WO2020259209A1

WO2020259209A1 - Fundus image recognition method, apparatus and device, and storage medium

Info

Publication number: WO2020259209A1
Application number: PCT/CN2020/093415
Authority: WO
Inventors: 楼文杰; 王立龙; 朱军明; 吕传峰
Original assignee: 平安科技（深圳）有限公司
Priority date: 2019-06-26
Filing date: 2020-05-29
Publication date: 2020-12-30
Also published as: CN110400289A; CN110400289B

Abstract

The present application relates to the technical field of artificial intelligence. Provided are a fundus image recognition method, apparatus and device, and a storage medium. The method comprises: acquiring a fundus image; extracting first target data from the fundus image, and carrying out redundancy elimination processing on the first target data to obtain a first central fovea feature; generating a macular region mask according to the first central fovea feature; capturing a macular region in the fundus image by means of the macular region mask to obtain a macular region image; and recognizing an age-related macular degeneration focus feature in the macular region image, and classifying the macular region image according to the age-related macular degeneration focus feature. A macular region image is cut out from a fundus image by means of a mask generation model, the macular region image is classified according to an AMD focus feature in the macular region image, and feature data in the macular region image is obvious and can be easily recognized, thus effectively improving the accuracy of fundus image classification.

Description

Fundus image recognition method, device, equipment and storage medium

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on June 26, 2019, the application number is 201910560716.X, and the invention title is "Fundus Image Recognition Method, Apparatus, Equipment and Storage Medium". The reference is incorporated in this application.

Technical field

This application relates to the field of artificial intelligence, in particular to fundus image recognition methods, devices, equipment and storage media.

Background technique

Age-related Macular Degeneration, also known as Age-related Macular Degeneration (AMD), is one of the major blindness diseases in adults over 50. In fundus images, AMD lesions are often manifested as drusen in the macular area and abnormal pigmentation of the retina (such as hyperpigmentation or depigmentation).

The inventor found that currently, only a few machine learning algorithms are used to identify image features to detect age-related macular degeneration. At present, the commonly used method is to use mirror inspection, which is a technique used to capture fundus images and analyze different types of skin problems through these images. Fundus imaging technology can improve the accuracy in the process of diagnosing age-related macular degeneration. It uses optical magnification technology and polarized light technology to obtain images of age-related macular degeneration. Compared with the traditional microscopic imaging technology, the mirror image technology makes the fundus structure that cannot be observed by the naked eye appear more clearly in the image. However, if the doctor is not experienced, the accuracy of diagnosing age-related macular degeneration is still very low.

The inventor also found that with the development of technology, in clinical diagnosis, doctors have developed a variety of different diagnostic criteria based on the surface characteristics and growth characteristics of age-related macular degeneration. Among them, the more widely used diagnostic criteria include pattern analysis. , ABCD principle and seven-point inspection method, but these feature extraction methods are more complicated. In actual use, they generally need to rely on manual manual operations to complete, and these manual feature extraction methods can easily lead to the loss of some feature information, making diagnosis The recognition performance is not ideal, and the recognition is mainly based on the doctor's experience. If the doctor has insufficient experience, misjudgments will occur, so further improvement is needed.

technical problem

The main purpose of this application is to solve the technical problems of high labor cost and low efficiency of the existing AMD diagnosis methods. It proposes a fundus image recognition method that combines deep learning with fundus image recognition and detects the original fundus through the image quality recognition model The image quality of the image to obtain a fundus image that is easy to identify; the macular area image is cut out from the fundus image through the mask generation model, and the macular area image is classified according to the AMD lesion feature in the macular area image, and the features in the macular area image The data is obvious and easy to identify, thereby effectively improving the accuracy of fundus image classification.

Technical solutions

A method for recognizing a fundus image includes: acquiring a fundus image; extracting first target data from the fundus image through a convolutional layer of a convolutional neural network, and processing the first target data through a pooling layer of the convolutional neural network. The target data undergoes de-redundancy processing to obtain the first foveal feature; a macular area mask is generated according to the first foveal feature; the size of the macular area mask is the same as the size of the fundus image, and the macular area The mask includes a target interception area composed of a logic 1 array; the target interception area is used to intercept the macular area in the fundus image; the target interception area and the macular area in the fundus image are subjected to a bitwise AND operation , To obtain an image of the macular area; identify the age-related macular degeneration lesion feature in the macular area image through a lesion recognition model, and classify the macular area image according to the age-related macular degeneration lesion feature to obtain an image category.

Based on the same technical concept, this application also provides a fundus image recognition device, including:

The acquisition module is used to acquire fundus images.

The processing module is configured to extract first target data from the fundus image through the convolutional layer of the convolutional neural network, and perform de-redundancy processing on the first target data through the pooling layer of the convolutional neural network, Obtain a first foveal feature; generate a macular area mask according to the first foveal feature; the size of the macular area mask is the same as the size of the fundus image, and the macular area mask includes an array of logic 1 The target interception area; the target interception area is used to intercept the macular area in the fundus image; perform a bitwise AND operation between the target interception area and the macular area in the fundus image to obtain the macular area image; The lesion recognition model recognizes the features of the age-related macular degeneration lesion in the image of the macular area, and classifies the image of the macular area according to the feature of the age-related macular degeneration lesion to obtain an image category.

Based on the same technical concept, the present application also provides a computer device, including an input and output unit, a memory, and a processor. The memory stores computer-readable instructions that are executed by the processor. , Enabling the processor to execute the steps in the above-mentioned fundus image recognition method.

Based on the same technical concept, the present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, one or more processors can execute the above Steps in fundus image recognition method.

Beneficial effect

The beneficial effects of the present application: the first central foveal feature in the fundus image is recognized through the mask generation model, the macular area mask is generated according to the first central foveal feature, and the macular area in the fundus image is intercepted through the macular area mask , Get the macular area image, cut out the macular area image from the fundus image through the mask generation model; classify the macular area image according to the AMD lesion feature in the macular area image, the feature data in the macular area image is obvious and easy to identify, thus Effectively improve the accuracy of fundus image classification; in addition, in fundus images, the fovea is the easiest area to identify the macular area, and the amount of data is small; according to the first fovea feature, the macular area mask can be generated simply and accurately, Improve the data processing speed in the image recognition process.

Description of the drawings

FIG. 1 is a schematic flowchart of a method for recognizing fundus images in an embodiment of the application.

Fig. 2 is a schematic structural diagram of a fundus image recognition device in an embodiment of the application.

Fig. 3 is a schematic structural diagram of a computer device in an embodiment of the application.

The best mode of the invention

It should be understood that the specific embodiments described herein are only used to explain the application, and not used to limit the application.

Those skilled in the art can understand that unless specifically stated otherwise, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of this application refers to the presence of the features, procedures, steps, operations, elements and/or components, but does not exclude the presence or addition of one or more other features, Procedures, steps, operations, elements, components, and/or groups of them.

Figure 1 is a flowchart of a fundus image recognition method in some embodiments of this application. The fundus image recognition method is executed by a fundus image recognition device, which can be a computer or a mobile phone, etc., as shown in Figure 1, which can include The following steps S1-S5:

S1. Obtain a fundus image.

The fundus is the tissue behind the eyeball, the inner membrane of the eyeball, which contains the retina, optic papilla, macula, and central retinal artery and vein.

Fundus images are generated using fundus photography technology. Fundus photography is an effective method for early detection of occult eye diseases. It is of great significance for guiding the diagnosis and treatment of fundus diseases and assessing general health. Fundus photography is a fundus examination that has been used more clinically in recent years. It uses photographic equipment such as a digital camera to connect with a fundus lens. With the help of the fundus lens, the blood vessels on the retina behind the eyeball and the optic nerve can be clearly observed. Image of the retina. Fundus images can objectively record the morphological changes of the retina at the posterior pole of the fundus, and have good objectivity and comparability.

In some embodiments, before step S1, the fundus image recognition method further includes the following steps S11-S14:

S11. Obtain an original fundus image.

S12. Input the original fundus image to the image quality recognition model.

The image quality recognition model is used to identify the image clarity of the original fundus image, and identify the original fundus image that is fuzzy, too dark, too bright, lens contamination or abnormal angle, so as to ensure that a clear fundus image is finally obtained.

The image quality recognition model is pre-trained using training samples. In the process of training the image quality recognition model, the following loss function is used to calculate the error output by the image quality recognition model:

Among them, M represents the number of categories, M is an integer greater than or equal to 1; y _c represents the indicator variable, if the category c is the same as the category marked by the training sample, then y _c is equal to 1, otherwise y _c is equal to 0; p _c represents The predicted probability value of the training sample belonging to category c. The loss function calculates the error of the classification result of the image quality recognition model relative to the category marked by the training sample, and uses the reverse transmission method to update the parameters of each layer of the image quality recognition model according to the error. Repeat the above training until the internal network of the image quality recognition model converges, and the classification accuracy of the training samples reaches the preset requirements.

S13: Determine whether the image quality of the original fundus image is qualified according to the image clarity of the image quality recognition model.

The image quality of the original fundus image can be divided into blur, too dark, too bright, lens pollution, abnormal angle, and qualified. Among them, the fuzzy category, the too dark category, the over bright category, the lens pollution category and the angle abnormal category are all unqualified categories.

S14. If the quality of the original fundus image is qualified, set the original fundus image as the fundus image; if the quality of the original fundus image is unqualified, prompt to input a new original fundus image.

Affected by exposure or other noise, the image quality of the original fundus image taken by the camera may not be clear, and it is not sufficient for the recognition processing of the fundus image. In this embodiment, a number of the original fundus images are acquired in advance, and the image quality of the currently input original fundus image is recognized. If the image quality does not meet the preset standard, it will prompt to input other original fundus images. To ensure that the subsequent fundus image recognition results are accurate.

S2. Extract the first target data from the fundus image through the convolutional layer of the convolutional neural network, and perform de-redundancy processing on the first target data through the pooling layer of the convolutional neural network to obtain the first Foveal feature.

The first target data is data used to identify the fovea, for example, including the shape, size, color, reflective point, and position relative to the eyeball of the fovea. The fundus image is initially screened through the convolutional layer of the convolutional neural network to obtain the preliminary feature data of the fovea, that is, the first target data, and then the first target data is further processed through the pooling layer of the convolutional neural network Screening, filtering out the color of the fovea and the position of the fove relative to the eyeball, etc., to obtain data that facilitates the identification of the fovea, that is, the first fovea feature, which includes the shape, color, and Reflective point.

Pre-set a 3*3 width convolutional neural network. The convolutional neural network starts from the position of the starting pixel of the fundus image and uses 1 pixel as the step to gradually traverse the fundus image data and run the convolution operation , Extract the first central fovea feature in the fundus image. The convolutional neural network stitches the first central foveal feature into continuous data. Obtaining feature information from an image by a convolutional neural network belongs to the prior art, and will not be repeated here.

The posterior pole of the retina has a shallow funnel-shaped depression with a diameter of about 2mm, called the macula, which is named after the area is rich in lutein. There is a small fovea in the center of the macular fovea. There is no blood vessel in the macular area. However, because the pigment epithelial cells contain more pigments, the color is darker under the ophthalmoscope. There are reflective spots in the fovea, which is called foveal reflex. The most sensitive part of the retina is also the most recognizable area in the fundus image. Therefore, determining the location of the fovea in the fundus image can accurately and quickly determine the area of the macular area.

S3, generating a macular area mask according to the first central fovea feature.

The first central concave feature includes the brightness, shape, and pixel difference of the central concave from the surrounding area.

The size of the macular region mask is the same as the size of the fundus image. The macular region mask includes a target interception area and a shielding area. The target interception area is used to intercept the macular area in the fundus image; the shielding area is an area excluding the macular area in the macular area mask.

The macular area mask is a binary mask indicating the macular area in the fundus image, in the form of a sheet with only black (represented by logic 0) and white (represented by logic 1) the same size as the fundus image Image. The black area is the shielding area. The white area is the target capture area, and represents the macula area in the fundus image. In this way, the macular area in the fundus image can be distinguished from other areas through the mask.

In some embodiments, step S3 includes the following steps S31-S33:

S31. Determine the area where the central cavity is located in the fundus image according to the shape, color, and reflective point of the central cavity, and obtain a pixel point coordinate set of the area where the central cavity is located.

The first fovea feature is identified, that is, the area where the fovea is located in the fundus image is determined, and the coordinates of all pixels in the area where the fovea is located are obtained to obtain the pixel point coordinate set.

S32. Extract the coordinates of the center pixel point from the pixel point coordinate set, and generate a macular area mask of the same size as the fundus image according to the fundus image.

In step S32, extracting the coordinates of the center pixel point from the pixel point coordinate set includes the following steps:

Traverse the pixel point coordinate set to obtain the smallest abscissa, the largest abscissa, the smallest ordinate, and the largest ordinate in the pixel point coordinate set. According to the minimum abscissa and the maximum abscissa, the abscissa of the center pixel is obtained; and according to the minimum ordinate and the maximum ordinate, the ordinate of the center pixel is obtained.

In some embodiments, the coordinates of the central pixel point are obtained by using the following formula:

Wherein, x _mid represents the abscissa of the central pixel, y _mid represents the ordinate of the central pixel; x _min represents the smallest abscissa in the set of pixel coordinates, and x _max represents the set of pixel coordinates The largest abscissa; y _min represents the smallest ordinate in the pixel point coordinate set; y _max represents the largest ordinate in the pixel point coordinate set.

S33. Using the coordinates of the central pixel point as a center, generate the target interception area in a regular shape on the macular area mask to obtain the macular area mask.

The target interception area is centered on the coordinates of the central pixel point, and the length S is the total length of the contour.

Optionally, the target interception area is a square, and the expression of the length S is:

S=4L ₁

L ₁ ＝2*2*l＝4l

Wherein, L ₁ represents the side length of the target capturing area of the square; l represents the major axis of the optic disc in the fundus image.

Optionally, the target interception area is circular, and the expression of the length S is:

S=2*π*L ₂

L ₂ ＝2*l

Wherein, L ₂ represents the radius of the circular target capture area; l represents the major axis of the optic disc in the fundus image.

Clinically, drusen in the area around the fovea with the fovea as the center and twice the length of the optic disc as the radius has the most clinical statistical value. The fovea is the most recognizable feature of the macular area. In this embodiment, the fovea is used as a reference point to generate a macular area mask corresponding to the macular area, and the macular area in the fundus image is intercepted through the macular area mask. The method of identifying the macular area is simple and accurate, and the features of the lesion carried by the intercepted macular area are more obvious.

In some embodiments, the macular region mask is generated by a mask generation model; the first foveal feature is the input data of the mask generation model, and the macular region mask is the mask generation model. Output result;

Before step S1, the fundus image recognition method further includes the following steps S01-S04:

S01: Create the mask generation model.

S02: Preprocess multiple fundus images for training respectively to obtain multiple preprocessed training images.

The preprocessing includes image noise reduction, image size adjustment, image rotation, and image flipping.

S03. Obtain macular region mask samples corresponding to each preprocessed training image.

Manually pre-set macular area mask samples for each fundus image for training. The mask generation model calculates the deviation between the actual generated macular area mask and the pre-set macular area mask sample during the training process, and does it according to the magnitude of the deviation Self-parameter adjustment to achieve the purpose of training.

S04. Extract the second target data in each preprocessed training image through the convolutional layer of the convolutional neural network, and perform de-redundancy processing on each second target data through the pooling layer of the convolutional neural network. , Obtain a plurality of second foveal features, respectively input each second foveal feature into the mask generation model, use each macular region mask sample as the output reference of the mask generation model, and compare the mask The generation model is trained to make the parameters of the mask generation model converge.

The training of the mask generation model by taking each macular region mask sample as the output reference of the mask generation model includes the following steps S041-S042:

S041. Taking the macular region mask sample as a reference, the error of the macular region training mask output by the mask generation model is calculated through a loss function.

Optionally, the functional expression of the loss function is:

J_loss=-J(A, B);

Wherein, J_loss represents the error, A represents the macular area mask sample, B represents the macular area training mask output by the mask generation model, and J(A, B) represents the similarity coefficient (or called the jaccard coefficient). Its functional meaning, given two sets A and B, the jaccard coefficient is defined as the ratio of the size of the intersection of A and B to the size of the union. The larger the jaccard value, the higher the similarity. The loss function value J_loss is a negative jaccard coefficient, that is, J_loss=-J(A, B).

S042. Use a reverse conduction method to adjust the parameters of each layer of the entire mask generation model according to the error.

In this embodiment, there are 2595 training images, of which 80% are used for training and 20% are used for training verification. The training image is preset to a size of 128*128, and rotated by 90, 180, 270 degrees, and horizontal and vertical flip operations are performed for data enhancement. When the mask generation model is trained, the Adam optimizer is used to control the learning speed, the initial learning rate is set to 0.0001, and the parameters of each layer of the segmentation model are updated using the reverse transmission law.

In some embodiments, step S3 specifically includes: inputting the first foveal feature into the mask generation model, and outputting the macular region mask corresponding to the first foveal feature through the mask generation model. membrane.

In the foregoing embodiment, the macular region mask is generated using a pre-trained mask generation model. Generally, the more the number of training images, the higher the recognition accuracy of the mask generation model after training. The use of reverse conduction method to train the mask generation model has the advantages of fast training speed and easy implementation.

S4. Perform a bitwise AND operation on the target captured area and the macular area in the fundus image to obtain a macular area image.

The image of the macular area can be extracted from the original fundus image by performing a bitwise AND operation on the target intercepted area and the corresponding values of the shielding area and the fundus image.

S5: Recognizing the features of the age-related macular degeneration focus in the image of the macular area through a focus recognition model, and classifying the images of the macular area according to the feature of the age-related macular degeneration focus to obtain an image category.

According to the severity of AMD reflected in the macular area images, the macular area images are classified into ‘non-urgent’, ‘general emergency’, ‘urgent’ and ‘very urgent’.

The lesion classification model is obtained through training. Manually classify multiple macular area image samples used for classification training; input the labeled macular area image samples into the lesion classification model; the lesion classification model continuously updates its own parameters of each layer according to the macular area image samples until the lesion classification model The internal network converges.

In the foregoing embodiment, the image quality of the original fundus image is detected by the image quality recognition model to obtain the fundus image that is easy to recognize; the macular region image is cut out from the fundus image by the mask generation model; The AMD lesion feature classifies the image of the macular area to realize the automatic recognition of AMD lesions, which improves the diagnostic efficiency of AMD lesions and reduces labor costs.

Based on the same technical concept, the present application also provides a fundus image recognition device, which can be used to automatically recognize AMD lesion features in the macular area in the fundus image, and can provide a reference for the diagnosis of AMD. The device in the embodiment of the present application can implement the steps corresponding to the method for recognizing fundus images performed in the embodiment corresponding to FIG. 1 above. The functions realized by the device can be realized by hardware, or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above functions, and the modules may be software and/or hardware. As shown in Figure 2, the device includes an acquisition module 1 and a processing module 2. For the functional realization of the processing module 2 and the acquisition module 2, reference may be made to the operations performed in the embodiment corresponding to FIG. 1, which will not be repeated here. The processing module 2 can be used to control the receiving and sending operations of the acquiring module 1.

The acquisition module 1 is used to acquire fundus images.

The processing module 2 is configured to extract first target data from the fundus image through a convolutional layer of a convolutional neural network, and perform de-redundancy on the first target data through a pooling layer of the convolutional neural network. After remaining processing, the first foveal feature is obtained; the macular area mask is generated according to the first foveal feature; the size of the macular area mask is the same as the size of the fundus image, and the macular area mask includes the logic 1 The target capture area composed of an array; the target capture area is used to capture the macular area in the fundus image; the bitwise AND operation is performed on the target capture area and the macular area in the fundus image to obtain the macular area Image; Identify the focal features of the age-related macular degeneration in the image of the macular area through a lesion recognition model, and classify the image of the macular area according to the features of the age-related macular degeneration to obtain an image category.

In some embodiments, the processing module 2 is also used to obtain an original fundus image by using the obtaining module 1; and input the original fundus image into an image quality recognition model, and the image quality recognition model is used to identify the original fundus image. Picture definition; according to the picture definition output by the picture quality recognition model, judge whether the picture quality of the original fundus image is qualified; if the picture quality of the original fundus image is qualified, set the original fundus image to Describe the fundus image.

In some embodiments, the first central concave feature includes the shape, size, color, and reflective point of the central concave. The processing module 2 is specifically configured to determine the area of the central cavity in the fundus image according to the shape, color, and reflective point of the central cavity, and obtain the pixel point coordinate set of the area where the central cavity is located; Extracting the coordinates of the central pixel point from the pixel point coordinate set, and generating the macular area mask according to the fundus image; taking the coordinates of the central pixel point as the center, generating a regular shape on the macular area mask Target intercept area.

In some embodiments, the processing module is specifically configured to traverse the pixel point coordinate set to obtain the minimum abscissa, maximum abscissa, minimum ordinate, and maximum ordinate in the pixel point coordinate set; according to the minimum abscissa And the maximum abscissa to obtain the abscissa of the central pixel point; and according to the minimum ordinate and the maximum ordinate to obtain the ordinate of the central pixel point.

In some implementation manners, the expression of the coordinates of the center pixel point is:

In some embodiments, the macular region mask is generated by a mask generation model; the first foveal feature is the input data of the mask generation model, and the macular region mask is the mask generation model. Output the result.

The processing module 2 is also used to create the mask generation model; preprocess multiple fundus images for training to obtain multiple preprocessed training images; the preprocessing includes image noise reduction, image size adjustment, and image Rotate; respectively obtain the macular region mask samples corresponding to each preprocessing training image; extract the second target data in each preprocessing training image through the convolutional layer of the convolutional neural network, and use the convolutional neural network The pooling layer respectively performs de-redundancy processing on each second target data to obtain a plurality of second foveal features, and input each second foveal feature into the mask generation model to mask the sample with each macula As an output reference of the mask generation model, the mask generation model is trained to make the parameters of the mask generation model converge.

The processing module 2 is further configured to input the first central foveal feature into the mask generation model, and output the macular region mask corresponding to the first central foveal feature through the mask generation model.

In some embodiments, the processing module 2 is specifically configured to use the macular area mask sample as a reference to calculate the error of the macular area training mask output by the mask generation model through loss function calculation; adopt the reverse conduction method according to the The error adjusts the parameters of each layer of the entire mask generation model.

The functional expression of the loss function is:

J_loss=-J(A, B);

Wherein, J_loss represents the error, A represents the macular area mask sample, B represents the macular area training mask output by the mask generation model, and J(A, B) represents the similarity coefficient.

Based on the same technical concept, the present application also provides a computer device, as shown in FIG. 3, the computer device includes an input output unit 31, a processor 32, and a memory 33. The memory 33 stores computer readable instructions, When the computer-readable instructions are executed by the processor 32, the processor executes the steps of the fundus image recognition method in the foregoing embodiments.

The physical device corresponding to the acquisition module 1 shown in FIG. 2 is the input and output unit 31 shown in FIG. 3, which can realize part or all of the functions of the acquisition module 1, or realize the same or similar functions as the acquisition module 1. Features.

The physical device corresponding to the processing module 2 shown in FIG. 2 is the processor 32 shown in FIG. 3, and the processor 32 can implement part or all of the functions of the processing module 2 or implement the same or similar functions as the processing module 2.

Based on the same technical concept, this application also provides a storage medium storing computer-readable instructions. The computer-readable storage medium may be non-volatile or volatile. The computer-readable instructions are When executed by one or more processors, one or more processors are caused to execute the steps of the fundus image recognition method in the foregoing embodiments.

Through the description of the above embodiments, those skilled in the art can clearly understand that the method of the above embodiments can be implemented by means of software plus the necessary general hardware platform. Of course, it can also be implemented by hardware, but in many cases the former is better.的实施方式。 Based on this understanding, the technical solution of this application essentially or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM/RAM), including Several instructions are used to make a terminal (which may be a mobile phone, a computer, a server, or a network device, etc.) execute the method described in each embodiment of the present application.

The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are only illustrative and not restrictive. Those of ordinary skill in the art are Under the enlightenment of this application, without departing from the purpose of this application and the scope of protection of the claims, many forms can be made, any equivalent structure or equivalent process transformation made by using the content of the description and drawings of this application, or Directly or indirectly used in other related technical fields, these are all protected by this application.

Claims

A method for recognizing fundus images, which includes:

Acquire fundus images;

The first target data is extracted from the fundus image through the convolutional layer of the convolutional neural network, and the first target data is de-redundant through the pooling layer of the convolutional neural network to obtain the first central fovea feature;

Generating a macular area mask according to the first central fovea feature; the size of the macular area mask is the same as the size of the fundus image, and the macular area mask includes a target interception area composed of a logical 1 array; The target interception area is used to intercept the macular area in the fundus image;

Performing a bitwise AND operation on the target intercepted area and the macular area in the fundus image to obtain a macular area image;

Identifying the features of the age-related macular degeneration lesions in the macular area image through a lesion recognition model, classifying the macular area images according to the features of the age-related macular degeneration lesions, to obtain an image category.
The method for recognizing age-related macular degeneration according to claim 1, wherein, before the obtaining the fundus image, the method further comprises:

Obtain the original fundus image;

Inputting the original fundus image to an image quality recognition model, where the image quality recognition model is used to recognize the picture clarity of the original fundus image;

Judging whether the image quality of the original fundus image is qualified according to the image clarity output by the image quality recognition model;

If the quality of the original fundus image is qualified, the original fundus image is set as the fundus image.
2. The fundus image recognition method according to claim 2, wherein the image quality recognition model is pre-trained using training samples, and in the process of training the image quality recognition model, the following loss function is used to calculate the image quality recognition model Output error:

Among them, M represents the number of categories, M is an integer greater than or equal to 1; y c represents the indicator variable, if the category c is the same as the category marked by the training sample, then y c is equal to 1, otherwise y c is equal to 0; p c represents The predicted probability value of the training sample belonging to category c.
The method for recognizing fundus images according to claim 1, wherein the first foveal feature includes the shape, color, and reflective point of the fovea; and generating a macular area mask according to the first foveal feature comprises:

Determining the area where the central cavity is located in the fundus image according to the shape, color, and reflective point of the central cavity, and obtaining a pixel point coordinate set of the area where the central cavity is located;

Extracting the coordinates of the center pixel point from the pixel point coordinate set, and generating the macular area mask according to the fundus image;

Taking the coordinates of the central pixel point as the center, generating the target interception area in a regular shape on the macula mask.
The method for recognizing fundus images according to claim 4, wherein said extracting the coordinates of the center pixel point from the pixel point coordinate set comprises:

Traverse the pixel point coordinate set to obtain the smallest abscissa, the largest abscissa, the smallest ordinate, and the largest ordinate in the pixel point coordinate set;

Obtaining the abscissa of the central pixel according to the minimum abscissa and the maximum abscissa; and obtaining the ordinate of the central pixel according to the minimum ordinate and the maximum ordinate;

The expression of the abscissa of the central pixel is:

Wherein, x mid represents the abscissa of the center pixel, x min represents the minimum abscissa, and x max represents the maximum abscissa;

The expression of the ordinate of the center pixel is:

Wherein, y mid represents the ordinate of the center pixel point; y min represents the minimum ordinate; y max represents the maximum ordinate.
The method for recognizing a fundus image according to claim 1, wherein, before the obtaining the fundus image, the method further comprises:

Creating a mask generation model; the mask generation model is used to generate the macular region mask;

Preprocessing multiple fundus images for training to obtain multiple preprocessed training images; the preprocessing includes image noise reduction, image size adjustment, and image rotation;

Obtain respectively the macular area mask samples corresponding to each pre-processed training image;

The second target data in each preprocessed training image is extracted through the convolutional layer of the convolutional neural network, and each second target data is de-redundantly processed through the pooling layer of the convolutional neural network to obtain A plurality of second foveal features, each of the second foveal features are respectively input to the mask generation model, and each macular region mask sample is used as the output reference of the mask generation model, and the mask generation model Training so that the parameters of the mask generation model converge;

Correspondingly, the generating a macular area mask according to the first central fovea feature includes:

The first foveal feature is input to the mask generation model, and the macular region mask corresponding to the first foveal feature is output through the mask generation model.
The method for recognizing fundus images according to claim 6, wherein the training the mask generation model using each macular region mask sample as an output reference of the mask generation model comprises:

Taking the macular region mask sample as a reference, the error of the macular region training mask output by the mask generation model is calculated through a loss function;

Adjusting the parameters of each layer of the entire mask generation model according to the error by using a reverse conduction method;

The functional expression of the loss function is:

J_loss=-J(A, B)

Wherein, J_loss represents the error, A represents the macular area mask sample, B represents the macular area training mask output by the mask generation model, and J(A, B) represents the similarity coefficient.
A fundus image recognition device, which includes:

Obtaining module for obtaining fundus images;

The processing module is configured to extract first target data from the fundus image acquired by the acquisition module through the convolutional layer of the convolutional neural network, and perform processing on the first target through the pooling layer of the convolutional neural network. Data is de-redundantly processed to obtain a first foveal feature; a macular area mask is generated according to the first foveal feature; the first foveal feature includes the brightness and shape of the fovea and the pixel difference with the surrounding area; The size of the macular area mask is the same as the size of the fundus image, and the macular area mask includes a target capture area composed of a logical 1 array; the target capture area is used to capture the macular area in the fundus image Perform a bitwise AND operation on the target captured area and the macular area in the fundus image to obtain an image of the macular area; identify the focal feature of age-related macular degeneration in the image of the macular area through a lesion recognition model, The age-related macular degeneration lesion feature classifies the macular region image to obtain the image category.
8. The fundus image recognition device according to claim 8, wherein the processing module is further configured to obtain an original fundus image by using the acquisition module; and input the original fundus image into an image quality recognition model, the image quality recognition model Used to identify the picture definition of the original fundus image; determine whether the picture quality of the original fundus image is qualified according to the picture definition output by the picture quality recognition model; if the picture quality of the original fundus image is qualified, then The original fundus image is set as the fundus image.
8. The fundus image recognition device according to claim 8, wherein the first central fovea features include the shape, size, color, and reflective point of the central fovea; the processing module is specifically configured to perform according to the shape, color, and reflective point of the central fovea. Point to determine the area where the fovea is located in the fundus image, and obtain the pixel point coordinate set of the area where the fovea is located; extract the coordinates of the center pixel point from the pixel point coordinate set, and according to the The fundus image generates the macular area mask; taking the coordinates of the central pixel point as the center, a regular-shaped target interception area is generated on the macular area mask.
A computer device, which includes an input and output unit, a memory, and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes a Fundus image recognition method;

Wherein, the fundus image recognition method includes the following steps:

Acquire fundus images;

The first target data is extracted from the fundus image through the convolutional layer of the convolutional neural network, and the first target data is de-redundant through the pooling layer of the convolutional neural network to obtain the first central fovea feature;

Generating a macular area mask according to the first central fovea feature; the size of the macular area mask is the same as the size of the fundus image, and the macular area mask includes a target interception area composed of a logical 1 array; The target interception area is used to intercept the macular area in the fundus image;

Performing a bitwise AND operation on the target intercepted area and the macular area in the fundus image to obtain a macular area image;

Identifying the features of the age-related macular degeneration lesions in the macular area image through a lesion recognition model, classifying the macular area images according to the features of the age-related macular degeneration lesions, to obtain an image category.
The computer device according to claim 11, wherein, before said obtaining the fundus image, further comprising:

Obtain the original fundus image;

Inputting the original fundus image to an image quality recognition model, where the image quality recognition model is used to recognize the picture clarity of the original fundus image;

Judging whether the image quality of the original fundus image is qualified according to the image clarity output by the image quality recognition model;

If the quality of the original fundus image is qualified, the original fundus image is set as the fundus image.
11. The computer device according to claim 11, wherein the first central foveal feature includes the shape, color, and reflective point of the central fovea; and the generating of the macular area mask according to the first central foveal feature comprises:

Determining the area where the central cavity is located in the fundus image according to the shape, color, and reflective point of the central cavity, and obtaining a pixel point coordinate set of the area where the central cavity is located;

Extracting the coordinates of the center pixel point from the pixel point coordinate set, and generating the macular area mask according to the fundus image;

Taking the coordinates of the central pixel point as the center, generating the target interception area in a regular shape on the macula mask.
The computer device according to claim 13, wherein said extracting the coordinates of the center pixel point from the pixel point coordinate set comprises:

Traverse the pixel point coordinate set to obtain the smallest abscissa, the largest abscissa, the smallest ordinate, and the largest ordinate in the pixel point coordinate set;

Obtaining the abscissa of the central pixel according to the minimum abscissa and the maximum abscissa; and obtaining the ordinate of the central pixel according to the minimum ordinate and the maximum ordinate;

The expression of the abscissa of the central pixel is:

Wherein, x mid represents the abscissa of the center pixel, x min represents the minimum abscissa, and x max represents the maximum abscissa;

The expression of the ordinate of the center pixel is:

Wherein, y mid represents the ordinate of the center pixel; y min represents the minimum ordinate; y max represents the maximum ordinate.
The computer device according to claim 11, wherein, before said obtaining the fundus image, further comprising:

Creating a mask generation model; the mask generation model is used to generate the macular region mask;

Preprocessing multiple fundus images for training to obtain multiple preprocessed training images; the preprocessing includes image noise reduction, image size adjustment, and image rotation;

Obtain respectively the macular area mask samples corresponding to each pre-processed training image;

The second target data in each preprocessed training image is extracted through the convolutional layer of the convolutional neural network, and each second target data is de-redundantly processed through the pooling layer of the convolutional neural network to obtain A plurality of second foveal features, each of the second foveal features are respectively input to the mask generation model, and each macular region mask sample is used as the output reference of the mask generation model, and the mask generation model Training so that the parameters of the mask generation model converge;

Correspondingly, the generating a macular area mask according to the first central fovea feature includes:

The first foveal feature is input to the mask generation model, and the macular region mask corresponding to the first foveal feature is output through the mask generation model.
A storage medium storing computer-readable instructions, wherein when the computer-readable instructions are executed by one or more processors, the one or more processors execute a fundus image recognition method;

Wherein, the fundus image recognition method includes the following steps:

Acquire fundus images;

The first target data is extracted from the fundus image through the convolutional layer of the convolutional neural network, and the first target data is de-redundant through the pooling layer of the convolutional neural network to obtain the first central fovea feature;

Generating a macular area mask according to the first central fovea feature; the size of the macular area mask is the same as the size of the fundus image, and the macular area mask includes a target interception area composed of a logical 1 array; The target interception area is used to intercept the macular area in the fundus image;

Performing a bitwise AND operation on the target intercepted area and the macular area in the fundus image to obtain a macular area image;

Identifying the features of the age-related macular degeneration lesions in the macular area image through a lesion recognition model, classifying the macular area images according to the features of the age-related macular degeneration lesions, to obtain an image category.
The storage medium according to claim 16, wherein before said obtaining the fundus image, further comprising:

Obtain the original fundus image;

Inputting the original fundus image to an image quality recognition model, where the image quality recognition model is used to recognize the picture clarity of the original fundus image;

Judging whether the image quality of the original fundus image is qualified according to the image clarity output by the image quality recognition model;

If the quality of the original fundus image is qualified, the original fundus image is set as the fundus image.
The storage medium according to claim 16, wherein the first central foveal feature includes the shape, color, and reflective point of the central fovea; and the generating of the macular area mask according to the first central foveal feature comprises:

Determining the area where the central cavity is located in the fundus image according to the shape, color, and reflective point of the central cavity, and obtaining a pixel point coordinate set of the area where the central cavity is located;

Extracting the coordinates of the center pixel point from the pixel point coordinate set, and generating the macular area mask according to the fundus image;

Taking the coordinates of the central pixel point as the center, generating the target interception area in a regular shape on the macula mask.
The storage medium according to claim 18, wherein said extracting the coordinates of the center pixel point from the pixel point coordinate set comprises:

Traverse the pixel point coordinate set to obtain the smallest abscissa, the largest abscissa, the smallest ordinate, and the largest ordinate in the pixel point coordinate set;

Obtaining the abscissa of the central pixel according to the minimum abscissa and the maximum abscissa; and obtaining the ordinate of the central pixel according to the minimum ordinate and the maximum ordinate;

The expression of the abscissa of the central pixel is:

Wherein, x mid represents the abscissa of the center pixel, x min represents the minimum abscissa, and x max represents the maximum abscissa;

The expression of the ordinate of the center pixel is:

Wherein, y mid represents the ordinate of the center pixel point; y min represents the minimum ordinate; y max represents the maximum ordinate.
The storage medium according to claim 16, wherein before said obtaining the fundus image, further comprising:

Creating a mask generation model; the mask generation model is used to generate the macular region mask;

Preprocessing multiple fundus images for training to obtain multiple preprocessed training images; the preprocessing includes image noise reduction, image size adjustment, and image rotation;

Obtain respectively the macular area mask samples corresponding to each pre-processed training image;

The second target data in each preprocessed training image is extracted through the convolutional layer of the convolutional neural network, and each second target data is de-redundantly processed through the pooling layer of the convolutional neural network to obtain A plurality of second foveal features, each of the second foveal features are respectively input to the mask generation model, and each macular region mask sample is used as the output reference of the mask generation model, and the mask generation model Training so that the parameters of the mask generation model converge;

Correspondingly, the generating a macular area mask according to the first central fovea feature includes:

The first foveal feature is input to the mask generation model, and the macular region mask corresponding to the first foveal feature is output through the mask generation model.