Neural network-based glaucoma recognition device and recognition method

Abstract

The disclosure describes a neural network-based glaucoma recognition device and method, which includes an input unit, a preprocessing unit, a segmentation unit, a feature extraction unit, and a classification unit. The input unit is used to receive a first image. The preprocessing unit is used to preprocess the first image to obtain a first preprocessed image. The segmentation unit is used to input the first preprocessed image into an artificial neural network based on deep learning to generate an optic disc region image and an optic cup region image. The feature extraction unit obtains multiple glaucoma features based on the optic disc region image and the optic cup region image corresponding to the first preprocessed image. The classification unit is used to input feature information including glaucoma features corresponding to the first preprocessed image into a classifier based on machine learning for classification to obtain a glaucoma classification result. According to the present scheme, the accuracy of glaucoma recognition can be improved.

Description

[0001] This application is a divisional application of the patent application with the application number 2020107013737, the application date of July 18, 2020, and the invention name of training method, training device, identification method and identification system for glaucoma identification. TECHNICAL FIELD

[0002] The present disclosure specifically relates to a device and method for identifying glaucoma based on neural networks. BACKGROUND

[0003] At present, glaucoma has become the second leading cause of blindness in the world. There are more than 10 million patients with primary glaucoma worldwide, and more than 10% of them may develop bilateral blindness. If glaucoma is not treated early, it may develop into irreversible blindness, so early glaucoma screening is of great significance.

[0004] Among the technologies for glaucoma screening, fundus camera technology provides an economical and accurate way for early glaucoma screening. Medical research has shown that the cup-to-disc ratio (the ratio of the optic cup radius to the optic disc radius, simply referred to as cup-to-disc ratio) determined by fundus imaging can detect glaucoma in the early stage. With the development of artificial intelligence technology in recent years, artificial intelligence technology can be used to calculate the cup-to-disc ratio to achieve automatic identification of glaucoma, such as the method for automatic identification of glaucoma described in the patent document (CN109829877A). In the above patent document, an image processing algorithm is first used to preliminarily locate the optic disc in the fundus image, and a deep convolutional neural network is used to segment the optic disc region and the optic cup region from the preliminarily located region, then the cup-to-disc ratio is calculated and it is determined whether glaucoma exists in the fundus image.

[0005] However, in the above glaucoma identification method, a complex image processing algorithm is needed to preliminarily locate the optic disc, and the accuracy of the optic disc location will affect the accuracy of the subsequent segmentation of the optic cup or optic disc, and thus affect the accuracy of the cup-to-disc ratio calculation. In addition, in other existing glaucoma identification methods, the cup-to-disc ratio is used to identify glaucoma, and no other features of the optic cup or optic disc extracted by the convolutional neural network are used for glaucoma identification, so the accuracy of glaucoma identification still needs to be improved. SUMMARY

[0006] The present disclosure aims to provide a training method, training device, identification method and identification system for glaucoma identification that can accurately identify glaucoma in view of the above existing situation.

[0007] To this end, the first aspect of the present disclosure provides a training device based on glaucoma identification, comprising: an acquisition module that acquires fundus images and glaucoma classification labels thereof, preprocessed fundus images obtained by preprocessing the fundus images, and labeled images obtained by labeling the fundus images, the labeled images including optic disc labeled images labeling optic disc regions and optic cup labeled images labeling optic cup regions; an image segmentation network that is an artificial neural network based on deep learning and is trained by the preprocessed fundus images, the labeled images, and a spatial weighting map to output probabilities of each pixel point in the preprocessed fundus images belonging to the optic disc and belonging to the optic cup, and to generate optic disc region images and optic cup region images based on the probabilities of each pixel point in the preprocessed fundus images belonging to the optic disc and belonging to the optic cup, wherein the spatial weighting map is generated by weighting each pixel point in the preprocessed fundus images based on a preset distance threshold and an optic disc distance, the optic disc distance being the shortest distance from each pixel point in the preprocessed fundus images to the optic disc region in the optic disc labeled images, in the training of the artificial neural network, a loss function of each pixel point in the preprocessed fundus images is weighted based on the spatial weighting map to obtain a first loss function, and the artificial neural network is optimized based on the first loss function; a feature extraction module that acquires glaucoma features based on the optic disc region images and the optic cup region images; and a classifier that is trained by feature information including the glaucoma features and the glaucoma classification labels based on machine learning to output probabilities of belonging to glaucoma, in the training of the classifier, a second loss function is obtained, and the classifier is optimized based on the second loss function.

[0008] In the present disclosure, the image segmentation network is trained based on the preprocessed fundus images, the labeled images, and the spatial weighting map, the loss function of each pixel point in the preprocessed fundus images is weighted based on the spatial weighting map to obtain the first loss function, the image segmentation network is trained and optimized using the first loss function, the optic disc region images and the optic cup region images obtained using the image segmentation network are used to obtain the glaucoma features, the classifier is trained based on feature information including the glaucoma features and the glaucoma classification labels, and the classifier is optimized based on the second loss function to obtain a classifier capable of identifying glaucoma. In this case, the image segmentation network can be trained without preliminary positioning of the optic disc, the problem of inaccurate segmentation of the optic cup and the optic disc caused by inaccurate positioning of the optic disc is improved, and the features extracted based on the image segmentation network can be used to identify glaucoma in combination with the image segmentation network and the classifier. Thus, the accuracy of glaucoma identification can be improved.

[0009] In addition, in the training device for glaucoma identification according to the first aspect of the present disclosure, optionally, the glaucoma features include at least one of a vertical diameter of the optic disc in the optic disc region image and a vertical diameter of the optic cup in the optic cup region image, a horizontal diameter of the optic disc in the optic disc region image and a horizontal diameter of the optic cup in the optic cup region image, an area of the optic disc in the optic disc region image and an area of the optic cup in the optic cup region image. In this case, multiple features can be extracted based on the optic disc region image and the optic cup region image for glaucoma identification. Thus, the accuracy of glaucoma identification can be improved.

[0010] In addition, in the training device for glaucoma identification according to the first aspect of the present disclosure, optionally, by comparing the optic disc distance of each pixel point in the preprocessed fundus image with the preset distance threshold, the weight of the pixel point with the optic disc distance less than the preset distance threshold is a first preset value, and the weight of the pixel point with the optic disc distance greater than or equal to the preset distance threshold is a second preset value, wherein the first preset value is greater than the second preset value. In this case, the influence of the optic disc region can be improved, and the image segmentation network can be trained without preliminary positioning of the optic disc. Thus, the problem of inaccurate segmentation of the optic cup and the optic disc caused by inaccurate positioning of the optic disc can be improved, and the accuracy of glaucoma identification can be improved.

[0011] In addition, in the training device for glaucoma identification according to the first aspect of the present disclosure, optionally, the training device is optimized based on a total loss function determined according to the first loss function and the second loss function. In this case, the training device can be optimized using the total loss function. Thus, the accuracy of glaucoma identification can be improved.

[0012] In addition, in the training device for glaucoma identification according to the first aspect of the present disclosure, optionally, the feature information further includes at least one of age, gender, and medical history. In this case, the classifier can be trained based on different combinations of feature information. Thus, a classifier with better performance can be obtained.

[0013] The second aspect of the present disclosure provides a training method for glaucoma recognition, which comprises: obtaining fundus images and glaucoma classification labels thereof, preprocessed fundus images obtained by preprocessing the fundus images, and labeled images obtained by labeling the fundus images, wherein the labeled images comprise optic disc labeled images in which optic disc regions are labeled and optic cup labeled images in which optic cup regions are labeled; training a deep learning-based artificial neural network based on the preprocessed fundus images, the labeled images, and a spatial weighting map to output probabilities of each pixel point in the preprocessed fundus images belonging to the optic disc and belonging to the optic cup, and generating optic disc region images and optic cup region images based on the probabilities of each pixel point in the preprocessed fundus images belonging to the optic disc and belonging to the optic cup, wherein the spatial weighting map is generated by weighting each pixel point in the preprocessed fundus images based on a preset distance threshold and an optic disc distance, the optic disc distance being the shortest distance from each pixel point in the preprocessed fundus images to the optic disc region in the optic disc labeled images, in the training of the artificial neural network, a loss function of each pixel point in the preprocessed fundus images is weighted based on the spatial weighting map to obtain a first loss function, and the artificial neural network is optimized based on the first loss function; obtaining glaucoma features based on the optic disc region images and the optic cup region images; and training a machine learning-based classifier based on feature information comprising the glaucoma features and the glaucoma classification labels to output probabilities of belonging to glaucoma, in the training of the classifier, a second loss function is obtained, and the classifier is optimized based on the second loss function.

[0014] In the present disclosure, the artificial neural network is trained based on the preprocessed fundus images, the labeled images, and the spatial weighting map, the loss function of each pixel point in the preprocessed fundus images is weighted based on the spatial weighting map to obtain the first loss function, the artificial neural network is trained and optimized using the first loss function, the optic disc region images and the optic cup region images obtained using the artificial neural network are used to obtain the glaucoma features, the classifier is trained based on the feature information comprising the glaucoma features and the glaucoma classification labels, and the classifier is optimized based on the second loss function to obtain a classifier capable of recognizing glaucoma. In this case, the artificial neural network can be trained without preliminary positioning of the optic disc, the problem of inaccurate segmentation of the optic cup and the optic disc caused by inaccurate positioning of the optic disc can be improved, and the features extracted based on the artificial neural network can be used to recognize glaucoma in combination with the artificial neural network and the classifier. Therefore, the accuracy of glaucoma recognition can be improved.

[0015] In addition, in the training method for glaucoma identification, the glaucoma features can include at least one of a vertical diameter of the optic disc in the optic disc region image, a vertical diameter of the optic cup in the optic cup region image, a horizontal diameter of the optic disc in the optic disc region image, a horizontal diameter of the optic cup in the optic cup region image, an area of the optic disc in the optic disc region image, and an area of the optic cup in the optic cup region image. In this case, multiple features can be extracted based on the optic disc region image and the optic cup region image for glaucoma identification. Thus, the accuracy of glaucoma identification can be improved.

[0016] In addition, in the training method for glaucoma identification, the glaucoma features can include at least one of a vertical diameter of the optic disc in the optic disc region image, a vertical diameter of the optic cup in the optic cup region image, a horizontal diameter of the optic disc in the optic disc region image, a horizontal diameter of the optic cup in the optic cup region image, an area of the optic disc in the optic disc region image, and an area of the optic cup in the optic cup region image. In this case, multiple features can be extracted based on the optic disc region image and the optic cup region image for glaucoma identification. Thus, the accuracy of glaucoma identification can be improved.

[0017] In addition, in the training method for glaucoma identification, the glaucoma features can include at least one of a vertical diameter of the optic disc in the optic disc region image, a vertical diameter of the optic cup in the optic cup region image, a horizontal diameter of the optic disc in the optic disc region image, a horizontal diameter of the optic cup in the optic cup region image, an area of the optic disc in the optic disc region image, and an area of the optic cup in the optic cup region image. In this case, multiple features can be extracted based on the optic disc region image and the optic cup region image for glaucoma identification. Thus, the accuracy of glaucoma identification can be improved.

[0018] In addition, in the training method for glaucoma identification, the glaucoma features can include at least one of a vertical diameter of the optic disc in the optic disc region image, a vertical diameter of the optic cup in the optic cup region image, a horizontal diameter of the optic disc in the optic disc region image, a horizontal diameter of the optic cup in the optic cup region image, an area of the optic disc in the optic disc region image, and an area of the optic cup in the optic cup region image. In this case, multiple features can be extracted based on the optic disc region image and the optic cup region image for glaucoma identification. Thus, the accuracy of glaucoma identification can be improved.

[0019] In a third aspect, the present disclosure provides a glaucoma recognition method, which comprises: receiving a fundus image; preprocessing the fundus image to obtain a preprocessed fundus image; inputting the preprocessed fundus image into an artificial neural network obtained by the training method to obtain the probability of each pixel point in the preprocessed fundus image belonging to the optic disc and the probability of each pixel point belonging to the optic cup, and generating an optic disc region image and an optic cup region image based on the probability of each pixel point in the preprocessed fundus image belonging to the optic disc and the probability of each pixel point belonging to the optic cup; obtaining glaucoma features based on the optic disc region image and the optic cup region image; and inputting feature information comprising the glaucoma features into a classifier obtained by the training method for classification to obtain a glaucoma classification result. In the present disclosure, the received fundus image is preprocessed to obtain a preprocessed fundus image, the preprocessed fundus image is segmented by an artificial neural network obtained by the training method to generate an optic disc region image and an optic cup region image, glaucoma features are obtained based on the optic disc region image and the optic cup region image, and a classifier obtained by the training method is used to obtain a glaucoma classification result based on feature information comprising the glaucoma features. In this case, the preprocessed fundus image can be segmented without preliminary positioning of the optic disc, the problem of inaccurate segmentation of the optic cup and the optic disc caused by inaccurate positioning of the optic disc can be improved, and the artificial neural network and the classifier can be combined to recognize glaucoma by using features extracted based on the artificial neural network. Therefore, the accuracy of glaucoma recognition can be improved.

[0020] In a fourth aspect, the disclosure provides a glaucoma recognition system for glaucoma recognition, comprising: an input unit configured to receive a fundus image; a preprocessing unit configured to preprocess the fundus image to obtain a preprocessed fundus image; a segmentation unit configured to input the preprocessed fundus image into an artificial neural network obtained by the training method to obtain probabilities of each pixel point in the preprocessed fundus image belonging to an optic disc and belonging to an optic cup, and generate an optic disc region image and an optic cup region image based on the probabilities of each pixel point in the preprocessed fundus image belonging to the optic disc and belonging to the optic cup; a feature extraction unit configured to obtain glaucoma features based on the optic disc region image and the optic cup region image; and a classification unit configured to input feature information comprising the glaucoma features into a classifier obtained by the training method for classification to obtain a glaucoma classification result. In the disclosure, the preprocessing unit preprocesses the fundus image received by the input unit to obtain a preprocessed fundus image, the segmentation unit segments the preprocessed fundus image using the artificial neural network obtained by the training method to generate an optic disc region image and an optic cup region image, the feature extraction unit obtains glaucoma features based on the optic disc region image and the optic cup region image, and the classification unit obtains a glaucoma classification result based on the feature information comprising the glaucoma features using the classifier obtained by the training method. In this case, the preprocessed fundus image can be segmented without preliminary positioning of the optic disc, the problem of inaccurate segmentation of the optic cup and the optic disc caused by inaccurate positioning of the optic disc can be improved, and the artificial neural network and the classifier can be combined to recognize glaucoma using features extracted based on the artificial neural network. Therefore, the accuracy of glaucoma recognition can be improved.

[0021] According to the disclosure, a training method, a training device, a recognition method, and a recognition system for glaucoma recognition capable of accurately recognizing glaucoma are provided. BRIEF DESCRIPTION OF DRAWINGS

[0022] Embodiments of the disclosure will now be explained in further detail by way of example only with reference to the accompanying drawings, in which:

[0023] Figure 1 FIG. 1 is a schematic diagram of an electronic device illustrating a glaucoma recognition system according to an example of the disclosure.

[0024] Figure 2 FIG. 2 is a flowchart illustrating a training method for glaucoma recognition according to an example of the disclosure.

[0025] Figure 3 FIG. 3 is a schematic diagram illustrating annotation of a fundus image to form an annotated image according to an example of the disclosure.

[0026] Figure 4 FIG. 4 is a flowchart illustrating a training method for glaucoma recognition according to an example of the disclosure.

[0027] Figure 5 is a block diagram illustrating a training apparatus for glaucoma recognition involved in an example of the present disclosure.

[0028] Figure 6 is a block diagram illustrating a training apparatus for glaucoma recognition involved in an example of the present disclosure.

[0029] Figure 7 is a flowchart illustrating a recognition method for glaucoma recognition involved in an example of the present disclosure.

[0030] Figure 8 is a block diagram illustrating a recognition system for glaucoma recognition involved in an example of the present disclosure.

[0031] Label Explanation:

[0032] 1…electronic device, 110…input device, 120…server, 121…processor, 122…memory, 130…output device, P210…fundus image, P220…optic disc annotation image, P230…optic cup annotation image, 2…training apparatus, 210…acquisition module, 220…image segmentation network, 230…feature extraction module, 240…classifier, 250…optimization module, 3…recognition system, 310…input unit, 320…preprocessing unit, 330…segmentation unit, 340…feature extraction unit, 350…classification unit. DETAILED DESCRIPTION

[0033] Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the following description, the same parts are given the same symbol, and repeated description is omitted. In addition, the drawings are merely schematic diagrams, and the proportions of the sizes of the components with respect to each other or the shapes of the components, etc. can be different from actual ones.

[0034] Figure 1 is a block diagram illustrating a training apparatus for glaucoma recognition involved in an example of the present disclosure.

[0035] In some examples, with reference to Figure 1 , a recognition system for glaucoma recognition involved in the present disclosure (which can also be referred to simply as "recognition system") can be implemented with the aid of an electronic device 1. As Figure 1 indicated, the electronic device 1 can include an input device 110, a server 120, and an output device 130. The input device 110 is used to receive data. The server 120 can be used to process the data received by the input device 110 to obtain a processing result. The output device 130 is used to display the processing result obtained by the server 120.

[0036] The input device 110 can include, but is not limited to, a keyboard, a mouse, a touch screen, a scanner, a camera, and the like. The server 120 can include one or more processors 121 and one or more memories 122. The processor 121 can include a central processing unit, a graphics processing unit, and any other electronic component capable of processing data and capable of executing computer program instructions. The memory 122 can be used to store computer program instructions. The output device 130 can include, but is not limited to, a display, a printer, a projector, a plotter, and the like.

[0037] In some examples, the identification system can be stored in the memory 122 in the form of computer program instructions. The processor 121 classifies the fundus image received by the input device 110 to obtain the glaucoma classification result by executing the computer program instructions stored in the memory 122, and displays the obtained glaucoma classification result through the output device 130.

[0038] Hereinafter, the training method of glaucoma identification according to an example of the present disclosure will be described in detail. Figure 2 The training method of glaucoma identification according to an example of the present disclosure will be described in detail. Figure 2 is a flowchart illustrating the training method of glaucoma identification according to an example of the present disclosure.

[0039] In the present embodiment, the training method of glaucoma identification can include obtaining fundus images and glaucoma classification labels thereof, pre-processing and labeling the fundus images (step S210), training an artificial neural network based on deep learning (step S220), obtaining glaucoma features based on optic disc region images and optic cup region images (step S230), and training a classifier based on machine learning (step S240). In this case, the artificial neural network can be trained without preliminary positioning of the optic disc, the problem of inaccurate segmentation of the optic cup and the optic disc caused by inaccurate positioning of the optic disc can be improved, and the features extracted based on the artificial neural network can be used in combination with the artificial neural network and the classifier to identify glaucoma. Thus, the accuracy of glaucoma identification can be improved.

[0040] Figure 3 is a schematic diagram illustrating the labeling of fundus images to form labeled images according to an example of the present disclosure. Among them, Figure 3 (a) shows the fundus image P210, Figure 3 (b) shows the optic disc labeled image P220, Figure 3 (c) shows the optic cup labeled image P230.

[0041] In some examples, as described above, in step S210, the fundus images and the glaucoma classification labels thereof can be obtained, and the fundus images can be pre-processed and labeled.

[0042] In some examples, in step S210, the fundus image can be acquired. The fundus image can be an image about the fundus taken by a fundus camera or other fundus photographing device. As an example of the fundus image, for example Figure 3 (a) shows a fundus image P210 taken by a fundus camera.

[0043] In some examples, the fundus image can be a color fundus image. The color fundus image can clearly present rich fundus information such as optic disc, optic cup, macula, blood vessels, etc. In addition, the fundus image can also be one of images in modes such as RGB mode, CMYK mode, Lab mode or grayscale mode, etc.

[0044] In some examples, the fundus image can include the area of the optic disc and the optic cup. In medicine, the optic disc and the optic cup have clear anatomical definitions. That is, the optic disc is defined as the edge of the posterior opening of the sclera, and is bounded by the inner edge of the scleral ring. The optic cup is defined as the range between the scleral lamina cribrosa and the retinal plane.

[0045] In some examples, a plurality of fundus images can constitute a training data set. The training data set can include a training set and a testing set. For example, 500,000-2,000,000 fundus images from cooperating hospitals and removing patient information can be selected as the training set, and 5,000-20,000 fundus images can be selected as the testing set.

[0046] In addition, in some examples, in step S210, the glaucoma classification label of the fundus image can be acquired. In some examples, the glaucoma classification label can be formed by classifying the fundus images of patients by a plurality of clinical experts. Specifically, after collecting fundus images from cooperating hospitals and removing patient information, the fundus images can be classified according to the consultation results of more than 3 clinical experts to form the glaucoma classification label about the fundus images.

[0047] In some examples, the glaucoma classification label can include two types of labels of glaucoma and non-glaucoma. By letting the training method involved in the present disclosure learn two types of glaucoma classification of glaucoma and non-glaucoma, the training method has the ability to judge whether there is glaucoma lesion in the fundus image of the patient. In addition, in the present embodiment, the training method can also be further learned to judge which glaucoma lesion the fundus image of the patient is and to grade.

[0048] In addition, in some examples, in step S210, the preprocessed fundus image can be acquired. The preprocessed fundus image can be obtained by preprocessing the fundus image. In some examples, the preprocessing can include cropping and normalizing the fundus image.

[0049] In some examples, the fundus image can be cropped. Generally, since the fundus image obtained in step S210 can have different image formats and sizes, the fundus image can be cropped using manual processing or an image processing algorithm to convert the fundus image into a fixed standard form. The fixed standard form means that the image includes the fundus region of interest and the proportion of the optic disc region in the fundus region is consistent, and the format of the image is uniform and the size is the same.

[0050] In some examples, the size of the pre-processed fundus image can be uniform, such as 512x512 or 1024x1024 pixels.

[0051] In some examples, the fundus image can be normalized. Generally, since the fundus image obtained in step S210 can have non-uniform brightness and large contrast differences, the fundus image can be normalized to overcome the differences between different fundus images. In some examples, the normalization can be z-score standardization, so that the mean of the processed fundus image is 0 and the standard deviation is 1. In some examples, the normalization can be max-min normalization, which linearly transforms the original pixel values of the fundus image to map the pixel values to between 0 and 1.

[0052] In addition, in some examples, in step S210, the fundus image can be denoised, grayscale processed, etc. In some examples, the fundus image can be scaled, flipped, translated, etc. In this case, the amount of data for training the artificial neural network can be increased. Thus, the generalization ability of the artificial neural network can be improved. In addition, in some examples, in step S210, a labeled image can be obtained. The labeled image can be obtained by labeling the fundus image. In some examples, the labeled image can include an optic disc labeled image that labels the optic disc region and a cup labeled image that labels the cup region (see Figure 3 ).

[0053] In some examples, the optic disc labeled image and the cup labeled image in the labeled image can be merged into one labeled image as the true value of the artificial neural network, or can be separated into two labeled images as the true value of the artificial neural network.

[0054] In some examples, as described above, the labeled image can include an optic disc labeled image and a cup labeled image. For example, as shown in Figure 3 In some examples, the optic disc region in the fundus image P210 can be labeled to obtain an optic disc labeled image P220. The optic disc labeled image P220 can include an optic disc region A1 (see Figure 3(b)). In some examples, the optic cup region in the fundus image P210 can be labeled, thereby obtaining an optic cup labeled image P230. The optic cup labeled image P230 can contain the optic cup region A2 (see Figure 3 (c).

[0055] In some examples, the optic disc region and the optic cup region in the fundus image can be manually labeled by experienced doctors. In this way, the accuracy of the optic disc region and the optic cup region labeling can be improved. In some examples, the fundus image can be labeled using a dataset labeling tool, for example, the dataset labeling tool can be LabelImg tool (image labeling tool).

[0056] In some examples, while the fundus image is preprocessed, the labeled image can also be preprocessed. In this way, the size of the labeled image and the preprocessed fundus image size can always be consistent, thereby being more conducive to artificial neural network training.

[0057] In step S220, the artificial neural network based on deep learning can be trained. In some examples, the preprocessed fundus image and the labeled image obtained in step S210 can be obtained and combined with the spatial weighting map to train the artificial neural network based on deep learning. In other examples, the artificial neural network can also be trained directly with the fundus image.

[0058] In some examples, the artificial neural network based on deep learning can be trained based on the preprocessed fundus image, the labeled image and the spatial weighting map. In some examples, the artificial neural network based on deep learning can be an image semantic segmentation artificial neural network. For example, it can be an artificial neural network based on Unet network or its improved type. The Unet network is an image semantic segmentation artificial neural network, which can include a feature extraction and up-sampling part. The feature extraction part can include multiple encoding layers, for example, it can include five encoding layers (first encoding layer, second encoding layer, third encoding layer, fourth encoding layer and fifth encoding layer), and the first encoding layer takes the preprocessed fundus image as input. In some examples, the encoding layer can contain a series of convolutional layers, batch normalization layers, activation layers and max-pooling layers. The up-sampling part can include multiple decoding layers, for example, it can include four decoding layers (first decoding layer, second decoding layer, third decoding layer, fourth decoding layer), and the fourth decoding layer is used to output the probability of each pixel point in each preprocessed fundus image belonging to the optic disc and the probability of belonging to the optic cup. In this case, the artificial neural network can obtain the probability of each pixel point in the preprocessed fundus image belonging to the optic disc and the probability of belonging to the optic cup.

[0059] In some examples, the spatial weighting map can be generated by weighting each pixel in the preprocessed fundus image based on the preset distance threshold and the optic disc distance. In some examples, the spatial weighting map can be consistent with the size of the preprocessed fundus image. In the present disclosure, the spatial weighting map can be an image or a matrix. If the spatial weighting map is an image, the spatial weighting map can be consistent with the size of the preprocessed fundus image, and the pixel value of the spatial weighting map at the position corresponding to each pixel in the preprocessed fundus image is the weight of each pixel in the preprocessed fundus image. If the spatial weighting map is a matrix, the row and column of the matrix can be consistent with the size of the preprocessed fundus image. For example, if the size of the preprocessed fundus image is 512x512, the spatial weighting map is a 512x512 matrix. The element value of the spatial weighting map at the position corresponding to each pixel in the preprocessed fundus image is the weight of each pixel in the preprocessed fundus image.

[0060] In some examples, the preset distance threshold can be set according to the size of the optic disc region in the optic disc annotation image. In some examples, the preset distance threshold can be 0.1 times the diameter (vertical diameter or horizontal diameter) of the optic disc region in the optic disc annotation image. In some examples, the optic disc distance can be the shortest distance from each pixel in the preprocessed fundus image to the optic disc region in the optic disc annotation image.

[0061] As described above, the spatial weighting map can be generated by weighting each pixel in the preprocessed fundus image based on the preset distance threshold and the optic disc distance. Specifically, in some examples, the optic disc distance of each pixel in the preprocessed fundus image can be compared with the preset distance threshold. The weight of the pixel with the optic disc distance less than the preset distance threshold is set as a first preset value. The weight of the pixel with the optic disc distance greater than or equal to the preset distance threshold is set as a second preset value. The first preset value is greater than the second preset value. In this case, the influence of the optic disc region can be improved, and the artificial neural network can be trained without preliminary positioning of the optic disc. Thus, the problem of inaccurate optic cup and optic disc segmentation caused by inaccurate optic disc positioning can be improved, and the accuracy of glaucoma identification can be improved.

[0062] In some examples, the first preset value can be 0.8 to 1. For example, the first preset value can be 0.82, 0.85, 0.9, or 0.95, etc. The second preset value can be 0 to 0.2. For example, the second preset value can be 0.02, 0.05, 0.1, or 0.15, etc.

[0063] In some examples, in the training of the artificial neural network, the loss function of each pixel in the preprocessed fundus image can be weighted based on the spatial weighting map to obtain a first loss function, and the artificial neural network is optimized based on the first loss function.

[0064] Generally, the loss function can be used to calculate the loss, measure the pros and cons of the model prediction. The difference between the prediction value of the model based on artificial neural network and the true value of a single sample can be referred to as loss. The smaller the loss, the better the model. A single sample in the present disclosure can refer to each pixel point in the preprocessed fundus image.

[0065] In some examples, the Adam (adaptive moment estimation) optimization algorithm can be used to optimize the loss function. For example, the initial learning rate can be set to 0.001, and when the loss function stops decreasing during continuous multiple rounds of training of the artificial neural network, the learning rate can be reduced. In this way, the efficiency of training the artificial neural network can be improved, and the memory can be saved.

[0066] In some examples, the loss function can be a predefined loss function. In some examples, the loss function can be a cross-entropy loss function, a Dice loss function, etc. The cross-entropy loss function is a function that measures the difference between the true distribution and the predicted distribution, and the Dice loss function is a set similarity measure function. In this way, the appropriate loss function can be selected according to the needs, and the efficiency of training the artificial neural network can be improved.

[0067] Specifically, taking the cross-entropy loss function as an example, the loss function loss of each pixel point is: i,j

[0068]

[0069] where c represents the class of each pixel point in the preprocessed fundus image, and the predicted class includes two classes of optic cup or optic disc. (i,j) represents the coordinates of the pixel point in the preprocessed fundus image. represents the value of the pixel point with coordinates (i,j) in the optic cup annotation image or the optic disc annotation image, as the true value of the pixel point with coordinates (i,j) in the preprocessed fundus image. represents the predicted value of the pixel point with coordinates (i,j) in the preprocessed fundus image. α c is the weight of each class.

[0070] In some examples, the optic disc distance of each pixel point in the preprocessed fundus image can be compared with the preset distance threshold. The weight of the pixel point with the optic disc distance less than the preset distance threshold is set to a first preset value. The first preset value can be 1, for example. The weight of the pixel point with the optic disc distance greater than or equal to the preset distance threshold is set to a second preset value. The second preset value can be 0, for example. Then the pixel value or element value w i,j ​(i.e. the weight of the loss function of each pixel point in the pre-processed fundus image) is:

[0071]

[0072] where d i,j is the shortest distance from the pixel point (i, j) to the optic disc region in the optic disc region image. D is a preset distance threshold. In some examples, the preset distance threshold can be set as 0.1 times of the diameter (vertical diameter or horizontal diameter) of the optic disc region in the optic disc annotation image.

[0073] In some examples, the loss function of each pixel point can be spatially weighted by using the spatial weighting map, and thus the first loss function L1 of the artificial neural network is obtained:

[0074] L1 = å i,j (w i,j * loss i,j ) …… equation (3)

[0075] where w i,j is the weight of the pixel point with coordinates (i, j) in the pre-processed fundus image, and loss i,j is the loss function of the pixel point with coordinates (i, j) in the pre-processed fundus image. Thus, the artificial neural network can be trained based on the first loss function to optimize the output of the artificial neural network.

[0076] In addition, in some examples, in step S220, the optic disc region image and the optic cup region image can be generated based on the probability that each pixel point in the pre-processed fundus image belongs to the optic disc and the probability that each pixel point belongs to the optic cup. Specifically, in some examples, the optic disc region image and the optic cup region image can be generated based on the probability that each pixel point in the pre-processed fundus image belongs to the optic disc and the probability that each pixel point belongs to the optic cup in the fourth decoding layer of the Unet network described above. For example, the probability of the optic disc can be white when the probability is 100%, black when the probability is 0%, and gray when the probability is 50%. In some examples, the optic disc region image and the optic cup region image can also be grayscale images.

[0077] In step S230, the optic disc region image and the optic cup region image obtained in step S220 can be acquired, and the glaucoma features can be acquired based on the optic disc region image and the optic cup region image.

[0078] In some examples, the glaucoma features can be obtained from the optic disc region image and the optic cup region image. In some examples, the glaucoma features can include at least one of a vertical diameter of the optic disc in the optic disc region image and a vertical diameter of the optic cup in the optic cup region, a horizontal diameter of the optic disc in the optic disc region image and a horizontal diameter of the optic cup in the optic cup region, an area of the optic disc in the optic disc region image and an area of the optic cup in the optic cup region. In this case, a plurality of features can be extracted based on the optic disc region image and the optic cup region image for the identification of glaucoma. In this way, the accuracy of the glaucoma identification can be improved.

[0079] In step S240, the machine learning based classifier can be trained. In some examples, the machine learning based classifier can be trained based on the feature information and the glaucoma classification label to output a probability of belonging to glaucoma. The feature information can include the glaucoma features. The glaucoma features can be obtained by step S230. The glaucoma classification label can be obtained by step S210. The glaucoma classification label can include both glaucoma and non-glaucoma labels.

[0080] In some examples, the feature information can further include at least one of age, gender, medical history. In some examples, the feature information can be combined to form a plurality of feature combinations, and the classifier can be trained based on the plurality of feature combinations to obtain the classifier performance corresponding to each feature combination. For example, the feature combination can include the area of the optic disc in the optic disc region image, the area of the optic cup in the optic cup region, age, etc. In this case, the classifier can be trained based on different combinations of the feature information. In this way, the classifier with better performance can be obtained.

[0081] In addition, in some examples, the machine learning based classifier can include, but is not limited to, a classifier based on a random forest algorithm, a support vector machine algorithm, and a logistic regression algorithm, or an artificial neural network based on deep learning.

[0082] In some examples, in step S240, the probability of belonging to glaucoma can be output. In some examples, the probability of glaucoma can be the probability of whether the fundus image has glaucoma lesions. In some examples, a probability threshold can be set. For example, when the probability of glaucoma is greater than the probability threshold, it is identified as glaucoma. When the probability of glaucoma is less than or equal to the probability threshold, it is identified as non-glaucoma.

[0083] In addition, in some examples, in step S240, a second loss function can be obtained, and the classifier can be optimized based on the second loss function. In some examples, the second loss function can be a cross-entropy loss function. In this way, the classifier can be optimized using the second loss function.

[0084] The second loss function L2 can be as shown in equation (4):

[0085] L2 = -Y*log(Y pred ) - (1-Y)*log(1-Y pred ) …… Equation (4)

[0086] wherein Y represents the true value of the fundus image (i.e., the glaucoma classification label). Ypred represents the predicted value of the glaucoma of the fundus image output by the classifier.

[0087] Figure 4 is a flowchart illustrating a training method for glaucoma identification involved in examples of the present disclosure. In some examples, as shown in Figure 4 the training method can further include optimizing the trained artificial neural network and the classifier based on a total loss function (step S250). In this case, the total loss function can be used for optimization during the training of the artificial neural network and the classifier. Thus, the accuracy of glaucoma identification can be improved.

[0088] In some examples, the total loss function in step S250 can be determined according to the first loss function and the second loss function. In some examples, the first loss function and the second loss function can be assigned weights respectively, and the sum of the first loss function and the second loss function after assigning the weights is taken as the total loss function.

[0089] In some examples, by performing steps S210 to S240, the trained artificial neural network and the classifier can be obtained. In this case, the trained artificial neural network and the classifier can be further optimized based on the total loss function.

[0090] For example, taking the first loss function L1 and the second loss function L2 described above as an example, the total loss function L can be obtained based on the first loss function L1 and the second loss function L2, as shown in the following equation (5):

[0091] L = βL1 + γL2 …… Equation (5)

[0092] In the equation, β is the weight of the first loss function, and γ is the weight of the second loss function. In this case, the trained artificial neural network and the classifier can be optimized using the total loss function. Thus, the accuracy of glaucoma identification can be improved. In some examples, the artificial neural network and the classifier can also be trained directly based on the total loss function.

[0093] The training device for glaucoma identification of the present disclosure will be described in detail below. Figure 5 The training device for glaucoma identification of the present disclosure will be described in detail below. Figure 5 is a block diagram illustrating a training device for glaucoma identification involved in examples of the present disclosure.

[0094] In some examples, the various components of the training apparatus 2 correspond to the various steps of the training method described above and can be functional modules needed to implement the various steps of the training method described above. As shown in FIG. 2, the training apparatus 2 can include an acquisition module 210, an image segmentation network 220, a feature extraction module 230, and a classifier 240. Figure 5

[0095] In some examples, the acquisition module 210 can be configured to acquire fundus images and glaucoma classification labels thereof, and can be configured to pre-process and label the fundus images. The image segmentation network 220 can be an artificial neural network based on deep learning. The image segmentation network 220 can obtain optic disc region images and optic cup region images. The feature extraction module 230 can be configured to obtain glaucoma features based on the optic disc region images and the optic cup region images. The classifier 240 can be a classifier based on machine learning. The classifier 240 can obtain a probability of belonging to glaucoma. In this case, the image segmentation network can be trained without preliminary positioning of the optic disc, the problem of inaccurate segmentation of the optic cup and the optic disc caused by inaccurate positioning of the optic disc can be improved, and the features extracted based on the image segmentation network can be used to identify glaucoma in combination with the image segmentation network and the classifier. Thus, the accuracy of glaucoma identification can be improved.

[0096] In some examples, the acquisition module 210 can acquire fundus images. The fundus images can be images of the fundus taken by a fundus camera or other fundus photographing device. The fundus images can be one of images in RGB mode, CMYK mode, Lab mode, or grayscale mode, etc. In some examples, the acquisition module 210 can acquire glaucoma classification labels of the fundus images. The glaucoma classification labels can be two types of labels, glaucoma and non-glaucoma. In some examples, the acquisition module 210 can obtain pre-processed fundus images by pre-processing the fundus images. In some examples, the acquisition module 210 can obtain labeled images by labeling the fundus images. The labeled images can include optic disc labeled images labeled with optic disc regions and optic cup labeled images labeled with optic cup regions. For specific descriptions, see step S210, which will not be described here.

[0097] ​In some examples, the image segmentation network 220 can be an artificial neural network based on deep learning. In some examples, the image segmentation network 220 can be trained by pre-processing the fundus image, labeling the image, and spatially weighting the image. In some examples, the image segmentation network 220 can output probabilities that each pixel in the pre-processed fundus image belongs to the optic disc and belongs to the optic cup. In some examples, the image segmentation network 220 can generate an optic disc region image and an optic cup region image based on the probabilities that each pixel in the pre-processed fundus image belongs to the optic disc and belongs to the optic cup. For a detailed description, refer to step S220, which will not be repeated here.

[0098] In some examples, in the training of the image segmentation network 220, a loss function for each pixel in the pre-processed fundus image can be weighted based on the spatially weighted image to obtain a first loss function. In some examples, the artificial neural network can be optimized based on the first loss function. In some examples, the spatially weighted image can be generated by weighting each pixel in the pre-processed fundus image based on a pre-set distance threshold and the optic disc distance. Specifically, in some examples, the optic disc distance of each pixel in the pre-processed fundus image can be compared with the pre-set distance threshold. The weight of the pixel with the optic disc distance less than the pre-set distance threshold is set as a first pre-set value. The weight of the pixel with the optic disc distance greater than or equal to the pre-set distance threshold is set as a second pre-set value. The first pre-set value is greater than the second pre-set value. In this case, the influence of the optic disc region can be improved and the image segmentation network can be trained without preliminary positioning of the optic disc. Thus, the problem of inaccurate segmentation of the optic cup and the optic disc caused by inaccurate positioning of the optic disc can be improved, thereby improving the accuracy of glaucoma recognition. In some examples, the optic disc distance can be the shortest distance from each pixel in the pre-processed fundus image to the optic disc region in the optic disc labeled image. For a detailed description, refer to step S220, which will not be repeated here.

[0099] In some examples, the feature extraction module 230 can obtain glaucoma features based on the optic disc region image and the optic cup region image. In some examples, the glaucoma features can include at least one of the vertical diameter of the optic disc in the optic disc region image and the vertical diameter of the optic cup in the optic cup region, the horizontal diameter of the optic disc in the optic disc region image and the horizontal diameter of the optic cup in the optic cup region, the area of the optic disc in the optic disc region image and the area of the optic cup in the optic cup region. In this case, multiple features can be extracted based on the optic disc region image and the optic cup region image for glaucoma recognition. Thus, the accuracy of glaucoma recognition can be improved. For a detailed description, refer to step S230, which will not be repeated here.

[0100] In some examples, classifier 240 can be a machine learning-based classifier. In some examples, classifier 240 can be trained using feature information and glaucoma classification labels to output the probability of belonging to glaucoma. Feature information may include glaucoma features. In some examples, a second loss function can be obtained during classifier training, and the classifier can be optimized based on the second loss function. In some examples, feature information may also include at least one of age, gender, and medical history. In this case, the classifier can be trained based on different combinations of feature information. Thus, a classifier with better performance can be obtained. For a detailed description, please refer to step S240, which will not be repeated here.

[0101] Figure 6 This is a block diagram illustrating a training apparatus for glaucoma recognition as described in the examples of this disclosure. Figure 6 As shown, in some examples, the training device 2 also includes an optimization module 250.

[0102] In some examples, the optimization module 250 can optimize the training device 2 based on the total loss function. In some examples, the total loss function can be determined based on the first loss function and the second loss function. In some examples, weights can be assigned to the first loss function and the second loss function respectively, and the sum of the weighted first loss function and the second loss function can be used as the total loss function. In this case, the training device can be optimized using the total loss function. This can improve the accuracy of glaucoma identification. For a detailed description, please refer to step S250, which will not be repeated here.

[0103] The following combination Figure 7 The glaucoma identification method disclosed herein (also referred to simply as the "identification method") is described in detail. Figure 7 This is a flowchart illustrating an identification method for glaucoma identification as described in the examples of this disclosure.

[0104] In this embodiment, such as Figure 7 As shown, the identification method may include receiving a fundus image (step S310), preprocessing the fundus image (step S320), segmenting the preprocessed fundus image into an optic disc region image and an optic cup region image using an artificial neural network (step S330), obtaining glaucoma features (step S340), and obtaining a glaucoma classification result using a classifier (step S350). In this case, the preprocessed fundus image can be segmented without initial optic disc localization, which can improve the problem of inaccurate optic cup and optic disc segmentation caused by inaccurate optic disc localization. Furthermore, by combining the artificial neural network and the classifier, features extracted based on the artificial neural network can be used to identify glaucoma. Therefore, the accuracy of glaucoma identification can be improved.

[0105] In step S310, an eye fundus image can be received. The eye fundus image can be an image about an eye fundus taken by an eye fundus camera or other eye fundus photographing device. In other examples, the eye fundus image can be a picture pre-stored in a user terminal. The user terminal can include, but is not limited to, a notebook computer, a tablet computer, a mobile phone, a desktop computer, and the like. In some examples, the eye fundus image can be a color eye fundus image. The eye fundus image can be one of images in modes such as an RGB mode, a CMYK mode, a Lab mode, a grayscale mode, and the like.

[0106] In some examples, the identification method can be stored in a server in the form of a computer program, and the server can receive the eye fundus image by executing the computer program stored in the server.

[0107] In step S320, the eye fundus image received in step S310 can be acquired and pre-processed to obtain a pre-processed eye fundus image. In some examples, in the pre-processing process, operations such as cropping and normalization can be included. In this way, the eye fundus image can be converted into an image in a fixed standard form and the differences between different eye fundus images can be overcome. The fixed standard form means that the image includes an eye fundus region of interest and the proportion of the optic disc region in the eye fundus region is consistent, and the format of the image is uniform and the size is the same. In some examples, the size of the eye fundus image after pre-processing can be uniformly 512x512 or 1024x1024 pixels. In some examples, the eye fundus image can be subjected to noise reduction, grayscale processing, and the like. The cropping and normalization operations in the identification method can be analogous to the related descriptions of the cropping and normalization operations in step S210 of the training method described above.

[0108] In step S330, the pre-processed eye fundus image generated in step S320 can be input into the artificial neural network obtained using the training method described above to obtain the probability of each pixel point in the pre-processed eye fundus image belonging to the optic disc and the probability of each pixel point belonging to the optic cup. In some examples, the optic disc region image and the optic cup region image can be generated based on the probability of each pixel point in the pre-processed eye fundus image belonging to the optic disc and the probability of each pixel point belonging to the optic cup. In this way, based on the artificial neural network, the pre-processed eye fundus image can be segmented into the optic disc region image and the optic cup region image. The optic disc region image and the optic cup region image in the identification method can be analogous to the related descriptions of the optic disc region image and the optic cup region image in step S220 of the training method described above.

[0109] In step S340, glaucoma features can be obtained based on the optic disc region image and optic cup region image obtained in step S330. In some examples, the glaucoma features may include at least one of the following: the vertical diameter of the optic disc in the optic disc region image and the vertical diameter of the optic cup in the optic cup region; the horizontal diameter of the optic disc in the optic disc region image and the horizontal diameter of the optic cup in the optic cup region; and the area of ​​the optic disc in the optic disc region image and the area of ​​the optic cup in the optic cup region. In this case, multiple features can be extracted based on the optic disc region image and the optic cup region image for glaucoma identification. This improves the accuracy of glaucoma identification.

[0110] In step S350, the feature information can be input into the classifier obtained using the training method described above for classification to obtain a glaucoma classification result. In some examples, the feature information may include the glaucoma features obtained in step S340. In some examples, the feature information may also include at least one of age, gender, and medical history. In some examples, the glaucoma classification result can be either glaucoma or non-glaucoma.

[0111] The following combination Figure 8 The glaucoma identification system (also referred to simply as the "identification system") disclosed herein is described in detail. The identification system involved in this disclosure is used to implement the above-described identification method. Figure 8 This is a block diagram illustrating a glaucoma identification system as described in the examples of this disclosure.

[0112] In this embodiment, such as Figure 8 As shown, the glaucoma recognition system 3 may include an input unit 310, a preprocessing unit 320, a segmentation unit 330, a feature extraction unit 340, and a classification unit 350.

[0113] In some examples, input unit 310 can be used to receive fundus images. Preprocessing unit 320 can be used to preprocess the fundus images. Segmentation unit 330 can be used to segment the preprocessed fundus images into optic disc region images and optic cup region images using an artificial neural network. Feature extraction unit 340 can acquire glaucoma features. Classification unit 350 can be used to obtain glaucoma classification results using a classifier. In this case, segmentation of the preprocessed fundus images can be performed without initial optic disc localization, which can improve the problem of inaccurate optic cup and optic disc segmentation caused by inaccurate optic disc localization. Furthermore, by combining artificial neural networks and classifiers, features extracted based on artificial neural networks can be used to identify glaucoma. Thus, the accuracy of glaucoma identification can be improved.

[0114] In some examples, the input unit 310 can receive an eye fundus image. The eye fundus image can be an image about an eye fundus taken by an eye fundus camera or other eye fundus photographing device. In other examples, the eye fundus image can be a picture pre-stored in a user terminal. The user terminal can include, but is not limited to, a notebook computer, a tablet computer, a mobile phone, a desktop computer, etc. In some examples, the eye fundus image can be a color eye fundus image. The eye fundus image can be one of images in modes such as RGB mode, CMYK mode, Lab mode, or grayscale mode. For specific description, please refer to step S310, which will not be repeated here.

[0115] In some examples, the preprocessing unit 320 can obtain the eye fundus image received by the input unit 310 and pre-process the eye fundus image to obtain a pre-processed eye fundus image. In some examples, during the pre-processing, operations such as cutting and normalizing the eye fundus image can be included. In this way, the eye fundus image can be converted into an image in a fixed standard form and the differences between different eye fundus images can be overcome. In some examples, the eye fundus image can be subjected to noise reduction, grayscale processing, etc. For specific description, please refer to step S320, which will not be repeated here.

[0116] In some examples, the segmentation unit 330 can input the pre-processed eye fundus image generated in the preprocessing unit 320 into the artificial neural network obtained by using the above-mentioned training method, and can obtain the probability of each pixel point in the pre-processed eye fundus image belonging to the optic disc and the probability of each pixel point belonging to the optic cup. In some examples, the optic disc region image and the optic cup region image can be generated based on the probability of each pixel point in the pre-processed eye fundus image belonging to the optic disc and the probability of each pixel point belonging to the optic cup. In this way, based on the artificial neural network, the pre-processed eye fundus image can be segmented into the optic disc region image and the optic cup region image. For specific description, please refer to step S330, which will not be repeated here.

[0117] In some examples, the feature extraction unit 340 can obtain the glaucoma features based on the optic disc region image and the optic cup region image obtained by the segmentation unit 330. For specific description, please refer to step S340, which will not be repeated here.

[0118] In some examples, the classification unit 350 can input the feature information into the classifier obtained by using the above-mentioned training method to perform classification, so as to obtain the glaucoma classification result. In some examples, the feature information can include the glaucoma features obtained by the feature extraction unit 340. In some examples, the feature information further includes at least one of age, gender, and medical history. In some examples, the glaucoma classification result can be two classifications of glaucoma and non-glaucoma.

[0119] Although the present disclosure has been specifically shown and described with reference to the accompanying drawings and embodiments, it is understood that the above description is not in any way limiting on the present disclosure. Those skilled in the art can make modifications and changes to the present disclosure as needed without departing from the essential spirit and scope of the present disclosure, and such modifications and changes fall within the scope of the present disclosure.

Claims

1. A glaucoma identification device based on a neural network, characterized in that, The system includes an input unit, a preprocessing unit, a segmentation unit, a feature extraction unit, and a classification unit. The input unit receives a first image, which is a fundus image to be identified. The preprocessing unit preprocesses the first image to obtain a first preprocessed image, which is a preprocessed fundus image corresponding to the first image. The segmentation unit inputs the first preprocessed image into a deep learning-based artificial neural network to generate an optic disc region image and an optic cup region image. The artificial neural network is trained based on a second image, a labeled image corresponding to the second image, and a spatially weighted map generated from the optic disc distance of each pixel in the second preprocessed image corresponding to the second image, based on a preset distance threshold. The spatially weighted map is used to weight the loss function of each pixel in the second preprocessed image to obtain a classification image. The first loss function of the artificial neural network is optimized to train the artificial neural network without initial localization of the optic disc. The second image is a fundus image used to train the artificial neural network. The labeled images corresponding to the second image include an optic disc labeled image of the optic disc region and an optic cup labeled image of the optic cup region. The preset distance threshold is set according to the size of the optic disc region in the labeled image. The optic disc distance is the shortest distance from each pixel in the second preprocessed image to the optic disc region in the labeled image. The feature extraction unit obtains multiple glaucoma features based on the optic disc region image and the optic cup region image corresponding to the first preprocessed image. The classification unit is used to input the feature information including the glaucoma features corresponding to the first preprocessed image into a machine learning-based classifier for classification to obtain glaucoma classification results.

2. The identification device according to claim 1, characterized in that, The artificial neural network is based on the Unet network, or the artificial neural network is based on a modified type of the Unet network.

3. The identification device according to claim 1, characterized in that, The spatially weighted graph is an image or a matrix, and the spatially weighted graph has the same size as the second preprocessed image.

4. The identification device according to claim 1, characterized in that, The glaucoma features include at least one of the following: the vertical diameter of the optic disc in the optic disc region image and the vertical diameter of the optic cup in the optic cup region image; the horizontal diameter of the optic disc in the optic disc region image and the horizontal diameter of the optic cup in the optic cup region image; the area of ​​the optic disc in the optic disc region image and the area of ​​the optic cup in the optic cup region image.

5. The identification device according to claim 1, characterized in that, In the generated spatial weighted graph, the weight of pixels whose distance from the viewing disk is less than the preset distance threshold is set to a first preset value, and the weight of pixels whose distance from the viewing disk is greater than or equal to the preset distance threshold is set to a second preset value, wherein the first preset value is greater than the second preset value.

6. The identification device according to claim 1, characterized in that, The classifier is used to output the probability of belonging to glaucoma and is obtained by training based on feature information including glaucoma features corresponding to the second preprocessed image and glaucoma classification labels. The glaucoma features corresponding to the second preprocessed image are obtained by extracting the optic disc region image and optic cup region image generated by the artificial neural network for the second preprocessed image.

7. The identification device according to claim 6, characterized in that, The glaucoma classification labels are divided into two categories: glaucoma and non-glaucoma.

8. The identification device according to claim 1, characterized in that, During the training of the classifier, a second loss function is obtained and the classifier is optimized based on the second loss function. The classifier and the artificial neural network are optimized based on a total loss function, which is determined according to the first loss function and the second loss function.

9. A method for identifying glaucoma based on neural networks, characterized in that, include: Receive a first image, which is a fundus image to be identified; The first image is preprocessed to obtain a first preprocessed image, which is a preprocessed fundus image corresponding to the first image; The first preprocessed image is input into a deep learning-based artificial neural network to generate an optic disc region image and an optic cup region image. The artificial neural network is trained based on a second image, a corresponding labeled image of the second image, and a spatially weighted map generated by calculating the optic disc distance of each pixel in the second preprocessed image corresponding to the second image, based on a preset distance threshold. The spatially weighted map is used to weight the loss function of each pixel in the second preprocessed image to obtain a first loss function for optimizing the artificial neural network, so as to train the artificial neural network without performing preliminary optic disc localization. The second image is a fundus image used to train the artificial neural network. The labeled image corresponding to the second image includes an optic disc labeled image of the optic disc region and an optic cup labeled image of the optic cup region. The preset distance threshold is set according to the size of the optic disc region in the labeled image. The optic disc distance is the shortest distance from each pixel in the second preprocessed image to the optic disc region in the labeled image. Multiple glaucoma features are obtained based on the optic disc region image and optic cup region image corresponding to the first preprocessed image; and the feature information including the glaucoma features corresponding to the first preprocessed image is input into a machine learning-based classifier for classification to obtain glaucoma classification results.

Images