Image recognition model construction method, fundus image recognition method, electronic equipment and storage medium

Abstract

The invention relates to the technical field of image recognition, particularly provides an image recognition model construction method, an eye fundus image recognition method, electronic equipment and a storage medium, and aims to solve the problem that model construction is relatively difficult due to lack of labeled samples. Therefore, the image recognition model construction method comprises the steps that a training data set is acquired, and the training data set comprises multiple pieces of region-of-interest image data; an image feature extraction model is constructed, the image feature extraction model comprises an encoder and a generator, and the encoder is composed of an introspection variational auto-encoder, a compression excitation network and a deep separable convolutional network; training the image feature extraction model based on the training data set to obtain a trained image feature extraction model; and constructing an image recognition model based on the trained image feature extraction model. According to the method, the image recognition model capable of ensuring the recognition accuracy can be constructed based on the trained image feature extraction model under the condition that the number of marked samples is limited.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, in particular to an image recognition model construction method, an eye fundus image recognition method, an electronic device and a storage medium. BACKGROUND

[0002] Glaucoma is a chronic optic neuropathy caused by elevated intraocular pressure and insufficient blood supply to the optic nerve. If it is not discovered and treated in time, it will lead to visual dysfunction and even blindness. Currently, the diagnosis and treatment of glaucoma mainly relies on the professional knowledge and experience of doctors, which undoubtedly increases the workload of doctors in the case of tight medical resources. In order to alleviate this pressure and improve the efficiency of diagnosis, it is proposed to use deep learning technology to provide pre-treatment assistance for patients to achieve pre-screening of glaucoma. However, the current deep learning model for glaucoma image recognition generally uses supervised learning to construct, which requires a large number of labeled samples to be collected in advance, and it is difficult to collect a large number of samples with professional labels in practical applications. Especially in the field of medical health, it not only takes a lot of time for doctors to obtain a large number of samples with professional labels, but also the cost of labeled samples is relatively high.

[0003] Correspondingly, there is a need in the art for a new image recognition model construction scheme to solve the above problems. SUMMARY

[0004] In order to overcome the above defects, the present application is proposed to solve or at least partially solve the technical problem of difficulty in constructing a model due to the lack of labeled samples.

[0005] In a first aspect, an image recognition model construction method is provided, the method comprising: obtaining a training data set, the training data set comprising a plurality of region of interest image data; constructing an image feature extraction model, the image feature extraction model comprising an encoder and a generator, wherein the encoder comprises at least a self-supervised variational autoencoder, a compression excitation network and a depth separable convolutional network; unsupervised training the image feature extraction model based on the training data set to obtain a trained image feature extraction model; and constructing an image recognition model based on the trained image feature extraction model.

[0006] In one technical solution of the above image recognition model construction method, the obtaining of the training data set comprises: collecting eye fundus image data; pre-processing the eye fundus image data to obtain region of interest image data; and obtaining the training data set based on all region of interest image data.

[0007] In a technical solution of the above image recognition model construction method, the preprocessing of the fundus image data to obtain the region of interest image data comprises: obtaining a mask region of the fundus image data; determining a cup-disc region of the fundus image data based on the mask region; and cutting the fundus image data based on the cup-disc region to obtain the region of interest image data.

[0008] In a technical solution of the above image recognition model construction method, the generator comprises a first generator and a second generator; the unsupervised training of the image feature extraction model based on the training data set to obtain the trained image feature extraction model comprises: inputting the region of interest image data in the training data set into the image feature extraction model, and obtaining a feature vector conforming to a standard normal distribution by an encoder based on the region of interest image data; inputting the feature vector into the first generator to generate reconstructed image data; obtaining a random vector from the standard normal distribution and inputting the random vector into the second generator to generate pseudo-sample image data; constructing an encoder loss function and a generator loss function based on the region of interest image data, the reconstructed image data and the pseudo-sample image data; and performing unsupervised training on the image feature extraction model based on the encoder loss function and the generator loss function, determining weight parameters of the encoder loss function and weight parameters of the generator loss function, and obtaining the trained image feature extraction model.

[0009] In a technical solution of the above image recognition model construction method, the construction of the encoder loss function and the generator loss function based on the region of interest image data, the reconstructed image data and the pseudo-sample image data comprises: obtaining divergence loss of the region of interest image data, divergence loss of the pseudo-sample image data and divergence loss of the reconstructed image data respectively; determining introspection loss based on the divergence loss of the region of interest image data, the divergence loss of the pseudo-sample image data and the divergence loss of the reconstructed image data; determining reconstruction loss based on the reconstructed image data and the region of interest image data; constructing the encoder loss function based on the introspection loss and the reconstruction loss; and constructing the generator loss function based on the divergence loss of the pseudo-sample image data and the divergence loss of the reconstructed image data.

[0010] In a technical solution of the above image recognition model construction method, the construction of the image recognition model based on the trained image feature extraction model comprises: obtaining an encoder in the trained image feature extraction model as a recognition network; connecting a fully connected layer after the recognition network and setting an output dimension of the fully connected layer; and constructing the image recognition model based on the recognition network and the fully connected layer.

[0011] In one of the technical solutions of the image recognition model construction method, the method further comprises: obtaining a labeled sample data set, wherein the labeled sample data set comprises glaucoma sample data and non-glaucoma sample data; training the image recognition model based on the labeled sample data set, wherein in the training process, the weight parameters of the recognition network are fixed and unchanged, the full connection layer is trained based on the labeled sample data set, the network parameters of the full connection layer are determined, and a trained image recognition model is obtained.

[0012] In a second aspect, an eye fundus image recognition method is provided, comprising: obtaining a to-be-detected eye fundus image; preprocessing the to-be-detected eye fundus image to obtain a processed to-be-detected eye fundus image; inputting the processed to-be-detected eye fundus image into an image recognition model constructed by the method of any one of the technical solutions of the image recognition model construction method described above, and obtaining a detection result of the to-be-detected eye fundus image.

[0013] In a third aspect, an electronic device is provided, comprising at least one processor; and a memory connected in communication with the at least one processor; wherein the memory has stored therein a computer program, and the computer program, when executed by the at least one processor, implements the method of any one of the technical solutions of the image recognition model construction method described above.

[0014] In a fourth aspect, a computer-readable storage medium is provided, which has stored therein a plurality of program codes, the program codes being adapted to be loaded and run by a processor to execute the method of any one of the technical solutions of the image recognition model construction method described above.

[0015] The one or more technical solutions of the present application have at least one or more of the following beneficial effects:

[0016] The image recognition model construction method provided in the application comprises the following steps: obtaining a training data set, wherein the training data set comprises a plurality of region of interest image data; constructing an image feature extraction model, wherein the image feature extraction model comprises an encoder and a generator, and the encoder comprises at least a self-supervised variational autoencoder, a compression excitation network and a depth separable convolution network; performing unsupervised training on the image feature extraction model based on the training data set to obtain a trained image feature extraction model; and constructing an image recognition model based on the trained image feature extraction model. The image feature extraction model is trained by means of unsupervised learning based on the training data set, so that the trained image feature extraction model can be obtained, and the image recognition model can be constructed based on the trained image feature extraction model even if the number of labeled samples is limited. The labeling cost of samples required for constructing the image recognition model is significantly reduced, and the recognition accuracy of the image recognition model is ensured, thereby providing strong support for early screening and diagnosis of glaucoma. BRIEF DESCRIPTION OF DRAWINGS

[0017] The disclosure of the application will become more apparent from the following description in conjunction with the accompanying drawings. It is readily understood by those skilled in the art that the drawings are merely intended to illustrate the application and are not intended to limit the scope of protection of the application. In which:

[0018] Figure 1 is a main step flow diagram of an image recognition model construction method according to an embodiment of the application;

[0019] Figure 2 is a fundus image data schematic diagram according to an embodiment of the application;

[0020] Figure 3 is a mask region image schematic diagram according to an embodiment of the application;

[0021] Figure 4 is a region of interest image data schematic diagram according to an embodiment of the application;

[0022] Figure 5 is an encoder structure schematic diagram according to an embodiment of the application;

[0023] Figure 6 is a generator structure schematic diagram according to an embodiment of the application;

[0024] Figure 7 is a data processing flow diagram of an image feature extraction model according to an embodiment of the application;

[0025] Figure 8 is a display schematic diagram of each image data according to an embodiment of the application.

[0026] Figure 9 is a detailed step flow diagram of an image recognition model construction method according to an embodiment of the present application;

[0027] Figure 10 is a main step flow diagram of an eye fundus image recognition method according to an embodiment of the present application;

[0028] Figure 11 is a main structure diagram of an electronic device according to an embodiment of the present application.

[0029] Reference signs:

[0030] 11: memory; 12: processor. DETAILED DESCRIPTION

[0031] Some embodiments of the present application will be described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0032] In the description of the present application, "module" and "processor" can include hardware, software or a combination of both. A module can include hardware circuitry, various suitable sensors, communication ports, memories, and can also include a software part such as program code, and can be a combination of software and hardware. The processor can be a central processor, a microprocessor, an image processor, a digital signal processor or any other suitable processor. The processor has data and / or signal processing functions. The processor can be implemented in software, hardware or a combination of both. The computer readable storage medium includes any suitable medium that can store program code, such as a magnetic disk, a hard disk, an optical disk, a flash memory, a read-only memory, a random access memory, etc. The term "A and / or B" means all possible combinations of A and B, such as only A, only B or A and B. The term "at least one of A or B" or "at least one of A and B" has a similar meaning as "A and / or B", and can include only A, only B or A and B. The singular form of the term "one", "this" can also include the plural form.

[0033] Using deep learning technology to provide pre-diagnosis screening for patients with glaucoma can effectively improve the diagnosis efficiency. However, the current deep learning model for glaucoma image recognition generally uses a supervised learning method to construct, which requires a large number of labeled samples to be collected in advance, and it is difficult to collect a large number of samples with professional labels in real application. Especially in the field of medical health, to obtain a large number of samples with professional labels not only needs a large amount of time of doctors, but also the cost of labeled samples is relatively high.

[0034] To this end, this application provides a method for constructing an image recognition model, comprising: acquiring a training dataset, the training dataset including image data of multiple regions of interest; constructing an image feature extraction model, the image feature extraction model including an encoder and a generator, wherein the encoder includes at least an introspective variational autoencoder, a compressed excitation network, and a deep separable convolutional network; performing unsupervised training on the image feature extraction model based on the training dataset to obtain a trained image feature extraction model; and constructing an image recognition model based on the trained image feature extraction model. This application utilizes a training dataset to train the image feature extraction model through unsupervised learning, thereby enabling the construction of an image recognition model based on the trained image feature extraction model even with a limited number of labeled samples. This not only significantly reduces the labeling cost of the large number of labeled samples required in the construction of supervised learning models, but also ensures the recognition accuracy of the image recognition model, providing strong support for the early screening and diagnosis of glaucoma.

[0035] See appendix Figure 1 , Figure 1 This is a schematic flowchart illustrating the main steps of an image recognition model construction method according to an embodiment of this application. Figure 1 As shown, the image recognition model construction method in this application embodiment mainly includes the following steps S101 to S104.

[0036] Step S101: Obtain the training dataset, which includes image data of multiple regions of interest.

[0037] In this embodiment, the training dataset includes multiple regions of interest (ROI) image data. The ROI image data are images of the regions of interest obtained from fundus image data. All image data in the training dataset are unlabeled image data.

[0038] Step S102: Construct an image feature extraction model, which includes an encoder and a generator, wherein the encoder includes at least an introspective variational autoencoder, a compressed excitation network, and a depthwise separable convolutional network.

[0039] In this embodiment, the image feature extraction model includes an encoder and a generator. The encoder includes an introspective variational autoencoder, a compressed excitation network, and a depthwise separable convolutional network. The introspective variational autoencoder refers to the IntroVAE model (Introspective Variational Autoencoder), and the compressed excitation network refers to SeNet (Squeeze and Excitation Networks).

[0040] Step S103: Perform unsupervised training on the image feature extraction model based on the training dataset to obtain a trained image feature extraction model.

[0041] Step S104: Construct an image recognition model based on the trained image feature extraction model.

[0042] Based on the methods described in steps S101 to S104 above, this application obtains a training dataset, constructs an image feature extraction model, and trains the image feature extraction model using the training dataset through unsupervised learning to obtain a trained image feature extraction model. This enables the construction of an image recognition model based on the trained image feature extraction model even with a limited number of labeled samples. This not only significantly reduces the labeling cost of the large number of labeled samples required in the construction of supervised learning models, but also ensures the recognition accuracy of the image recognition model, providing strong support for the early screening and diagnosis of glaucoma.

[0043] The following sections will provide further explanation of steps S101-S104.

[0044] For step S101, a training dataset is obtained, which includes image data of multiple regions of interest.

[0045] In one embodiment, obtaining the training dataset includes: collecting fundus image data; preprocessing the fundus image data to obtain region of interest image data; and obtaining a training dataset based on all region of interest image data.

[0046] Specifically, a large amount of fundus image data is collected. All fundus images are unlabeled and may include normal fundus images as well as fundus images of glaucoma patients. The collected fundus image data may also contain other eye diseases. Therefore, the fundus image data needs to be preprocessed to obtain image data of the region of interest in order to avoid interference from information from other regions.

[0047] In one embodiment, the preprocessing of the fundus image data to obtain region of interest image data includes: acquiring a mask region of the fundus image data; determining the optic cup / optic disc region of the fundus image data based on the mask region; and cropping the fundus image data based on the optic cup / optic disc region to obtain the region of interest image data.

[0048] Specifically, fundus image data such as Figure 2 As shown, the image contains a lot of interfering information, such as a black background, markings on the image, and other interfering information that affects glaucoma identification. In order to improve the accuracy of identification, it is necessary to remove image information from other areas.

[0049] First, the size of the fundus image data is adjusted. Then, the segmentation model UNet is used to predict the mask region of the fundus image data. This mask region corresponds to the optic cup and optic disc in the fundus image data. UNet is a neural network structure commonly used in image segmentation tasks. It can learn and predict pixel-level masks of specific regions in an image, i.e., generate a mask region image in which the mask region is highlighted. The mask region image is as follows: Figure 3 As shown.

[0050] Then, the mask region obtained by U-Net is further refined with the Otsu thresholding algorithm to more accurately delineate the actual boundaries of the view cup and view disk. The Otsu thresholding algorithm is a global thresholding segmentation method used to divide the gray levels of an image into two parts to maximize the inter-class variance between the two parts.

[0051] Based on the refined mask region, the size of the cropping region is estimated according to the numerical values ​​at different locations in the mask region image, thus determining the optic cup and optic disc regions of the fundus image data. The fundus image data is then cropped based on these regions and a pre-trained cropping model, which is responsible for locating and extracting the region of interest (ROI) image data from the fundus image data. The size of the ROI image data can be set to 400*400 pixels, as shown in the image. Figure 4 As shown.

[0052] The above is an explanation of step S101. The following is a further explanation of step S102.

[0053] For step S102, an image feature extraction model is constructed. The image feature extraction model includes an encoder and a generator. The encoder includes at least an introspective variational autoencoder, a compressed excitation network, and a depthwise separable convolutional network.

[0054] Specifically, the image feature extraction model includes an encoder and a generator. The encoder includes at least an introspective variational autoencoder (IntroVAE), a squeeze-and-excitation network (SENet), and a depthwise separable convolutional network.

[0055] The encoder comprises a convolutional layer (Conv), batch normalization (BN), a rectified linear unit (ReLU) activation function, an average pooling layer (AvgPool), a residual network (ResNet), an average pooling layer (AvgPool), squeeze-and-excitation networks (SENet), a depthwise separable convolutional network, and a fully connected layer (FC). The residual network and the average pooling layer are iteratively configured as multiple modules. The encoder structure is as follows: Figure 5 As shown.

[0056] The generator consists of a fully connected layer (FC), a scaling network, a residual network (ResNet), and an upsampling layer. The residual network and upsampling layer are iteratively constructed modules. The generator's structure is as follows: Figure 6 As shown.

[0057] The above is an explanation of step S102. The following is a further explanation of step S103.

[0058] In step S103, the image feature extraction model is trained unsupervised based on the training dataset to obtain a trained image feature extraction model.

[0059] In one embodiment, the generator includes a first generator and a second generator; the step of unsupervised training of the image feature extraction model based on the training dataset to obtain a trained image feature extraction model includes: inputting region-of-interest (ROI) image data from the training dataset into the image feature extraction model, whereby the encoder obtains a feature vector conforming to a standard normal distribution based on the ROI image data; inputting the feature vector into the first generator to generate reconstructed image data; obtaining a random vector from the standard normal distribution and inputting it into the second generator to generate pseudo-sample image data; constructing an encoder loss function and a generator loss function based on the ROI image data, the reconstructed image data, and the pseudo-sample image data; and performing unsupervised training of the image feature extraction model based on the encoder loss function and the generator loss function to determine the weight parameters of the encoder loss function and the generator loss function, thereby obtaining a trained image feature extraction model.

[0060] See appendixFigure 7 , Figure 7 This is a schematic diagram of the data processing flow of an image feature extraction model according to an embodiment of this application.

[0061] like Figure 7 As shown, the region of interest image is input into the encoder, which processes the input region of interest image by extracting the image features through a series of convolutional layers, pooling layers, and other operations, and generates a feature vector that conforms to a standard normal distribution based on these features. This feature vector captures the main information of the region of interest image.

[0062] The feature vector generated by the encoder is input into the first generator to generate reconstructed image data. Then, a vector is randomly sampled from the normal distribution and input into the second generator, which generates pseudo-sample image data.

[0063] See appendix Figure 8 , Figure 8 This is a schematic diagram illustrating various image data according to an embodiment of this application.

[0064] like Figure 8 As shown, in this embodiment, the first four rows of image data are the original region of interest image data, the middle four rows are the reconstructed image data generated by the first generator, and the bottom four rows of image data are the pseudo-sample image data generated by the second generator based on random vectors of random sampling.

[0065] Based on the generated region-of-interest (ROI) image data, reconstructed image data, and pseudo-sample image data, the encoder loss function and generator loss function are determined. The encoder loss function typically measures the difference between the reconstructed image data and the ROI image data to improve the encoder's feature extraction capability.

[0066] The generator loss function consists of two parts: one part measures the difference between the pseudo-sample image data and the region of interest image data, and the other part measures the difference between the reconstructed image data and the region of interest image data.

[0067] An unsupervised training method is used to train an image feature extraction model based on encoder and generator loss functions. During training, the region-of-interest (ROI) images in the training dataset are all unlabeled. Through training, the image feature extraction model learns the weight parameters of the encoder and generator loss functions, resulting in a well-trained image feature extraction model.

[0068] In one implementation, constructing an encoder loss function and a generator loss function based on the region of interest (ROI) image data, the reconstructed image data, and the pseudo-sample image data includes: obtaining the divergence loss of the ROI image data, the divergence loss of the pseudo-sample image data, and the divergence loss of the reconstructed image data, respectively; determining an introspection loss based on the divergence loss of the ROI image data, the divergence loss of the pseudo-sample image data, and the divergence loss of the reconstructed image data; determining a reconstruction loss based on the reconstructed image data and the ROI image data; constructing an encoder loss function based on the introspection loss and the reconstruction loss; and constructing a generator loss function based on the divergence loss of the pseudo-sample image data and the divergence loss of the reconstructed image data.

[0069] Divergence loss refers to Kullback-Leibler divergence, a method of quantifying the difference between two probability distributions, also known as relative entropy. KL divergence measures the amount of information lost when using one distribution to approximate another.

[0070] Specifically, two encoders are connected after the first and second generators. These encoders are used to obtain the mean and variance of the pseudo-sample image data and the mean and variance of the reconstructed image data, respectively. Based on the mean and variance of the region of interest (ROI) image data, the mean and variance of the pseudo-sample image data, and the mean and variance of the reconstructed image data, the KL divergence loss of the ROI image data, the KL divergence loss of the pseudo-sample image data, and the KL divergence loss of the reconstructed image data are determined, respectively.

[0071] Based on the KL divergence loss of the region of interest image data, the KL divergence loss of the pseudo-sample image data, and the KL divergence loss of the reconstructed image data, the introspection loss is determined. The formula for calculating the introspection loss is as follows: Loss_margin =loss_real_kl+(F.relu(m-loss_rec_kl)+F.relu(m -loss_fake_kl))*0.5*weight_neg

[0072] Where Loss_margin is the introspection loss, loss_real_kl is the KL divergence loss of the region of interest image data, m is the positive boundary value, loss_rec_kl is the KL divergence loss of the reconstructed image data, loss_fake_kl is the KL divergence loss of the fake sample image data, and weight_neg is the weight.

[0073] Common methods for calculating reconstruction loss, such as Mean Squared Error (MSE), Mean Absolute Error (MAE), and feature-based loss functions, are used to determine the reconstruction loss based on the reconstructed image data and the region of interest image data. Then, the encoder loss function is determined based on the determined reconstruction loss and introspection loss. The encoder loss function is shown below: LossE=loss_rec*weight_rec+Loss_margin*weight_kl

[0074] Where loss_rec is the reconstruction loss, weight_rec is the weight of the reconstruction loss, loss_margin is the introspection loss, and weight_kl is the weight.

[0075] The generator loss function is constructed based on the divergence loss of pseudo-sample image data and the divergence loss of reconstructed image data. The generator loss function is as follows: LossG=(l0ssG_rec_kl+lossG_fake_kl)*0.5*weight_kl

[0076] Where lossG_rec_kl is the KL divergence loss of the reconstructed image data of the generator, lossG_fake_kl is the KL divergence loss of the fake sample image data of the generator, and weight_kl is the weight.

[0077] The above is an explanation of step S103. The following is a further explanation of step S104.

[0078] Regarding step S104, in one embodiment, constructing an image recognition model based on the trained image feature extraction model includes: obtaining the encoder in the trained image feature extraction model as a recognition network; connecting a fully connected layer after the recognition network, setting the output dimension of the fully connected layer, and constructing an image recognition model based on the recognition network and the fully connected layer.

[0079] See appendix Figure 9 , Figure 9 This is a detailed flowchart illustrating the steps of an image recognition model construction method according to an embodiment of this application.

[0080] like Figure 9As shown, this application first collects fundus image data and preprocesses the fundus image data to obtain a training dataset. Then, it trains an image feature extraction model based on the training dataset and builds an image recognition model based on the trained image feature extraction model. Specifically, the encoder network in the image feature extraction model is saved as a recognition network. A fully connected layer is added after the recognition network, and the output dimension of the fully connected layer is set to 2. The image recognition model is composed of the recognition network plus the fully connected layer.

[0081] In one embodiment, the method further includes: acquiring a labeled sample dataset, wherein the labeled sample dataset includes glaucoma sample data and non-glaucoma sample data; training the image recognition model based on the labeled sample dataset, wherein during the training process, the weight parameters of the recognition network are fixed, the fully connected layer is trained based on the labeled sample dataset, the network parameters of the fully connected layer are determined, and a trained image recognition model is obtained.

[0082] Specifically, the labeled sample dataset includes a small amount of labeled sample data, specifically glaucoma sample data and non-glaucoma sample data. Using transfer learning as a training method, a portion of the samples in the labeled sample dataset is extracted to train the image recognition model. The weight parameters of the recognition network are frozen, and the fully connected layers are trained with a focus. After multiple iterations, a small number of samples from the labeled sample dataset that were not used in model training are selected for testing and verification, thereby obtaining the final detection results and accuracy. Based on the detection results and accuracy, the image recognition model is evaluated. If the evaluation results meet the preset criteria, the network parameters of the fully connected layers are determined, resulting in the trained image recognition model. This model is then applied to actual glaucoma image pre-screening, and the model output is combined with doctors' experience for glaucoma pre-screening.

[0083] See appendix Figure 10 , Figure 10 This is a schematic flowchart of the main steps of a fundus image recognition method according to an embodiment of this application, as shown below. Figure 10 As shown, a fundus image recognition method of this application includes the following steps:

[0084] Step S201: Obtain the fundus image to be detected.

[0085] Step S202: Preprocess the fundus image to be detected to obtain a processed fundus image to be detected. Specifically, the preprocessing of the fundus image to be detected is the same as the preprocessing steps for the fundus image data described above.

[0086] Step S203: Input the processed fundus image to be detected into the image recognition model constructed by the method described in any one of the technical solutions of the above-described image recognition model construction method, and obtain the detection result of the fundus image to be detected.

[0087] This application utilizes a training dataset to train an image feature extraction model through unsupervised learning, resulting in a well-trained image feature extraction model. This enables the construction of an image recognition model based on the trained model, even with a limited number of labeled samples. This not only significantly reduces the labeling cost of the samples required to build the image recognition model but also ensures the recognition accuracy of the model. This provides strong support for the early screening and diagnosis of glaucoma, offers doctors more efficient auxiliary diagnosis and treatment, reduces their workload, and further improves diagnostic efficiency.

[0088] It should be noted that although the steps in the above embodiments are described in a specific order, those skilled in the art will understand that in order to achieve the effect of this application, different steps do not necessarily have to be executed in such an order. They can be executed simultaneously (in parallel) or in other orders. These adjusted solutions are equivalent to the technical solutions described in this application and therefore will also fall within the protection scope of this application.

[0089] Those skilled in the art will understand that all or part of the processes in the method of the above-described embodiment can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0090] Another aspect of this application provides an electronic device.

[0091] See appendix Figure 11 , Figure 11The diagram exemplarily illustrates a memory 11 and a processor 12 connected in communication via a bus. In an embodiment of an electronic device according to this application, the electronic device may include at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program that, when executed by the at least one processor, implements the methods described in any of the above embodiments. In some embodiments of this application, the electronic device may also include at least one sensor for sensing information. The sensor is communicatively connected to any type of processor mentioned in this application.

[0092] Another aspect of this application provides a computer-readable storage medium.

[0093] In one embodiment of a computer-readable storage medium according to this application, the computer-readable storage medium can be configured to store a program that performs the image recognition model construction method of the above-described method embodiments. This program can be loaded and run by a processor to implement the image recognition model construction method. For ease of explanation, only the parts related to the embodiments of this application are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of this application. The computer-readable storage medium can be a storage device comprising various electronic devices. Optionally, in the embodiments of this application, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0094] Furthermore, it should be understood that since the various modules are only provided to illustrate the functional units of the device described in this application, the physical devices corresponding to these modules may be the processor itself, or a part of the processor's software, hardware, or a combination of both. Therefore, the number of modules shown in the figures is merely illustrative.

[0095] Those skilled in the art will understand that the various modules in the device can be adaptively split or combined. Such splitting or combining of specific modules will not cause the technical solution to deviate from the principles of this application; therefore, the technical solutions after splitting or combining will fall within the protection scope of this application.

[0096] The technical solution of this application has been described above with reference to one embodiment shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

Claims

1. A method for constructing an image recognition model, characterized in that, The method includes: Obtain a training dataset, which includes image data of multiple regions of interest; An image feature extraction model is constructed, which includes an encoder and a generator, wherein the encoder includes at least an introspective variational autoencoder, a compressed excitation network, and a depthwise separable convolutional network; The image feature extraction model is trained unsupervised based on the training dataset to obtain a trained image feature extraction model. An image recognition model is constructed based on the trained image feature extraction model.

2. The image recognition model construction method according to claim 1, characterized in that, The acquisition of the training dataset includes: Collect fundus image data; The fundus image data is preprocessed to obtain the region of interest image data; The training dataset is obtained based on image data of all regions of interest.

3. The image recognition model construction method according to claim 2, characterized in that, The preprocessing of the fundus image data to obtain region of interest image data includes: Obtain the mask region of the fundus image data; The optic cup / optic disc region of the fundus image data is determined based on the mask region. The fundus image data is cropped based on the optic cup and optic disc region to obtain the region of interest image data.

4. The image recognition model construction method according to claim 1, characterized in that, The generator includes a first generator and a second generator; the unsupervised training of the image feature extraction model based on the training dataset to obtain a trained image feature extraction model includes: The region of interest (ROI) image data from the training dataset is input into the image feature extraction model, and the encoder obtains a feature vector that conforms to a standard normal distribution based on the ROI image data. The feature vector is input into the first generator to generate reconstructed image data; A random vector is obtained from a standard normal distribution and input into a second generator to generate pseudo-sample image data; Based on the region of interest image data, the reconstructed image data, and the pseudo-sample image data, an encoder loss function and a generator loss function are constructed. The image feature extraction model is trained unsupervised based on the encoder loss function and the generator loss function to determine the weight parameters of the encoder loss function and the generator loss function, thereby obtaining the trained image feature extraction model.

5. The image recognition model construction method according to claim 4, characterized in that, The construction of encoder loss function and generator loss function based on the region of interest image data, the reconstructed image data, and the pseudo-sample image data includes: The divergence loss of the region of interest image data, the divergence loss of the pseudo-sample image data, and the divergence loss of the reconstructed image data are obtained respectively. The introspection loss is determined based on the divergence loss of the region of interest image data, the divergence loss of the pseudo-sample image data, and the divergence loss of the reconstructed image data; The reconstruction loss is determined based on the reconstructed image data and the region of interest image data; Construct an encoder loss function based on the introspection loss and the reconstruction loss; A generator loss function is constructed based on the divergence loss of the pseudo-sample image data and the divergence loss of the reconstructed image data.

6. The image recognition model construction method according to claim 4, characterized in that, The construction of the image recognition model based on the trained image feature extraction model includes: Obtain the encoder from the trained image feature extraction model as the recognition network; A fully connected layer is connected after the recognition network, the output dimension of the fully connected layer is set, and an image recognition model is constructed based on the recognition network and the fully connected layer.

7. The image recognition model construction method according to claim 6, characterized in that, The method further includes: Obtain a labeled sample dataset, wherein the labeled sample dataset includes glaucoma sample data and non-glaucoma sample data; The image recognition model is trained based on the labeled sample dataset. During the training process, the weight parameters of the recognition network are kept constant. The fully connected layer is trained based on the labeled sample dataset to determine the network parameters of the fully connected layer, thereby obtaining the trained image recognition model.

8. A method for fundus image recognition, characterized in that, The method includes: Acquire the fundus image to be detected; The fundus image to be detected is preprocessed to obtain the processed fundus image to be detected; The processed fundus image to be detected is input into an image recognition model constructed using the image recognition model construction method as described in any one of claims 1 to 7, and the detection result of the fundus image to be detected is obtained.

9. An electronic device comprising at least one processor and at least one memory, said memory being adapted to store a plurality of program codes, characterized in that, The program code is adapted to be loaded and run by the processor to perform the image recognition model construction method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a plurality of program codes, characterized in that, The program code is adapted to be loaded and run by a processor to perform the image recognition model construction method according to any one of claims 1 to 7.

Images