elements in matrix O, and N represents the number of pictures participating in the prediction Number of pictures, prediction result i, j ∈ {0, 1, 2, 3, 4}, E _{i, j} represents the number of images that should appear for the first prediction as i and the second prediction as j, E is the expected prediction The N*N matrix of the result, E _i,j represent the matrix elements in the matrix E.

In this embodiment, through the scoring function To detect the recognition accuracy of the predetermined recognition model, so as to ensure that the recognition accuracy of the trained predetermined recognition model is maintained at a high level, so as to ensure accurate recognition of the degree of retinopathy of the patient.

The present invention further provides a grade detection system for the degree of retinopathy. Please refer to FIG. 2 , which is a schematic view of the operating environment of a preferred embodiment of the system 10 for detecting the degree of retinopathy in the present invention.

In this embodiment, the retinopathy level detection system 10 is installed and operated in the electronic device 1 . The electronic device 1 may include, but not limited to, a memory 11 , a processor 12 and a display 13 . Figure 2 only shows the electronic device 1 with the components 11-13, but it is to be understood that implementation of all of the illustrated components is not required and that more or fewer components may instead be implemented.

The storage 11 may be an internal storage unit of the electronic device 1 in some embodiments, such as a hard disk or a memory of the electronic device 1 . The memory 11 may also be an external storage device of the electronic device 1 in other embodiments, such as a plug-in hard disk equipped on the electronic device 1, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, flash memory card (Flash Card), etc. Further, the memory 11 may also include both an internal storage unit of the electronic device 1 and an external storage device. The memory 11 is used to store application software and various data installed in the electronic device 1 , such as program codes of the retinopathy level detection system 10 . The memory 11 can also be used to temporarily store data that has been output or will be output.

The processor 12 may be a central processing unit (Central Processing Unit, CPU) in some embodiments, a microprocessor or other data processing chips, for running the program code stored in the memory 11 or processing data, for example Execute the retinopathy degree grade detection system 10 and the like.

The display 13 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, Organic Light-Emitting Diode) touch panel, etc. in some embodiments. The display 13 is used for displaying information processed in the electronic device 1 and for displaying a visualized user interface, such as an application menu interface, an application icon interface, and the like. The components 11-13 of the electronic device 1 communicate with each other via a system bus.

Please refer to FIG. 3 , which is a functional block diagram of a preferred embodiment of the system 10 for detecting the degree of retinopathy in the present invention. In this embodiment, the retinopathy level detection system 10 can be divided into one or more modules, and the one or more modules are stored in the memory 11 and controlled by one or more processors (This embodiment is executed by the processor 12) to complete the present invention. For example, in FIG. 3 , the retinopathy level detection system 10 can be divided into an identification module 01 and a determination module 02 . The module referred to in the present invention refers to a series of computer program instruction segments capable of completing specific functions, which is more suitable than a program to describe the execution process of the retinopathy level detection system 10 in the electronic device 1 . The following description will specifically introduce the functions of the identification module 01 and the determination module 02.

Referring to FIG. 3 , FIG. 3 is a schematic diagram of functional modules of an embodiment of the system for detecting the degree of retinopathy in the present invention.

In one embodiment, the retinopathy grade detection system includes:

The recognition module 01 is configured to, after receiving the retinopathy picture to be recognized, use a predetermined recognition model to recognize the received retinopathy picture, and output the recognition result; wherein, the predetermined recognition model is a pre-passed A convolutional neural network model obtained by training a preset number of sample pictures marked with different grades of retinopathy.

In this embodiment, the retinopathy degree level detection system receives the retinopathy level detection request sent by the user including the retinopathy picture to be identified, for example, receives the retinopathy level level detection request sent by the user through a terminal such as a mobile phone, a tablet computer, and a self-service terminal device. Level detection request, such as receiving the level detection request of the degree of retinopathy sent by the user on the client pre-installed in the mobile phone, tablet computer, self-service terminal equipment, etc. The retinopathy degree grade detection request sent from the browser system.

Retinopathy degree level detection system, after receiving the user's retinopathy degree level detection request, uses the pre-trained recognition model to identify the received retinopathy pictures to be identified, and recognizes the received retinopathy level to be identified The recognition result of the image in the recognition model. The identification model can be continuously trained, learned, verified, optimized, etc. by identifying a large number of preset sample pictures marked with different levels of retinal lesions in advance, so as to train it to accurately identify the corresponding levels of different levels of retinal lesions. Annotated model of . For example, the recognition model may adopt a deep convolutional neural network model (Convolutional Neural Network, CNN) model and the like.

The determination module 02 is configured to determine the degree of retinopathy corresponding to the output recognition result according to the predetermined mapping relationship between the recognition result and the degree of retinopathy.

After the pre-trained deep convolutional neural network model is used to identify the received retinopathy pictures and obtain the recognition results, the corresponding output recognition results can be determined according to the predetermined mapping relationship between the recognition results and the degree of retinopathy. Retinopathy degree grade, the determined retinopathy degree grade is the retinopathy degree grade corresponding to the received retinopathy picture. For example, in one embodiment, the recognition results include a first recognition result (for example, marked as "0"), a second recognition result (for example, marked as "1"), a third recognition result (for example, marked as is "2"), the fourth recognition result (for example, marked as "3"), and the fifth recognition result (for example, marked as "4"), the grades of the degree of retinopathy include the first grade, the second grade, The third level, the fourth level and the fifth level. The mapping relationship between different recognition results and grades of retinopathy can be determined in advance, for example, the first recognition result corresponds to the first grade, the second recognition result corresponds to the second grade, the third recognition result corresponds to the third grade, and the fourth recognition result corresponds to The fourth level, the fifth recognition result corresponds to the fifth level. For example, specifically, the first level may correspond to normal and mild non-proliferative diabetic retinopathy, and the picture of retinopathy corresponding to the first level shows only individual hemangiomas, hard exudates, retinal hemorrhages and the like. The second grade can correspond to non-proliferative diabetic retinopathy without clinically significant macular edema. This second grade corresponds to retinopathy pictures showing microangiomas, hard exudates, retinal hemorrhages, looped or beaded veins. The third grade can correspond to non-proliferative diabetic retinopathy with clinically significant macular edema (C SME). Exudation, retinal hemorrhage. The fourth grade can correspond to proliferative retinopathy in the non-high-risk stage. The retinopathy picture corresponding to the fourth grade shows neovascularization in the outer area of the optic disc and proliferative changes in retinal microvascular formation in other areas. The fifth grade can correspond to the proliferative retinopathy in the high-risk stage, and the retinopathy picture corresponding to the fifth grade shows neovascularization in the optic disc region, vitreous body or preretinal hemorrhage.

In this way, after using the pre-trained recognition model to recognize the received retinopathy pictures and obtain the recognition results, the corresponding different grades of retinopathy can be determined according to the obtained different recognition results, so as to achieve a variety of refined Accurate identification of grades of retinopathy.

In this embodiment, a deep convolutional neural network model obtained by training based on a preset number of sample pictures marked with different grades of retinopathy is used to identify received retinopathy pictures, and determine the corresponding grade of retinopathy according to the recognition results . Since it only needs to recognize the received retinal lesion pictures according to the pre-trained deep convolutional neural network model, it does not need to perform complex feature extraction operations on eye images, which is simpler and can determine the corresponding different retinal lesions according to the recognition results The level of degree can effectively identify the degree of retinopathy of the patient in a fine-grained manner.

Further, in other embodiments, the training process of the predetermined recognition model is as follows:

A. Prepare corresponding presets for each preset grade of retinopathy (such as grade 1, grade 2, grade 3, grade 4 and grade 5, or mild, mild, moderate, severe, etc.) number of sample pictures, and mark the corresponding grade of retinal lesion for each sample picture;

B. Perform image preprocessing on each sample image to obtain a training image to be trained for the model. Model training is performed after image preprocessing such as scaling, cropping, flipping, and/or distorting operations are performed on each sample image to effectively improve the authenticity and accuracy of model training. For example, in one implementation manner, performing image preprocessing on each sample image may include:

Scale the length of the shorter side of each sample picture to a first preset size (for example, 640 pixels) to obtain a corresponding first picture, and randomly crop out a second preset size (for example, 256* 256 pixels) of the second image;

According to the standard parameter values corresponding to each predetermined preset type parameter (for example, color, brightness and/or contrast, etc.) (for example, the standard parameter value corresponding to color is a1, the standard parameter value corresponding to brightness is a2, and the standard parameter value corresponding to contrast The standard parameter value is a3), each predetermined preset type parameter value of each second picture is adjusted to the corresponding standard parameter value, and the corresponding third picture is obtained, so as to eliminate the external conditions when taking the sample picture as a medical picture The resulting pictures are unclear, improving the effectiveness of model training; for example, adjusting the brightness value of each second picture to the standard parameter value a2, and adjusting the contrast value of each second picture to the standard parameter value a3;

Flip each third picture in a preset direction (for example, horizontal and vertical directions), and perform a twist operation on each third picture according to a preset twist angle (for example, 30 degrees), to obtain the third picture corresponding to each third picture Four pictures, each fourth picture is a training picture of the corresponding sample picture. Among them, the function of flipping and twisting operations is to simulate various forms of pictures in actual business scenarios. Through these flipping and twisting operations, the scale of the data set can be increased, thereby improving the authenticity and practicability of model training.

C, divide all training pictures into the training set of the first proportion (for example, 50%), the verification set of the second proportion (for example, 25%);

D. Using the training set to train the predetermined recognition model;

E. Use the verification set to verify the accuracy of the trained recognition model. If the accuracy is greater than or equal to the preset accuracy, the training ends. Or, if the accuracy is less than the preset accuracy, then increase the degree of retinopathy corresponding to each grade. The number of sample pictures and re-execute the above steps B, C, D, E until the accuracy of the trained recognition model is greater than or equal to the preset accuracy.

Further, in other embodiments, the predetermined recognition model, that is, the deep convolutional neural network model includes an input layer and a plurality of network layers, and the network layers include a convolutional layer, a pooling layer, a fully connected layer, and a classification Optionally, the deep convolutional neural network model can also include a network layer (i.e. Dropout layer) with a mechanism for randomly discarding certain connection weights. The function of this network layer is to improve the recognition accuracy of the model.

In a specific embodiment, the deep convolutional neural network model consists of 1 input layer, 11 convolutional layers, 5 pooling layers, and 1 network layer with a mechanism of randomly discarding some connection weights (i.e. Dropout layer), a fully connected layer, and a classifier layer. The detailed structure of the deep convolutional neural network model is shown in Table 1 below:

Table 1

Among them: Layer Name represents the name of the network layer, Input represents the data input layer of the network, Conv represents the convolutional layer of the model, Conv1 represents the first convolutional layer of the model, MaxPool represents the maximum pooling layer of the model, and MaxPool1 represents the first A pooling layer based on the maximum value, Dropout means a network layer with a mechanism of randomly discarding some connection weights, Avgpool5 means the fifth pooling layer but uses the mean value for pooling, Fc means the fully connected layer in the model, Fc1 Indicates the first fully connected layer, Softmax indicates the Softmax classifier layer; Batch Size indicates the number of input images of the current layer; Kernel Size indicates the scale of the convolution kernel of the current layer (for example, Kernel Size can be equal to 3, indicating the scale of the convolution kernel is 3x3); Stride Size represents the moving step of the convolution kernel, that is, the distance to move to the next convolution position after one convolution is completed; Output Size represents the size of the feature map output by the network layer. It should be noted that the pooling methods of the pooling layer in this embodiment include but are not limited to Mean pooling (average sampling), Max pooling (maximum sampling), Overlapping (overlapping sampling), L2pooling (mean square sampling), Local Contrast Normalization (normalized sampling), Stochasticpooling (random sampling), Def-pooling (deformation constraint sampling), etc.

Further, in other embodiments, in order to improve the recognition accuracy of the model, each of the network layers (for example, a convolutional layer, a pooling layer, a network layer with a mechanism for randomly discarding certain connection weights, a fully connected layer, and a classifier) layer, etc.) corresponding to the activation function f(x) is:

f(x)=max(α*x,0)

Among them, α is the leakage rate, and x represents a numerical input of a neuron in the deep convolutional neural network model. In a preferred implementation of this embodiment, α is set to 0.5. After a comparative test of the same test data set, compared with other existing activation functions, the recognition accuracy of the deep convolutional neural network model is improved by about 3% through the activation function f(x) of this embodiment.

Further, in other embodiments, in order to improve the recognition accuracy of the model, each of the network layers (for example, a convolutional layer, a pooling layer, a network layer with a mechanism for randomly discarding certain connection weights, a fully connected layer, and a classifier) The corresponding cross entropy H(P,Q) is:

H(P,Q)＝H(P)+D _KL (P||Q)

Among them, P and Q are two probability distributions, H(P) is the expectation of probability distribution P, H(P)=-∑ _x∈X P(x)logP(x), x is the sample space X of probability distribution P Any sample in , P(x) represents the probability that sample x is selected; the expression of D _KL (P||Q) is x is any sample in the common sample space X of probability distribution P and Q, P(x) represents the probability that sample x is selected on probability distribution P, and Q(x) represents the probability that sample x is selected on probability distribution Q.

Further, in order to ensure the efficiency and accuracy of model training, the cross-entropy loss function L(x, :W) for:

Among them, x represents the input of the model, Represents the label corresponding to the input, W represents the preset model parameters, X represents the model input space, f(x:W) represents the output of the model after transforming the input x, ζ represents the normalization factor, ||W|| ² means to sum the matrix elements:

W _i+1 ＝W _i +ΔW _i+1

Among them, ΔW _i+1 represents the update increment of the weight matrix at time i+1, α is the potential energy item, β is the weight decay coefficient, γ is the learning rate of the model, and W _i represents the state value of the weight matrix at time i , _Di represents the i-th batch of input, Indicates the average gradient corresponding to the i-th batch of inputs.

In this embodiment, cross entropy can be used as a loss function in a neural network (machine learning). For example, P represents the distribution of real marks, and Q is the predicted mark distribution of the trained model. The cross entropy loss function can measure P and Q similarity to ensure the accuracy of model training. Moreover, cross entropy as a loss function can avoid the problem of reduced learning rate of the mean square error loss function during gradient descent, so it can ensure the efficiency of model training.

Further, in other embodiments, the deep convolutional neural network model includes at least one fully connected layer, and the initial value of each weight in the predetermined recognition model ranges from a preset weight range (for example, (0, 1) weight range) is determined by random sampling, the probability of the connection weight of the fully connected layer being discarded (Dropout) is set to a first preset value (for example, 0.5), and the weight decay coefficient in the cross quotient loss function It is set to a second preset value (for example, 0.0005), and the potential energy item in the cross quotient loss function is set to a third preset value (for example, 0.9).

Further, in other embodiments, the scoring function of the predetermined recognition model for:

in, O _{i, j} represents the number of pictures that are actually predicted as i for the first time and j for the second time, O represents an N*N matrix, O _{i, j} represents the matrix elements in matrix O, and N represents the number of pictures participating in the prediction Number of pictures, prediction result i, j ∈ {0, 1, 2, 3, 4}, E _{i, j} represents the number of images that should appear for the first prediction as i and the second prediction as j, E is the expected prediction The N*N matrix of the result, E _i,j represent the matrix elements in the matrix E.

In this embodiment, through the scoring function To detect the recognition accuracy of the predetermined recognition model, so as to ensure that the recognition accuracy of the trained predetermined recognition model is maintained at a high level, so as to ensure accurate recognition of the degree of retinopathy of the patient.

It should be noted that, in this document, the term "comprising", "comprising" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article or apparatus comprising a set of elements includes not only those elements, It also includes other elements not expressly listed, or elements inherent in the process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus comprising that element.

Through the description of the above embodiments, those skilled in the art can clearly understand that the methods in the above embodiments can be implemented by means of software plus a necessary general-purpose hardware platform, and of course can also be implemented by hardware, but in many cases the former is better implementation. Based on such an understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art can be embodied in the form of software products, and the computer software products are stored in a storage medium (such as ROM/RAM, disk, CD) contains several instructions to make a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in various embodiments of the present invention.

The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, and the scope of rights of the present invention is not limited thereto. The serial numbers of the above embodiments of the present invention are for description only, and do not represent the advantages and disadvantages of the embodiments. In addition, although a logical order is shown in the flowcharts, in some cases the steps shown or described may be performed in an order different from that shown or described herein.

Those skilled in the art can realize the present invention with many variants without departing from the scope and spirit of the present invention, for example, the features of one embodiment can be used in another embodiment to obtain another embodiment. All modifications, equivalent replacements and improvements made within the technical conception of the application of the present invention shall fall within the scope of rights of the present invention.

Claims (10)

1. A grade detection system for the degree of retinopathy, characterized in that, the grade detection system for the degree of retinopathy comprises: The recognition module is configured to, after receiving the retinopathy picture to be recognized, use a predetermined recognition model to recognize the received retinopathy picture, and output a recognition result; A convolutional neural network model obtained by training a preset number of sample pictures marked with different levels of retinopathy; The determination module is configured to determine the degree of retinopathy corresponding to the output recognition result according to the predetermined mapping relationship between the recognition result and the degree of retinopathy. 2. the retinopathy degree grade detection system as claimed in claim 1, is characterized in that, the training process of described predetermined identification model is as follows: A. Set a corresponding preset number of sample pictures for each preset degree of retinopathy, and mark the corresponding degree of retinopathy for each sample picture; B. Perform image preprocessing on each sample image to obtain a training image to be trained by the model; C, divide all training pictures into the training set of the first proportion and the verification set of the second proportion; D. Using the training set to train the predetermined recognition model; E. Use the verification set to verify the accuracy of the trained recognition model. If the accuracy is greater than or equal to the preset accuracy, the training ends. Or, if the accuracy is less than the preset accuracy, then increase the degree of retinopathy corresponding to each grade. number of sample images and re-execute steps B, C, D, and E above. 3. the retinopathy degree grade detection system as claimed in claim 2, is characterized in that, described step B comprises: Scaling the shorter side length of each sample picture to a first preset size to obtain a corresponding first picture, and randomly cropping a second picture of a second preset size on each first picture; Adjusting each predetermined preset type parameter value of each second picture to a corresponding standard parameter value according to a standard parameter value corresponding to each predetermined preset type parameter, so as to obtain a corresponding third picture; Perform a flip operation in a preset direction on each third picture, and perform a twist operation on each third picture according to a preset twist angle to obtain a fourth picture corresponding to each third picture, and use each fourth picture as a model to be trained training pictures. 4. the retinopathy degree grade detection system as claimed in claim 1 or 2, is characterized in that, described depth convolutional neural network model comprises input layer and a plurality of network layers, and described network layer comprises convolution layer, pooling Layer, fully connected layer and classifier layer, the activation function f(x) corresponding to each described network layer is: f(x)=max(α*x,0) Wherein, α is a preset leakage rate, and x represents a numerical input of a neuron in the deep convolutional neural network model. 5. the retinopathy degree grade detection system as claimed in claim 4, is characterized in that, the corresponding cross entropy H (P, Q) of each described network layer is: H(P,Q)＝H(P)+D _KL (P||Q) Among them, P and Q are two probability distributions, H(P) is the expectation of probability distribution P, H(P)=-Σ _x∈X P(x)logP(x), x is the sample space X of probability distribution P Any sample in , P(x) represents the probability that sample x is selected; the expression of D _KL (P||Q) is x is any sample in the common sample space X of probability distribution P and Q, P(x) represents the probability that sample x is selected on probability distribution P, and Q(x) represents the probability that sample x is selected on probability distribution Q. 6. The retinopathy degree grade detection system as claimed in claim 1 or 2, wherein the scoring function of the predetermined identification model for: in, O _{i, j} represents the number of pictures that are actually predicted as i for the first time and j for the second time, O represents an N*N matrix, O _{i, j} represents the matrix elements in matrix O, and N represents the number of pictures participating in the prediction Number of pictures, prediction result i, j ∈ {0, 1, 2, 3, 4}, E _{i, j} represents the number of images that should appear for the first prediction as i and the second prediction as j, E is the expected prediction The N*N matrix of the result, E _i,j represent the matrix elements in the matrix E. 7. The retinopathy degree grade detection system as claimed in claim 4, wherein the cross-entropy loss function corresponding to each said network layer for: <mrow> <mi>L</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mover> <mi>y</mi> <mo>^</mo> </mover> <mo>:</mo> <mi>W</mi> <mo>)</mo> </mrow> <mo>=</mo> <mo>-</mo> <mfrac> <mn>1</mn> <mi>n</mi> </mfrac> <msub> <mi>&amp;Sigma;</mi> <mrow> <mi>x</mi> <mo>&amp;Element;</mo> <mi>X</mi> </mrow> </msub> <mrow> <mo>(</mo> <mover> <mi>y</mi> <mo>^</mo> </mover> <mo>*</mo> <mi>log</mi> <mi>f</mi> <mo>(</mo> <mrow> <mi>x</mi> <mo>:</mo> <mi>W</mi> </mrow> <mo>)</mo> <mo>+</mo> <mo>(</mo> <mrow> <mn>1</mn> <mo>-</mo> <mover> <mi>y</mi> <mo>^</mo> </mover> </mrow> <mo>)</mo> <mo>*</mo> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mo>(</mo> <mrow> <mn>1</mn> <mo>-</mo> <mi>f</mi> <mrow> <mo>(</mo> <mrow> <mi>x</mi> <mo>:</mo> <mi>W</mi> </mrow> <mo>)</mo> </mrow> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <mo>+</mo> <mi>&amp;zeta;</mi> <mo>|</mo> <mo>|</mo> <mi>W</mi> <mo>|</mo> <msup> <mo>|</mo> <mn>2</mn> </msup> </mrow> Among them, x represents the input of the model, Represents the label corresponding to the input, W represents the preset model parameters, X represents the model input space, f(x:W) represents the output of the model after transforming the input x, ζ represents the normalization factor, ||W|| ² means to sum the matrix elements: <mrow> <msub> <mi>&amp;Delta;W</mi> <mrow> <mi>i</mi> <mo>+</mo> <mn>1</mn> </mrow> </msub> <mo>=</mo> <mi>&amp;alpha;</mi> <mo>*</mo> <msub> <mi>&amp;Delta;W</mi> <mi>i</mi> </msub> <mo>-</mo> <mi>&amp;beta;</mi> <mo>*</mo> <mi>&amp;gamma;</mi> <mo>*</mo> <msub> <mi>W</mi> <mi>i</mi> </msub> <mo>-</mo> <mi>&amp;gamma;</mi> <mo>*</mo> <mo>&lt;</mo> <mfrac> <mrow> <mi>d</mi> <mi>L</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mover> <mi>y</mi> <mo>^</mo> </mover> <mo>:</mo> <mi>W</mi> <mo>)</mo> </mrow> </mrow> <mrow> <mi>d</mi> <mi>W</mi> </mrow> </mfrac> <mo>|</mo> <msub> <mi>W</mi> <mi>i</mi> </msub> <msub> <mo>&gt;</mo> <msub> <mi>D</mi> <mi>i</mi> </msub> </msub> </mrow> W _i+1 ＝W _i +ΔW _i+1 Among them, ΔW _i+1 represents the update increment of the weight matrix at time i+1, α is the potential energy item, β is the weight decay coefficient, γ is the learning rate of the model, and W _i represents the state value of the weight matrix at time i , _Di represents the i-th batch of input, Indicates the average gradient corresponding to the i-th batch of inputs. 8. The retinal lesion degree grade detection system as claimed in claim 7, wherein the predetermined recognition model comprises at least one fully connected layer, and the initial value of each weight in the predetermined recognition model is from a predetermined The set weight range is determined by random sampling, the probability that the connection weight of the fully connected layer is discarded is set to a first preset value, and the weight decay coefficient in the cross-entropy loss function is set to a second preset value, so The potential energy item in the cross quotient loss function is set to the third preset value. 9. A detection method for grades of retinopathy, characterized in that said method comprises the following steps: After receiving the picture of retinopathy to be identified, use a predetermined recognition model to identify the received picture of retinopathy, and output the recognition result; The convolutional neural network model obtained by training the preset number of sample pictures of degree level; According to the predetermined mapping relationship between the recognition result and the degree of retinopathy, the degree of retinopathy corresponding to the output recognition result is determined. 10. the detection method of degree of retinopathy grade as claimed in claim 9, is characterized in that, the training process of described predetermined identification model is as follows: A. Set a corresponding preset number of sample pictures for each preset degree of retinopathy, and mark the corresponding degree of retinopathy for each sample picture; B. Perform image preprocessing on each sample image to obtain a training image to be trained by the model; C, divide all training pictures into the training set of the first proportion and the verification set of the second proportion; D. Using the training set to train the predetermined recognition model; E. Use the verification set to verify the accuracy of the trained recognition model. If the accuracy is greater than or equal to the preset accuracy, the training ends. Or, if the accuracy is less than the preset accuracy, then increase the degree of retinopathy corresponding to each grade. number of sample images and re-execute steps B, C, D, and E above.