CN114334124A - A pathological myopia detection system based on deep neural network - Google Patents

Abstract

The invention provides a pathological myopia detection system based on a deep neural network, which comprises a computer memory, a computer processor and an executable program, wherein the executable program is used for receiving a fundus image, transmitting the fundus image into a trained pathological myopia detection network model and finally outputting a result; the pathological myopia detection network model comprises a focus detector, a pathological degree classifier and a pathological myopia discriminator; the lesion degree classifier is used for judging the lesion grade of the fundus image; if the image lesion grade is abnormal, the fundus image is sent to a lesion detector; the focus detector is used for detecting the focus type and position in the fundus image; and the pathological myopia discriminator judges whether the fundus image has pathological myopia according to the output results of the pathological degree classifier and the focus detector. The system can predict, classify and position the whole pathological degree of the eyeground and judge whether the eyeground suffers from pathological myopia or not.

Description

A pathological myopia detection system based on deep neural network

【Technical field】

The invention relates to the technical field of ophthalmic imaging, in particular to a pathological myopia detection system based on a deep neural network.

【Background technique】

Pathological myopia generally refers to a disease with a diopter greater than -6D and/or an eye axis greater than 26.5mm, and may be accompanied by degenerative diseases such as retroscleral staphyloma. About 3% of the world's population suffers from pathological myopia, which is one of the three major blindness diseases in the world, especially in Asians, and mostly occurs in young and middle-aged people. Pathological myopia generally has more characteristic fundus lesions. Based on previous studies, pathological myopic fundus lesions are divided into five grades: normal, leopard-striped fundus, diffuse choroidal atrophy, patchy choroidal atrophy, and macular atrophy. Pathological myopia is generally accompanied by complications, including choroidal neovascularization (CNV), lacquer cracks, and Fuchs spots, which have a greater impact on vision. In clinical diagnosis, the existence of these lesions is an important factor in judging the grade of pathological myopia.

At this stage, fundus camera imaging is the most commonly used ophthalmology clinical examination method, but there are shortcomings such as poor repeatability and different judgment standards for doctors when diagnosing fundus images. Therefore, there are many studies devoted to processing fundus images using machine learning methods such as deep learning to assist doctors in diagnosis.

For example, Chinese patent document with publication number CN111046835A discloses a fundus photography multi-disease detection system based on regional feature set neural network, the system includes a semantic segmentation sub-network and a plurality of classifiers. The operating steps of the system are as follows: obtaining the original fundus photos to be tested and inputting the multiple fundus disease detection network model. The Chinese patent document with publication number CN110163839A discloses a method for recognizing a leopard-striped fundus image. The method obtains a fundus image, and uses a machine learning model to classify the fundus image to obtain a classification result, and the classification result is used to represent the fundus. The degree of saliency of the leopard-like features of the image.

Existing fundus image analysis methods based on machine learning algorithms such as deep learning focus on the entire fundus to predict the degree or type of lesions, but lack information on the location and type of specific lesions. At the same time, for pathological myopia, a fundus disease that has received increasing attention in recent years, there is currently a lack of mature deep learning-based detection methods.

[Content of the invention]

The purpose of the present invention is to solve the problems in the prior art, and propose a pathological myopia detection system based on a deep neural network. Classification and localization, combined with the above information to determine whether the fundus has pathological myopia, and provide information on the location and type of lesions.

In order to achieve the above object, the present invention proposes a pathological myopia detection system based on a deep neural network, comprising a computer memory, a computer processor that can communicate with the computer memory, and is stored in the computer memory and can be used in the computer. An executable program executed on the processor, the executable program is used to receive the fundus image and transfer it to the trained pathological myopia detection network model, and finally output the result; the pathological myopia detection network model includes Lesion detector, lesion degree classifier and pathological myopia discriminator;

The lesion grade classifier is used to judge the lesion grade of the fundus image; if the lesion grade of the image is abnormal, the fundus image is sent to the lesion detector;

The lesion detector is used for detecting the type and position of the lesion in the fundus image;

The pathological myopia discriminator discriminates whether the fundus image suffers from pathological myopia according to the output results of the lesion degree classifier and the lesion detector.

Preferably, the specific training process of the pathological myopia detection network model includes the following steps:

S1. Fundus images are collected and marked as a data set for training. The data set includes a fundus lesion degree data set and a lesion detection data set.

S2. Preprocess the fundus images in the dataset to make them meet the network input requirements;

S3. Train the pathological myopia detection network model. The training process is divided into two stages:

In the first stage, use the fundus lesion degree dataset to train the lesion degree classifier;

In the second stage, the lesion detector is trained using the lesion detection dataset.

Preferably, the specific steps of the first stage are:

S31. The network parameters of the lesion degree classifier are all randomly initialized;

S32. The initial learning rate is set to 0.001, and the learning rate is decayed by cosine annealing;

S33. Using the stochastic gradient descent algorithm, all the data in the training set go through the network to complete one round, and after the network converges, the parameters of this part are fixed, and the training is completed.

Preferably, the specific steps of the second stage are:

S34. The parameters of the lesion detector backbone network are initialized from the lesion degree classifier network trained in the first stage, and the rest of the network parameters of the lesion detector are randomly initialized;

S35. After the network converges, a complete pathological myopia detection network model is obtained;

S36. Input the collected fundus image into the pathological myopia detection network model. The pathological myopia detection network model will predict whether the fundus image has pathological myopia. If the fundus image has pathological myopia, the pathological myopia detection network model will also output The location and type of lesions assist doctors in diagnosis.

Preferably, the fundus lesion degree data set is marked as the lesion degree of the fundus image, including five lesion grades: grade 0, grade 1, grade 2, grade 3, and grade 4, respectively representing: normal, leopard-striped fundus, diffuse choroidal atrophy, patchy choroidal atrophy, and macular atrophy, the labeling of the lesion detection dataset includes the location of the lesion and the type of the lesion, the location of the lesion includes the center point coordinates, width and height of the lesion, and the type of the lesion includes CNV , paint cracks and Fuchs spots.

Preferably, the specific prediction process of the pathological myopia detection network model includes the following steps:

a. Preprocess the fundus image to be predicted to make it meet the input requirements of the pathological myopia detection network model;

b. The fundus image is first input to the lesion degree classifier network to predict the degree of lesion. If the prediction result is normal: level 0, the prediction is over, and the pathological myopia detection network model will output non-PM(0);

c. If the prediction result is abnormal: level 1, level 2, level 3, level 4, the fundus image will enter the lesion detector network, and the lesions detected by the lesion detector network will be marked, and the type of lesions and predictions will be marked. probability;

d. The pathological myopia discriminator integrates the output information of the lesion detector network and the lesion degree classifier network to determine whether the fundus suffers from pathological myopia;

e. The pathological myopia detection network model outputs the lesion degree, lesion location and type of the input fundus image, and whether the fundus image belongs to pathological myopia.

Preferably, in step d, the specific method for judging whether the fundus suffers from pathological myopia is: if the degree of fundus lesions is normal: grade 0, or the degree of fundus lesions is leopard-striped fundus: grade 1, but no lesions are detected , it is judged as non-pathological myopia, and other cases are judged as pathological myopia.

Preferably, the lesion degree classifier includes a deep residual network, a fully connected layer and a Softmax classification layer.

Preferably, the lesion detector includes a backbone network, an RPN network, a ROI-Pooling layer and a classifier.

Preferably, the pathological myopia discriminator is composed of logical judgment sentences.

Beneficial effects of the present invention:

1. Compared with the existing fundus image analysis model based on neural network, the present invention can not only predict the overall lesion degree or type of the fundus picture, but also predict the specific lesion location and type.

2. The present invention can simultaneously predict the lesion grade of the fundus image, detect the position of the lesion in the fundus image, and combine the two information to judge whether the sample is pathological myopia.

The features and advantages of the present invention will be explained in detail by way of examples.

【Detailed ways】

The present invention is a pathological myopia detection system based on a deep neural network, which is composed of the following parts: a computer memory, a computer processor and an executable program. The executable program receives the fundus image, transmits it to the trained pathological myopia detection network model, and finally outputs the result.

The pathological myopia detection network model includes a lesion detector, a lesion degree classifier and a pathological myopia discriminator. The lesion degree classifier is used to determine the lesion grade of the fundus image (see Table 1 for specific grades); if the lesion grade of the fundus image is not 0, the fundus image is sent to the lesion detector, which is used to detect the fundus. The type and location of the lesion in the image; the pathological myopia discriminator discriminates whether the fundus image suffers from pathological myopia according to the output results of the lesion degree classifier and the lesion detector.

Table 1

The lesion degree classifier includes a deep residual network (ResNet-101), a fully connected layer (FCLayer), and a Softmax classification layer. The structure of ResNet-101 includes the first convolutional layer, the second convolutional layer group, the third convolutional layer group, the fourth convolutional layer group, and the fifth convolutional layer group.

The lesion detector borrows the Faster-RCNN target detection network structure, which specifically includes a backbone network (ResNet-101), an RPN network, a ROI-Pooling layer and a classifier.

The pathological myopia discriminator is composed of logical judgment sentences.

The specific training process of the pathological myopia detection network model is as follows:

(1) Collect fundus images and label them as a training dataset. The dataset is divided into two parts. One is the fundus lesion degree dataset, which is labeled as the fundus image lesion degree (0, 1, 2) , 3, 4); the second is the lesion detection dataset, the annotation of the lesion detection dataset includes: lesion location (coordinates of center point, width, height) and lesion type (CNV, paint crack, Fuchs spot).

(2) Preprocess the fundus images in the dataset to make them meet the network input requirements.

(3) The pathological myopia detection network model is trained, and the training process is divided into two stages:

The first stage: using the fundus lesion degree dataset to train the lesion degree classifier.

This part of the network parameters are initialized randomly.

The initial learning rate is set to 0.001, and the learning rate decay is adopted with cosine annealing.

Using the stochastic gradient descent algorithm, all the data in the training set go through the network to complete one round, and after the network converges, the parameters of this part are fixed, and the training is completed.

Stage 2: The lesion detector is trained using the lesion detection dataset.

In the lesion detector network, the parameters of the backbone network ResNet101 are initialized from the lesion degree classifier network trained in stage one. The rest of the network parameters are initialized randomly.

The selection of training method and loss function are all references (S.Ren, K.He, R.Girshick and J.Sun, "Faster R-CNN: Towards Real-Time Object Detection with Region ProposalNetworks," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137-1149, 1 June 2017, doi: 10.1109/TPAMI.2016.2577031).

After the network converges, a complete pathological myopia detection network model is obtained.

Input the collected fundus images into the pathological myopia detection network model, and the network will predict whether the fundus image has pathological myopia.

The specific prediction process of the pathological myopia detection network model is as follows:

The fundus images to be predicted are preprocessed to meet the input requirements of the network model.

The fundus image will first be input into the lesion degree classifier network to predict the degree of lesion. If the prediction result is normal (0), the prediction is over, and the pathological myopia detection network model will output non-PM (0).

If the predicted result is abnormal (1, 2, 3, 4), the fundus image will enter the lesion detection network, and the lesions detected by the network will be framed by a box, and the type of the lesion and the predicted probability will be marked.

The pathological myopia discriminator integrates the output information of the lesion detector network and the lesion degree classifier network to determine whether the fundus has pathological myopia. The specific method is: if the degree of fundus lesions is normal (0) or the degree of fundus lesions is leopard print If no lesions were detected, it was judged as non-pathological myopia, and in other cases, it was judged as pathological myopia.

Finally, the pathological myopia detection network will output the lesion degree, lesion location and type of the input fundus image, and whether the fundus image belongs to pathological myopia.

The pathological myopia detection network model of the present invention realizes the grading of myopic macular degeneration, has high grading efficiency and high precision, and can clinically target pathological myopia, a disease that is increasingly concerned, and provide powerful AI technology-assisted support helps doctors screen and diagnose diseases quickly and effectively.

The above-mentioned embodiments are illustrative of the present invention, not limitations of the present invention, and any scheme after simple transformation of the present invention belongs to the protection scope of the present invention.

Claims (10)

1. A pathological myopia detection system based on a deep neural network, comprising a computer memory, a computer processor that can communicate with the computer memory, and an executable executable stored in the computer memory and executable on the computer processor The program is characterized in that: the executable program is used to receive the fundus image, and transmit it to the trained pathological myopia detection network model, and finally output the result; the pathological myopia detection network model includes a lesion detector. , lesion degree classifier and pathological myopia discriminator; The lesion grade classifier is used to judge the lesion grade of the fundus image; if the lesion grade of the image is abnormal, the fundus image is sent to the lesion detector; The lesion detector is used for detecting the type and position of the lesion in the fundus image; The pathological myopia discriminator discriminates whether the fundus image suffers from pathological myopia according to the output results of the lesion degree classifier and the lesion detector. 2. a kind of pathological myopia detection system based on deep neural network as claimed in claim 1 is characterized in that: the concrete training process of described pathological myopia detection network model comprises the following steps: S1. Fundus images are collected and marked as a data set for training. The data set includes a fundus lesion degree data set and a lesion detection data set. S2. Preprocess the fundus images in the dataset to make them meet the network input requirements; S3. Train the pathological myopia detection network model. The training process is divided into two stages: In the first stage, use the fundus lesion degree dataset to train the lesion degree classifier; In the second stage, the lesion detector is trained using the lesion detection dataset. 3. a kind of pathological myopia detection system based on deep neural network as claimed in claim 2 is characterized in that: the concrete steps of the first stage are: S31. The network parameters of the lesion degree classifier are all randomly initialized; S32. The initial learning rate is set to 0.001, and the learning rate is decayed by cosine annealing; S33. Using the stochastic gradient descent algorithm, all the data in the training set go through the network to complete one round, and after the network converges, the parameters of this part are fixed, and the training is completed. 4. a kind of pathological myopia detection system based on deep neural network as claimed in claim 2 is characterized in that: the concrete steps of second stage are: S34. The parameters of the lesion detector backbone network are initialized from the lesion degree classifier network trained in the first stage, and the rest of the network parameters of the lesion detector are randomly initialized; S35. After the network converges, a complete pathological myopia detection network model is obtained; S36. Input the collected fundus image into the pathological myopia detection network model. The pathological myopia detection network model will predict whether the fundus image has pathological myopia. If the fundus image has pathological myopia, the pathological myopia detection network model will also output The location and type of lesions. 5. A kind of pathological myopia detection system based on deep neural network as claimed in claim 2, it is characterized in that: described fundus lesion degree data set is marked as fundus image lesion degree, including five lesion grades: grade 0, Level 1, Level 2, Level 3, and Level 4 represent: normal, leopard-like fundus, diffuse choroidal atrophy, patchy choroidal atrophy, and macular atrophy. The labeling of the lesion detection dataset includes the location of the lesion and the type of the lesion. The location of the lesion includes the center point coordinates, width and height of the lesion, and the type of the lesion includes CNV, lacquer crack and Fuchs spot. 6. a kind of pathological myopia detection system based on deep neural network as described in any one in claim 1 to 5, is characterized in that: described pathological myopia detection network model concrete prediction process comprises the following steps: a. Preprocess the fundus image to be predicted to make it meet the input requirements of the pathological myopia detection network model; b. The fundus image is first input to the lesion degree classifier network to predict the degree of lesion. If the prediction result is normal: level 0, the prediction is over, and the pathological myopia detection network model will output non-PM(0) c. If the prediction result is abnormal: level 1, level 2, level 3, level 4, the fundus image will enter the lesion detector network, and the lesions detected by the lesion detector network will be marked, and the type of lesions and predictions will be marked. probability; d. The pathological myopia discriminator integrates the output information of the lesion detector network and the lesion degree classifier network to determine whether the fundus suffers from pathological myopia; e. The pathological myopia detection network model outputs the lesion degree, lesion location and type of the input fundus image, and whether the fundus image belongs to pathological myopia. 7. a kind of pathological myopia detection system based on deep neural network as claimed in claim 6, it is characterized in that: in step d, the concrete method of judging whether this fundus suffers from pathological myopia is: if fundus lesion degree is normal , or the degree of fundus lesions is leopard-like fundus, but no lesions are detected, it is judged as non-pathological myopia, and other cases are judged as pathological myopia. 8. A deep neural network-based pathological myopia detection system as claimed in claim 1 or 2, wherein the lesion degree classifier comprises a deep residual network, a fully connected layer and a Softmax layer classification layer. 9. A deep neural network-based pathological myopia detection system as claimed in claim 1 or 2, wherein the lesion detector comprises a backbone network, an RPN network, a ROI-Pooling layer and a classification device. 10 . The pathological myopia detection system based on a deep neural network according to claim 1 or 2 , wherein the pathological myopia discriminator is composed of logical judgment sentences. 11 .