CN110516685A - Detection method of lens turbidity based on convolutional neural network - Google Patents

Abstract

Lenticular opacities degree detecting method based on convolutional neural networks, the method steps are as follows: (1), to testing image using illumination enhancing method to carry out pretreatment and form preprocessing image data；(2), the preprocessing image data in (1) step is substituted into lenticular opacities degree detecting learning model and realizes detection, the Inception-V3 model and parameter that the present invention is crossed using ImageNet pre-training, and be trained to obtain disaggregated model using the thought of transfer learning, it can be achieved to pass through cell phone application real-time perfoming lenticular opacities investigation and comparison after the completion of the system.

Description

Detection method of lens turbidity based on convolutional neural network

technical field

The invention provides a method for detecting the turbidity degree of a lens based on a convolutional neural network, which belongs to the field of image processing.

Background technique

The degree of lens turbidity can be used as an intermediate reference data in many fields. For example, cataracts can be analyzed through this data. At present, the automatic classification method for the degree of lens turbidity has made great progress, but there are still some problems. First of all, most of the existing methods focus on classification problems, while for feature extraction, they mainly use artificial features to find features, which is very subjective. Moreover, since the currently public and labeled data sets are only used to study the degree of lens turbidity, and the amount of data is small, it is impossible to use deep learning to automatically extract features, and the trained model cannot achieve the effect of general research.

Contents of the invention

Purpose of the invention:

The invention provides a method for detecting the turbidity degree of the lens based on a convolutional neural network, and its purpose is to solve the existing problems in the past.

Technical solutions

The lens turbidity detection method based on convolutional neural network is characterized in that: the method steps are as follows:

(1) The image to be tested is preprocessed by the illumination enhancement method to form preprocessed image data;

(2) Substituting the preprocessed image data in step (1) into the lens turbidity detection learning model to realize the detection.

(2) The construction method of the lens turbidity detection model in the step is as follows:

(2.1), construct the MSLPP data set; Then the image in the data set is preprocessed; And the preprocessed MSLPP data set is divided into training set, verification set and test set;

(2.2), using the data set formed from the preprocessed image in step (2.1) to perform model training to obtain a learning model for detecting the degree of lens turbidity.

(2.1) The construction method of the MSLPP data set in the step is as follows: collect the ocular lens images collected by the slit lamp in clinical practice, and divide them into three categories: normal, early lens opacity and lens opacity.

The sample collection environment of the MSLPP dataset is complex and diverse, and the samples contain a variety of types.

The complex and diverse environments include: bright environment, dark environment, and reflective conditions; the diverse types of samples include: samples for nuclear lens opacity, cortical lens opacity, and posterior capsular lens opacity.

The preprocessing method in the step (2.1) is: first use the illumination enhancement method to process the data set, and then use the data amplification method to process the data set processed by the illumination enhancement method.

The light enhancement method uses the following methods:

Compress the input image size to 299×299 pixels, and set the three-channel pixel of any point A(x,y) in the image Among them, R(x,y), G(x,y), and B(x,y) respectively represent the brightness values of the red, green, and blue channels in point A(x,y), and the brightness value ranges of the three channels Both are 0-255, the average pixel value of the image Expressed as:

like Then the pixel value of each point in the image becomes like Then the pixel value of each point in the image becomes like Then the pixel value of each point in the image remains unchanged.

2.1) There are the following three data amplification methods in the preprocessing of the step, and one, two or all of the three methods are selected:

(1) translation: the image after the illumination enhancement is translated up, down, left, and right respectively by 5-20 pixels (this place can be translated by 5-20 pixels, but the data adopted in the present invention is 12, and it is the same as that generated in Fig. 4 corresponding to the image); the image translation is amplified to the original multiple, and the multiple is the number of times of translation;

(2) Rotation: Rotate the image after the illumination enhancement (selectively, it can be clockwise only once, or only counterclockwise once, or clockwise and counterclockwise can be paralleled once) 5°-20° (but The data adopted in the present invention is 15 °, and corresponds to the image generated in Fig. 4); Make the image rotate and amplify to the original multiple, and this multiple is the number of times of rotation;

(3) Mirror image: Mirror the image after illumination enhancement up, down, left, and right once, that is, flip it up and down once, and flip it left and right once, so that the image mirror image is amplified to the original multiple, which is the number of mirror images;

If two or all of the three methods are selected, the image augmentation is only performed on the same light-enhanced image, and then the enhanced images are used together.

(2.2) Step model training uses the convolutional neural network for migration learning and then uses the preprocessed MSLPP data set to continue training the model after migration learning;

The convolutional neural network is mainly divided into three parts: the convolutional layer, the pooling layer and the fully connected layer; the final convolutional neural network structure is the Inception-V3 model proposed by Google based on GoogLeNet;

The steps of transfer learning and training are as follows: First, based on ImageNet (ImageNet is the name of a computer vision system recognition project, it is currently the largest database of image recognition in the world, the database is open source and can be used directly) image annotation data set, in Inception -Pre-train on the V3 model to extract a 2048-dimensional feature vector; this stage makes full use of knowledge transfer, uses pre-trained weights for feature extraction, and does not train the weight parameters of Inception-V3; then, input the feature vector A single-layer fully-connected neural network, using a single-layer fully-connected neural network containing a Softmax classifier, and then trained on the preprocessed MSLPP dataset (referring to images classified and labeled by ophthalmologists) to obtain final classification result.

The training process is as follows: (Here, "training" refers to all the training processes after preprocessing, and "training" in "the final classification result is obtained after the training of the classified lens cloudy crystal image" refers to transfer learning The subsequent training of the fine-tuning parameter part. Steps 1 and 2 correspond to using convolutional neural networks for transfer learning, and steps 3 and 4 correspond to using the preprocessed training set to continue training the model after transfer learning)

(1) Load the Inception-V3 model with the fully connected layer removed and the weight parameters obtained from pre-training with the ImageNet dataset;

(2) Add a fully connected layer structure to the initialized Inception-V3 model obtained in (1), and add a Dropout strategy to the fully connected layer (a method to prevent the model from overfitting), and set the ratio to 0.75 ;Extract a 2048-dimensional feature vector;

(3) Freeze all feature extraction layers except the fully connected layer, then set the learning rate to 0.001, use the preprocessed MSLPP data training set to train 1 epoch, and iterate 550 times;

(4) Unfreeze all layers, use fine-tuning transfer learning, continue to use the MSLPP data training set for training, use the stochastic gradient descent method, set the initial learning rate to 0.01, train 100 epochs, and each epoch iterates 550 times. One epoch, use the verification set to test the accuracy of the model. If the accuracy is higher than last time, save the training parameters. If the accuracy is lower, use the previously saved parameters to continue training. The number of batch samples batch_size is set to 32, and the momentum momentum is set to 0.9.

(The number of batch samples batch_size refers to the number of samples contained in a bit, which is a parameter that needs to be set during the training process.)

(The values mentioned in (2)(3)(4) are the specific parameters set when training the model. Although they can be adjusted within a certain range, adjusting the training parameters requires certain skills (according to the training set and verification Adjust the recall rate and loss function value of the set), if any adjustment within the range cannot guarantee the training results.)

(3) In the step of freezing the feature extraction layer, only the fully connected layer is selected for training, and the weights should be updated in a small range to avoid destroying the pre-trained features.

The above-mentioned small range means that when other layers are frozen and only the fully connected layer is changed, the weight coefficients inside the model do not change much during training. This sentence is an explanation of the method in step 3, not a parameter that needs to be adjusted manually.

A lens turbidity detection system based on a convolutional neural network, characterized in that: the system includes an image preprocessing module and a detection module;

The image preprocessing module preprocesses the image to be tested to form preprocessed image data; the image data in the image preprocessing module is substituted into the lens turbidity detection learning model in the detection module to realize detection;

The detection module includes MSLPP dataset component module, image preprocessing module and model training module;

The MSLPP data set component module constructs the MSLPP data set; then the image in the data set is preprocessed by the image preprocessing module;

The model training module performs model training on the data set formed from the image preprocessed by the image preprocessing module to obtain a learning model.

Advantages and effects:

The present invention provides a lens turbidity detection method based on a convolutional neural network. The present invention designs an automatic learning model of lens turbidity characteristics based on a convolutional neural network. The specific process is shown in FIG. 1 . First of all, in order to solve the problem of the current lack of data sets of lens turbidity, the present invention collects eye lens images collected by slit lamps in clinical practice, and the eye sampling technician divides them into three categories: normal, early lens turbidity and lens turbidity. class, build the MSLPP data set; then the image is preprocessed, the main operations are illumination enhancement and data volume amplification; finally, in order to solve the problem of automatically extracting depth features, the present invention utilizes the Inception-V3 model and parameters, and use the idea of transfer learning to train to obtain a classification model. After the completion of the system, real-time research and comparison of lens opacity can be realized through the mobile phone APP.

Description of drawings

Fig. 1 is the flow chart of automatic learning model of lens opacity feature based on convolutional neural network;

Figure 2 is a partial sample image of the MSLPP dataset;

Figure 3 is a comparison diagram before and after illumination enhancement;

Figure 4 is a comparison chart of data amplification;

Fig. 5 is a schematic diagram of transfer learning strategy training;

Figure 6 is a model visualization feature map;

Fig. 7 is a schematic diagram of an example of translation. Translation: For example: the left image is the original image, and the right image is the image after the image is shifted to the right by x pixels, and then moved down by y pixels.

Detailed ways

A method for detecting the degree of lens turbidity based on a convolutional neural network, the steps of which are as follows:

(1) The image to be tested is preprocessed by the illumination enhancement method to form preprocessed image data;

(2) Substituting the preprocessed image data in step (1) into the lens turbidity detection learning model to realize the detection.

(2) The construction method of the lens turbidity detection model in the step is as follows:

(2.1), construct the MSLPP data set; Then the image in the data set is preprocessed; And the preprocessed MSLPP data set is divided into training set, verification set and test set;

(2.2), using the data set formed from the preprocessed image in step (2.1) to perform model training to obtain a learning model for detecting the degree of lens turbidity.

(2.1) The construction method of the MSLPP data set in the step is as follows: collect the ocular lens images collected by the slit lamp in clinical practice, and divide them into three categories: normal, early lens opacity and lens opacity.

The preprocessing method in the step (2.1) is: first use the illumination enhancement method to process the data set, and then use the data amplification method to process the data set processed by the illumination enhancement method.

The light enhancement method uses the following methods:

Compress the input image size to 299×299 pixels, and set the three-channel pixel of any point A(x,y) in the image Among them, R(x,y), G(x,y), and B(x,y) respectively represent the brightness values of the red, green, and blue channels in point A(x,y), and the brightness value ranges of the three channels Both are 0-255, the average pixel value of the image Expressed as:

like Then the pixel value of each point in the image becomes like Then the pixel value of each point in the image becomes like Then the pixel value of each point in the image remains unchanged.

2.1) There are the following three data amplification methods in the preprocessing of the step, and one, two or all of the three methods are selected:

(1) Translation: translate the image after illumination enhancement by 5-20 pixels up, down, left, and right; make the image translation amplified to the original multiple, which is the number of translations;

(2) Rotation: Rotate the image after illumination enhancement by 5°-20° forward and backward; make the image rotation amplified to the original multiple, which is the number of rotations;

(3) Mirror image: Mirror the image after illumination enhancement up, down, left, and right once, that is, flip it up and down once, and flip it left and right once, so that the image mirror image is amplified to the original multiple, which is the number of mirror images;

If two or all of the three methods are selected, the image augmentation is only performed on the same light-enhanced image, and then the enhanced images are used together.

(2.2) Step model training uses the convolutional neural network for migration learning and then uses the preprocessed MSLPP data set to continue training the model after migration learning;

The steps of transfer learning and training are:

The first step is to pre-train on the Inception-V3 model based on the ImageNet image annotation dataset (the ImageNet dataset, which contains natural images), and extract a 2048-dimensional feature vector;

In the second step, the feature vector is input into a single-layer fully-connected neural network, and a single-layer fully-connected neural network including a Softmax classifier is used, and the final classification result is obtained after training on the preprocessed MSLPP data set.

The pre-training process on the Inception-V3 model in the first step is:

(1) Load the Inception-V3 model with the fully connected layer removed and the weight parameters obtained from pre-training with the ImageNet dataset;

(2) Add a fully connected layer structure to the initialized Inception-V3 model obtained in (1), and add a Dropout strategy to the fully connected layer, with the ratio set to 0.75; extract a 2048-dimensional feature vector; ImageNet dataset It is a public dataset that has been marked and belongs to public knowledge.

In the second step, a single-layer fully connected neural network including a Softmax classifier is used, and the final classification result is obtained after training with the preprocessed MSLPP data set as follows:

(1) Freeze all feature extraction layers except the fully connected layer (here only the fully connected layer corresponds to the single layer mentioned above), then set the learning rate to 0.001, and use the preprocessed MSLPP data training set to train 1 epoch, 550 iterations;

(2) Unfreeze all layers, use fine-tuning transfer learning, continue to use the MSLPP data training set for training, use the stochastic gradient descent method, set the initial learning rate to 0.01, train 100 epochs, and each epoch iterates 550 times. One epoch, use the verification set to test the accuracy of the model. If the accuracy is higher than last time, save the training parameters. If the accuracy is lower, use the previously saved parameters to continue training. The number of batch samples batch_size is set to 32, and the momentum momentum is set to 0.9.

(3) In the step of freezing the feature extraction layer, only the fully connected layer is selected for training, and the weights should be updated in a small range to avoid destroying the pre-trained features.

The system includes an image preprocessing module and a detection module;

The image preprocessing module preprocesses the image to be tested to form preprocessed image data; the image data in the image preprocessing module is substituted into the lens turbidity detection learning model in the detection module to realize detection;

The detection module includes MSLPP dataset component module, image preprocessing module and model training module;

The MSLPP data set component module constructs the MSLPP data set; then the image in the data set is preprocessed by the image preprocessing module;

The model training module performs model training on the data set formed from the image preprocessed by the image preprocessing module to obtain a learning model.

The present invention is further described in detail below:

The present invention designs a lens turbidity detection method based on a convolutional neural network, and the specific process is shown in FIG. 1 . First of all, in order to solve the problem of lack of current lens turbidity data sets, the present invention collects the ocular lens images collected by the slit lamp in clinic, and divides them into three categories by ophthalmologists: normal, early lens turbidity and lens turbidity, and constructs MSLPP data set; then the image is preprocessed, the main operations are illumination enhancement and data volume amplification; finally, in order to solve the problem of automatically extracting depth features, the present invention utilizes the Inception-V3 model and parameters pre-trained by ImageNet, and adopts The idea of transfer learning is trained to obtain a classification model. After the system is completed, it can realize real-time research on lens opacity through the mobile phone APP. And the data can be used as intermediate data for general reference of cortical lens opacity, nuclear lens opacity and posterior capsular lens opacity.

Dataset and Image Preprocessing

MSLPP dataset

The database is an important part of implementing a deep learning system, and a high-quality database can enhance the accuracy of system screening. However, due to the lack of large publicly available datasets of labeled slit-lamp ocular lens images, a dataset for lens opacity classification needs to be constructed. The data set used in the present invention is developed in cooperation with Shenyang Ai Luobo Intelligent Technology Co., Ltd. and Shenyang He's Ophthalmology Group, and it is named MSLPP (Marked Slit Lamp Picture Project) data set.

The MSLPP data set contains a total of 16,239 images, including 5,302 eye sample images of lens turbidity, 5,400 eye sample images of early lens turbidity, and 5,537 eye sample images of normal people. The images were collected from 2015 to 2018, from 2864 healthy people and 5532 lens opacities.

The images collected in this data set are all eye sample images taken by slit lamps. The slit lamps used are mainly desktop slit lamps and mobile phone slit lamps. Some sample images of the database are shown in Figure 2. It can be seen from the figure that when the slit light is focused to the pupil area, the inside is transparent or light yellow, which is a normal lens, as shown in Figure 2(a); the inside is transparent and yellow, and the light spot is slightly dark, which is the early lens Turbidity, as shown in Figure 2(b); if there is obvious turbidity inside, and the lesion can be seen, it means that the lens is cloudy, as shown in Figure 2(c).

According to the different parts of the eye lens opacity, it is divided into cortical lens opacity, nuclear lens opacity and posterior capsular lens opacity. In nuclear lens turbidity, the initial nucleus is yellow, and as the disease progresses, the color of the nucleus gradually becomes yellowish-brown, brown, brown-black or even black, as shown in Figure 2(c)-1; The formation of vacuoles and water gaps can be seen in the middle, and the water gaps expand from the periphery to the center, forming spoke-shaped opacities. Wedge-shaped opacities appear in the anterior and posterior cortex of the lens, which is feather-like. As the lens opacity deepens, the lens opacity increases until it becomes milky white Complete turbidity, as shown in Figure 2(c)-2; the posterior capsule lens is turbid, and a disk-shaped turbidity composed of many yellow dots, small vacuoles, and crystal-like particles under the posterior capsule can be seen under the slit lamp microscope. As shown in Figure 2(c)-3. In practice, nuclear lens opacities and cortical lens opacities are more common, while posterior capsular lens opacities are less common.

The MSLPP data set has the following characteristics:

(1) The samples were all collected from He's Eye Hospital, and the ophthalmologists used slit lamp equipment to take images of the eye crystals of the candidates to be screened, and the three attending physicians of He's Ophthalmology performed screening and classification;

(2) The sample collection environment is complex and diverse, involving bright environments, dark environments, and some reflections;

(3) The samples contain a variety of types, which can be used for nuclear lens opacities, cortical lens opacities and posterior capsular lens opacities.

light enhancement

Brightness is the most important part of image processing. During the collection of this data set, due to the complex and diverse actual screening environment, the brightness of the captured sample images varies greatly, which will affect the accuracy of deep learning. The impact is as follows:

Compress the input image size to 299×299, set the three-channel pixel of any point A(x,y) Then the average pixel value of the image Can be expressed as:

like Then the pixel value of each point in the image becomes like Then the pixel value of each point in the image becomes like Then the pixel value of each point in the image remains unchanged. Figure 3 is a comparison before and after illumination enhancement.

data augmentation

In order to avoid overfitting during model training, it is necessary to increase the number of samples during image preprocessing, which is conducive to improving the performance of the model and improving the accuracy of image classification. The data amplification mode that the present invention adopts has following three kinds:

(1) Translation: translate the image up, down, left, and right by 12 pixels

(2) Rotation: Rotate the image forward and reverse by 15°

(3) Mirror image: Mirror the image up, down, left, and right once

Due to the difference in light incidence between the desktop slit lamp and the portable slit lamp, the light of the desktop slit lamp is incident from the right at an angle of 30°, and the crystal section is displayed on the left side of the slit light, while the light of the hand-held slit lamp is incident from the left at an angle of 30° , the cut plane is displayed on the right side of the light. In order to eliminate the impact of this feature on the system, we will mirror the image left and right in the preprocessing. The specific comparison chart before and after amplification is shown in Figure 4, and the three lines from top to bottom are the comparison charts before and after the image quantity enhancement of lens opacity, early lens opacity and normal image respectively.

Models and Methods

convolutional neural network

The method adopted in the present invention is a convolutional neural network, which is mainly divided into three parts: a convolutional layer, a pooling layer and a fully connected layer. The final convolutional neural network structure is the Inception-V3 model proposed by Google based on GoogLeNet (ILSVRC2014, the champion model of the ImageNet Large Scale Visual Recognition Challenge of 2014 competition).

In the whole network, the convolutional layer and the pooling layer are used to extract the effective features in the image, the nonlinear activation function is introduced in the network, the dimension occupied by the effective features is reduced, and the advanced features that can represent the input image are output. Finally, the full connection The layers use these features to classify the input image to be screened. In addition to the basic network structure mentioned above, the present invention also adds a Dropout strategy in the last fully connected layer, which can effectively avoid over-fitting, improve the network generalization ability, and speed up the training process of the network.

transfer learning

In the field of medical images, the lack of a large number of publicly annotated datasets will be one of the difficulties in applying deep learning to medical image processing. In the case of insufficient samples, it is easy to cause a series of problems such as non-convergence in the model training process or poor generalization ability of the trained model. Therefore, the present invention adopts the method of transfer learning to solve the above problems.

The flow chart of the lens opacity classification method based on the convolutional neural network model transfer learning is shown in Figure 5. First, based on the ImageNet image annotation dataset, the Inception-V3 model is pre-trained to extract a 2048-dimensional feature vector. This stage makes full use of knowledge transfer, uses pre-trained weights for feature extraction, and does not train the weight parameters of Inception-V3. Compared with traditional methods, feature extraction is more efficient. Then, the feature vector is input into a single-layer fully connected neural network, because the trained Inception-V3 model has abstracted the original image into a feature vector that is easier to classify, so a single-layer fully connected neural network containing a Softmax classifier is used The network is trained on the classified cloudy crystal images of the lens to obtain the final classification result. At this stage, the input feature vector is mainly responsible for the training task of the classifier, so that the classifier can better complete the classification based on the extracted features.

experiment

Experimental data

The MSLPP dataset contains a total of 5302 turbid crystal images, 5400 early turbid crystal images, and 5537 normal crystal images. The present invention divides the data set into a training set, a verification set and a test set, wherein 500 samples are randomly selected from the three categories as a test set, and the remaining samples are randomly divided into a training set and a verification set according to a ratio of 6:1. Among them, the training set has a total of 12,630 images, including 4,083 lens opacities, 4,241 early lens opacities, and 4,306 normal images; the verification set has a total of 2,109 images, including 719 lens opacities, 659 early lens opacities, and 731 normal images. The number of pictures in each category is shown in Table 1.

Table 1 Comparison of the number of pictures under each category

The training set and validation set data are augmented, and the test set remains unchanged. After the expansion, the total number of training and verification sets increased from 14,739 to 132,651, of which 43,218 were lens slit lamp images of patients with lens opacity, and 44,100 were images of crystal slit lamp images of patients with early lens opacity. The number of crystal slit lamp images changed to 45333, and the specific number of each category is shown in Table 2.

Table 2 Quantity comparison before and after amplification

training process

All codes of the present invention are completed with Keras (Keras is a high-level neural network API written in pure Python) as the front end and TensorFlow (TensorFlow is the second-generation artificial intelligence learning system developed by Google) as the back end. The framework is based on Ubuntu16.04 (64-bit)+CUDA9.1+CUDNN9.0 system. The programming language used is python.

The training process is as follows:

(1) Load the Inception-V3 model with the fully connected layer removed and the weight parameters obtained by pre-training with the ImageNet dataset;

(2) Add a fully connected layer structure to the initialized Inception-V3 network, and add a Dropout strategy to the fully connected layer, and set the ratio to 0.75;

(3) Freeze all feature extraction layers except the fully connected layer, then set the learning rate to 0.001, use the preprocessed training set to train 1 epoch, and iterate 550 times; (the frozen feature extraction layer can only choose full connection Layer training, on the one hand, can prevent overfitting, on the other hand, since the training continues on a trained model, the weights should be updated in a small range to avoid destroying the pre-trained features)

(4) Unfreeze all layers, use fine-tune transfer learning (fine-tune), continue to use the MSLPP data set for training, use the stochastic gradient descent method, set the initial learning rate to 0.01, train 100 epochs, and each epoch iterates 550 Every time an epoch ends, use the verification set to test the accuracy of the model. If the accuracy is higher than last time, save the training parameters. If the accuracy is lower, use the previously saved parameters to continue training. The number of batch samples batch_size is set to 32, and the momentum momentum is set to 0.9.

Use quiver (a visualization tool for the Keras platform) to visualize the obtained model, and the obtained feature map is shown in Figure 6.

system assesment

After the model training is completed, the verification set is used to verify the model. The recall rate of lens opacity is 88.24%, the recall rate of early lens opacity is 86.63%, and the recall rate of normal lens is 97.51%. Then, we tested with the test set that the model had not been exposed to during training. The test set contained a total of 1,500 images, of which 500 were judged to be cloudy by the operator, 500 were early cloudy, and 500 were normal, which were classified by the system. The latter situation is shown in Table 3 below.

The present invention uses four commonly used indexes to evaluate the performance of the system: Accuracy, Recall, Precision and F1 index F1_meature. Accuracy is the overall measure of classification performance, which is the ratio of the number of correctly classified samples to the total number of samples; the recall rate is the proportion of all positive samples that are paired; the precision rate is the actual number of samples classified as positive samples. The proportion of positive examples; the F1 indicator is the harmonic mean of precision and recall. They are calculated as:

Among them, TP, TN, FP, and FN are the numbers of true positives, true negatives, false positives, and false negatives, respectively. Using the lens opacity sample as an example, a "true positive" means that the lens opacity sample was correctly classified as lens opacity. If a lens opacity sample is misclassified as another classification, we refer to this as a "false negative". The meanings of "true negative" and "false positive" are similar, with "true negative" meaning that other classified samples were not incorrectly classified as lens opacity, and "false positive" meaning that other classified samples were incorrectly classified as lens opacity.

According to the above evaluation method, the performance of the system is shown in Table 4 below.

Table 4 Model reliability judgment

A lens turbidity detection system based on a convolutional neural network, which includes an image preprocessing module and a detection module;

The image preprocessing module preprocesses the image to be tested to form preprocessed image data; the image data in the image preprocessing module is substituted into the lens turbidity detection learning model in the detection module to realize detection;

The detection module includes MSLPP dataset component module, image preprocessing module and model training module;

The MSLPP data set component module constructs the MSLPP data set; then the image in the data set is preprocessed by the image preprocessing module;

The model training module performs model training on the data set formed from the image preprocessed by the image preprocessing module to obtain a learning model.

Claims (10)

1. the lenticular opacities degree detecting method based on convolutional neural networks, it is characterised in that: the method steps are as follows:

(1), pretreatment is carried out using illumination enhancing method to testing image and forms preprocessing image data；

(2), the preprocessing image data in (1) step is substituted into lenticular opacities degree detecting learning model and realizes detection.

2. the lenticular opacities degree detecting method according to claim 1 based on convolutional neural networks, it is characterised in that:

(2) construction method of the lenticular opacities degree detecting model in step is as follows:

(2.1), MSLPP data set is constructed；Then the image concentrated to data pre-processes；And by pretreated MSLPP number It is divided into training set, verifying collection and test set according to collection；

(2.2), it is muddy that the data set progress model training formed using the image after the pretreatment in (2.1) step obtains crystalline lens Turbid degree detecting learning model.

3. the lenticular opacities degree detecting method according to claim 2 based on convolutional neural networks, it is characterised in that: (2.1) building mode of MSLPP data set is as follows in step: the eye crystal image acquired in clinic by slit-lamp is collected, and It is classified as normal, early stage lenticular opacities and lenticular opacities three classes.

4. the lenticular opacities degree detecting method according to claim 2 based on convolutional neural networks, it is characterised in that: (2.1) preprocess method in step are as follows: first data set is handled using illumination enhancing method, then to illumination enhancing method processing Data set afterwards is handled using data amplification.

5. the lenticular opacities degree detecting method according to claim 1 or 4 based on convolutional neural networks, feature exist In:

Illumination enhances method with the following method:

By 299 × 299 pixel of input image size boil down to, if the triple channel pixel of image any point A (x, y)For [R(x,y),G(x,y),B(x,y)]^T, wherein R (x, y), G (x, y), B (x, y) respectively represent red, green, blue three in point A (x, y) The brightness value in a channel, the range of luminance values in three channels are 0-255, then the average pixel value of imageIt indicates are as follows:

IfThen image each point pixel value becomesIfThen image each point pixel value becomes IfThen image each point pixel value is constant.

6. the lenticular opacities degree detecting method according to claim 5 based on convolutional neural networks, it is characterised in that:

2.1) the data amplification in the pretreatment of step has following three kinds, in three kinds of methods selection one of them, two or All:

(1) it translates: the enhanced image of illumination is translated to 5-20 pixel respectively up and down；So that image translation is expanded to Multiple originally, the multiple are the number of translation；

(2) it rotates: by the enhanced image of illumination along 5 ° -20 ° of reverse rotation；So that image rotation is expanded to original multiple, it should Multiple is the number of rotation；

(3) mirror image: by the enhanced image of illumination, each mirror image is primary up and down, that is, spins upside down once, and left and right overturning is primary, So that image mirrors are expanded to original multiple, which is the number of mirror image；

If expanded image all just for the enhanced figure of the same illumination when two or the whole of three kinds of methods of selection wherein As carrying out, then enhanced image is used together.

7. the lenticular opacities degree detecting method according to claim 2 based on convolutional neural networks, it is characterised in that: (2.2) step model training recycles pretreated MSLPP data set pair after carrying out transfer learning using convolutional neural networks Model after transfer learning continues to train；

Transfer learning and the step of training are as follows:

The first step carries out pre-training based on the data set of ImageNet image labeling on Inception-V3 model, extracts one The feature vector of a 2048 dimension；

Described eigenvector is inputted the full Connection Neural Network of a single layer by second step, is classified using one comprising Softmax The full Connection Neural Network of the single layer of device obtains final classification result after training using pretreated MSLPP data set.

8. the lenticular opacities degree detecting method according to claim 7 based on convolutional neural networks, it is characterised in that:

Pre-training process is carried out in the first step on Inception-V3 model are as follows:

(1) load has been removed the Inception-V3 model of full articulamentum and has been obtained with ImageNet data set pre-training Weight parameter；

(2) full articulamentum structure is added on Inception-V3 network after initialization, and is added in full articulamentum Dropout strategy, ratio are set as 0.75；Extract the feature vector of one 2048 dimension；

One full Connection Neural Network of single layer comprising Softmax classifier of use in second step, using pretreated The step of obtaining final classification result after the training of MSLPP data set is as follows:

(1) all feature extraction layers other than full articulamentum are freezed, learning rate is then set as 0.001, utilize pre- place After reason MSLPP data training set training 1 epoch, iteration 550 times；

(2) all layers are thawed, using fine tuning transfer learning, continues to be trained with MSLPP data training set, using boarding steps The method of decline is spent, initial learning rate is set as 0.01, trains 100 epoch, and each epoch iteration 550 times is every to terminate one Epoch collects test model accuracy rate using verifying, if accuracy rate was improved compared with last time, saves this training parameter, if accuracy rate drops It is low, then continue to train using previously stored parameter；It criticizes sample number batch_size and is set as 32, momentum momentum is set as 0.9。

9. the lenticular opacities degree detecting method according to claim 8 based on convolutional neural networks, it is characterised in that: (3) freeze feature extract layer only selects full articulamentum to be trained in step, should update weight in a small range, in order to avoid destroy The good feature of pre-training.

10. lenticular opacities degree-measuring system of the base according to claim 1 based on convolutional neural networks, feature exist In: the system includes image pre-processing module and detection module；

Image pre-processing module carries out pretreatment to testing image and forms preprocessing image data；Detection module is by image preprocessing Image data in module, which substitutes into, realizes detection in the lenticular opacities degree detecting learning model in detection module；

Detection module includes MSLPP data set construction part module, image pre-processing module and model training module；

MSLPP data set construction part module constructs MSLPP data set；Then image data concentrated by image pre-processing module It is pre-processed；

The data set that model training module forms the image after image pre-processing module pretreatment, which carries out model training, to be learned Practise model.