Training method for glaucoma image feature extraction based on artificial neural network

Abstract

The disclosure describes a training method for glaucoma image feature extraction based on an artificial neural network, including preparing fundus images and labeled images, the labeled images including optic disc labeled images labeling optic disc regions and optic cup labeled images labeling optic cup regions; preprocessing the fundus images to obtain preprocessed fundus images, and performing blood vessel detection to obtain blood vessel images; mixing and weighting the optic disc labeled images and the blood vessel images to generate a mixed weight distribution map, the mixed weight distribution map including an intra-optic disc blood vessel region, an intra-optic disc non-blood vessel region, an extra-optic disc blood vessel region and an extra-optic disc non-blood vessel region, the weight of the optic disc region being greater than the weight of the non-optic disc region, and the weight of the blood vessel region being greater than the weight of the non-blood vessel region; and training the artificial neural network based on the preprocessed fundus images and the labeled images using the mixed weight distribution map. Thus, the accuracy of the artificial neural network in recognizing glaucoma image features can be improved.

Description

[0001] This application is a divisional application of the patent application filed on July 18, 2020, with application number 2020107026436, entitled "Training Method and Training System for Glaucoma Image Feature Extraction". Technical Field

[0002] This disclosure relates to the field of biomedical engineering, and in particular to a training method for glaucoma image feature extraction based on artificial neural networks. Background Technology

[0003] Glaucoma is currently the second leading cause of blindness worldwide. There are over ten million people with primary glaucoma globally, and more than 10% of them may develop bilateral blindness. If glaucoma is not treated early, it can lead to irreversible blindness; therefore, early glaucoma screening is of paramount importance.

[0004] The main pathological process of glaucoma is the death of retinal ganglion cells and loss of axons, leading to the loss of nerve fibers at the optic disc margin, which in turn causes morphological changes in the optic disc, such as widening and deepening of optic disc cupping. Clinical medical research shows that the cup-to-disc ratio (CDR) in fundus images is a reliable indicator of optic disc cupping; therefore, glaucoma can be identified through the CDR in fundus images. In clinical medicine, existing identification methods include using artificial intelligence technology to process features in fundus images to identify the optic disc or optic cup, thereby identifying fundus lesions.

[0005] However, since there are generally only 4-6 pairs of first- or second-order retinal arteries and veins in the optic disc region, when using artificial intelligence technology to identify fundus images, this arterial and venous information is easily overlooked, resulting in the inability to accurately learn the course of small blood vessels, thus leading to the inability to accurately identify the optic disc or optic cup. Summary of the Invention

[0006] This disclosure is made in view of the above-mentioned state of the prior art, and its purpose is to provide a training method and training system for glaucoma image feature extraction based on artificial neural networks that can improve the accuracy of glaucoma image feature extraction by artificial neural networks.

[0007] To this end, the first aspect of this disclosure provides a training method for glaucoma image feature extraction based on an artificial neural network, comprising: preparing a fundus image and a labeled image, wherein the labeled image includes an optic disc labeled image of the optic disc region and an optic cup labeled image of the optic cup region; preprocessing the fundus image to obtain a preprocessed fundus image, and generating a vascular image containing vascular regions based on vascular detection results; mixing and weighting the optic disc labeled image and the vascular image to generate a mixed weight distribution map; and training the artificial neural network based on the preprocessed fundus image, the labeled image, and the mixed weight distribution map, wherein, during the mixed weighting, the weight of the optic disc region is greater than the weight of the non-optic disc region, and the weight of the vascular region is greater than the weight of the non-vascular region. In this disclosure, training the artificial neural network based on the preprocessed fundus image, the labeled image, and the mixed weight distribution map enables the artificial neural network to take into account both vascular regions and optic disc regions during training, optimize the learning of small blood vessel course, and suppress positive and negative sample imbalance, thereby improving the accuracy of the artificial neural network in extracting glaucoma image features.

[0008] Furthermore, in the training method for glaucoma image feature extraction based on artificial neural networks according to the first aspect of this disclosure, optionally, the optic disc region in the optic disc annotation image is dilated to form an optic disc dilated image; and the blood vessel region in the blood vessel image is dilated to form a blood vessel dilated image. In this case, an optic disc dilated image including the region near the optic disc and a blood vessel dilated image including the blood vessel boundaries are obtained through dilation processing.

[0009] Furthermore, in the training method for glaucoma image feature extraction based on artificial neural networks according to the first aspect of this disclosure, optionally, the optic disc edema image and the vascular edema image are mixed and weighted to generate the mixed weight distribution map. This can further improve the accuracy of glaucoma image feature extraction by the artificial neural network.

[0010] Furthermore, in the training method for glaucoma image feature extraction based on an artificial neural network according to the first aspect of this disclosure, optionally, during the training process, the coefficients of the loss function in the artificial neural network are obtained based on the mixed weight distribution map, and the artificial neural network is trained based on the loss function. In this case, optimizing the artificial neural network based on the loss function coefficients obtained from the mixed weight distribution map can suppress the imbalance between positive and negative samples, thereby further improving the accuracy of the artificial neural network in extracting glaucoma image features.

[0011] Furthermore, in the training method for glaucoma image feature extraction based on artificial neural networks according to the first aspect of this disclosure, optionally, the mixed weight distribution map includes a fundus region and a background region, and the weight of the background region is set to zero. This reduces the interference of the background region on the artificial neural network's recognition of glaucoma image features.

[0012] Furthermore, in the training method for glaucoma image feature extraction based on artificial neural networks according to the first aspect of this disclosure, the preprocessing may optionally include cropping and normalizing the fundus image. In this case, the cropping process can convert the fundus image into an image of a fixed standard form, and the normalization process can overcome the differences between different fundus images, thereby enabling the artificial neural network to extract glaucoma image features more conveniently.

[0013] Furthermore, in the training method for glaucoma image feature extraction based on artificial neural networks according to the first aspect of this disclosure, optionally, the mixed weight distribution map includes intradiscal vascular regions, intradiscal non-vascular regions, extradiscal vascular regions, and extradiscal non-vascular regions. This enables the artificial neural network to more accurately extract glaucoma image features from each region of the fundus image.

[0014] Furthermore, in the training method for glaucoma image feature extraction based on artificial neural networks according to the first aspect of this disclosure, optionally, the weight of the optic disc region is set as a first weight, the weight of the non-optic disc region as a second weight, the weight of the vascular region as a third weight, and the weight of the non-vascular region as a fourth weight. Then, the weight of the intra-optic disc vascular region is the first weight multiplied by the third weight, the weight of the intra-optic disc non-vascular region is the first weight multiplied by the fourth weight, the weight of the extra-optic disc vascular region is the second weight multiplied by the third weight, and the weight of the extra-optic disc non-vascular region is the second weight multiplied by the fourth weight. Thus, the weight values ​​of the intra-optic disc vascular region, the intra-optic disc non-vascular region, the extra-optic disc vascular region, and the extra-optic disc non-vascular region can be obtained according to the weights of the optic disc region, the non-optic disc region, the vascular region, and the non-vascular region, respectively.

[0015] Furthermore, in the training method for glaucoma image features based on artificial neural networks according to the first aspect of this disclosure, optionally, vascular region detection is performed based on Frangi filtering to form the vascular image. This allows for automatic identification of vascular regions, facilitating subsequent identification and processing of vascular regions by the artificial neural network.

[0016] The second aspect of this disclosure provides a training system for glaucoma image feature extraction based on an artificial neural network, comprising: an acquisition module for acquiring a fundus image and a labeled image, the labeled image including an optic disc labeled image of the optic disc region and an optic cup labeled image of the optic cup region; an image preprocessing module for preprocessing the fundus image to obtain a preprocessed fundus image; a blood vessel region detection module for detecting blood vessels in the preprocessed fundus image to form a blood vessel image; a mixed weight generation module for mixing and weighting the optic disc labeled image and the blood vessel image to generate a mixed weight distribution map; and a model training module for training the artificial neural network based on the preprocessed fundus image, the labeled image, and the mixed weight distribution map, wherein, during the mixed weighting, the weight of the optic disc region is greater than the weight of the non-optic disc region, and the weight of the blood vessel region is greater than the weight of the non-blood vessel region. In this disclosure, the artificial neural network is trained based on preprocessed fundus images, labeled images, and mixed weight distribution maps. This enables the artificial neural network to take into account both the vascular region and the optic disc region during training, optimize the learning of small blood vessel course, and suppress the imbalance between positive and negative samples. As a result, the accuracy of the artificial neural network in extracting features from glaucoma images can be improved.

[0017] Furthermore, in the training system for glaucoma image feature extraction based on artificial neural networks according to the second aspect of this disclosure, optionally, the optic disc region in the optic disc annotation image is dilated to form an optic disc dilated image; and the blood vessel region in the blood vessel image is dilated to form a blood vessel dilated image. In this case, an optic disc dilated image including the region near the optic disc and a blood vessel dilated image including the blood vessel boundaries are obtained through dilation processing.

[0018] Furthermore, in the training system for glaucoma image feature extraction based on artificial neural networks according to the second aspect of this disclosure, optionally, the optic disc edema image and the vascular edema image are mixed and weighted to generate the mixed weight distribution map. This further improves the accuracy of glaucoma image feature extraction by the artificial neural network.

[0019] Furthermore, in the training system for glaucoma image feature extraction based on an artificial neural network according to the second aspect of this disclosure, optionally, during the training process, the coefficients of the loss function in the artificial neural network are obtained based on the mixed weight distribution map, and the artificial neural network is trained based on the loss function. In this case, optimizing the artificial neural network based on the loss function coefficients obtained from the mixed weight distribution map can suppress the imbalance between positive and negative samples, thereby further improving the accuracy of the artificial neural network in extracting glaucoma image features.

[0020] Furthermore, in the training system for glaucoma image feature extraction based on artificial neural networks according to the second aspect of this disclosure, optionally, the mixed weight distribution map includes a fundus region and a background region, with the weight of the background region set to zero. This reduces the interference of the background region on the artificial neural network's recognition of glaucoma image features.

[0021] Furthermore, in the training system for glaucoma image feature extraction based on artificial neural networks according to the second aspect of this disclosure, optionally, vascular region detection is performed based on Frangi filtering to form the vascular image. This allows for automatic identification of vascular regions, facilitating subsequent identification and processing of vascular regions by the artificial neural network.

[0022] Furthermore, in the training system for glaucoma image feature extraction based on artificial neural networks according to the second aspect of this disclosure, optionally, the hybrid weight distribution map includes intradiscal vascular regions, intradiscal non-vascular regions, extradiscal vascular regions, and extradiscal non-vascular regions. This enables the artificial neural network to more accurately extract glaucoma image features from each region of the fundus image.

[0023] Furthermore, in the training system for glaucoma image feature extraction based on artificial neural networks according to the second aspect of this disclosure, optionally, the weight of the optic disc region is set as a first weight, the weight of the non-optic disc region as a second weight, the weight of the vascular region as a third weight, and the weight of the non-vascular region as a fourth weight. Then, the weight of the intra-optic disc vascular region is the first weight multiplied by the third weight, the weight of the intra-optic disc non-vascular region is the first weight multiplied by the fourth weight, the weight of the extra-optic disc vascular region is the second weight multiplied by the third weight, and the weight of the extra-optic disc non-vascular region is the second weight multiplied by the fourth weight. Thus, the weight values ​​of the intra-optic disc vascular region, the intra-optic disc non-vascular region, the extra-optic disc vascular region, and the extra-optic disc non-vascular region can be obtained according to the weights of the optic disc region, the non-optic disc region, the vascular region, and the non-vascular region, respectively.

[0024] According to this disclosure, a training method and system for glaucoma image feature extraction based on artificial neural networks are provided, which can improve the accuracy of glaucoma image feature extraction by artificial neural networks. Attached Figure Description

[0025] Embodiments of this disclosure will now be explained in further detail by way of example only with reference to the accompanying drawings, in which:

[0026] Figure 1 A schematic diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0027] Figure 2A block diagram of a training system for glaucoma image feature extraction based on an artificial neural network, according to an embodiment of the present disclosure, is shown.

[0028] Figure 3 A schematic diagram is shown illustrating the process of annotating fundus images to form an annotated image according to an embodiment of the present disclosure.

[0029] Figure 4 A schematic diagram of a vascular image formed by vascular region detection of a preprocessed fundus image according to an embodiment of the present disclosure is shown.

[0030] Figure 5 A schematic diagram illustrating the formation of a first hybrid weight distribution map according to an embodiment of this disclosure is shown.

[0031] Figure 6 A block diagram of the hybrid weight generation module of the training system according to an embodiment of the present disclosure is shown.

[0032] Figure 7 A schematic diagram of a diaphragm image according to an embodiment of the present disclosure is shown.

[0033] Figure 8 A schematic diagram of a blood vessel dilation image according to an embodiment of the present disclosure is shown.

[0034] Figure 9 A schematic diagram illustrating the formation of a second hybrid weight distribution map according to an embodiment of this disclosure is shown.

[0035] Figure 10 A flowchart is shown of a training method for glaucoma image feature extraction based on an artificial neural network, according to an embodiment of this disclosure.

[0036] Symbol explanation:

[0037] 1…Electronic device, 1a…Host, 1b…Display device, 1c…Input device, 10…Training system, 100…Acquisition module, 200…Image preprocessing module, 300…Vascular region detection module, 400…Hybrid weight generation module, 410…Optic disc edema module, 420…Vascular edema module, 500…Model training module, P1…Fundus image, P20…Annotated optic disc image, P30…Annotated optic cup image, P10…Preprocessed fundus image, P40…Vascular image, P21…Weighted annotated optic disc image, P41…Weighted vascular image, P3…First hybrid weight distribution map, P23…Optic disc edema image, P43…Vascular edema image P4…Second mixed weight distribution map, A1…optic disc region, A1'…non-optic disc region, A2…optic cup region, A3…optic disc bulging region, A3'…non-optic disc bulging region, A4…vascular region, A4'…non-vascular region, A5…vascular bulging region, A5'…non-vascular bulging region, A30…vascular region within the first optic disc, A30'…non-vascular region within the first optic disc, A31…vascular region outside the first optic disc, A31'…non-vascular region outside the first optic disc, A40…vascular region within the second optic disc, A40'…non-vascular region within the second optic disc, A41…vascular region outside the second optic disc, A41'…non-vascular region outside the second optic disc. Detailed Implementation

[0038] Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the following description, the same reference numerals are used for the same components, and repeated descriptions are omitted. Furthermore, the drawings are merely schematic diagrams, and the proportions of the components or the shapes of the components may differ from the actual figures.

[0039] Figure 1 A schematic diagram of an electronic device according to an embodiment of the present disclosure is shown. Figure 2 A block diagram of a training system for glaucoma image feature extraction based on an artificial neural network, according to an embodiment of the present disclosure, is shown.

[0040] In some examples, refer to Figure 1 and Figure 2 The training system (which may be simply referred to as the "training system") 10 for glaucoma image feature extraction based on artificial neural networks disclosed herein can be implemented using an electronic device 1 (such as a computer). In some examples, such as Figure 1 As shown, electronic device 1 may include host 1a, display device 1b, and input device 1c (e.g., mouse, keyboard). Host 1a may include one or more processors, memory, and a computer program stored in the memory. In this case, training system 10 may be stored in memory as a computer program.

[0041] In some examples, glaucoma image features can be glaucoma-related characteristics. Since the main pathological process of glaucoma is the death of retinal ganglion cells and loss of axons, leading to the loss of nerve fibers at the optic disc margin, this causes morphological changes in the optic disc, such as widening or deepening of optic disc cupping. Clinical medical research shows that the cup-to-disc ratio (CDR) in fundus images is a reliable indicator of optic disc cupping. Therefore, glaucoma image features can include the optic cup and optic disc. In other examples, glaucoma image features can be the cup-to-disc ratio (CDR).

[0042] In some examples, one or more processors may include a central processing unit, an image processing unit, and any other electronic components capable of processing input data. For example, a processor may execute instructions and programs stored in memory.

[0043] As described above, the training system 10 can be implemented using program instructions and algorithms encoded in a computer program. Alternatively, the training system 10 can also be stored in the memory of a cloud server. In some examples, the cloud server may be leased, thereby reducing server maintenance costs. In other examples, the cloud server may be self-built. In this case, the memory can be located on a self-built server, ensuring data confidentiality and preventing data leakage from clients or patients.

[0044] In some examples, the training system 10 may use one or more artificial neural networks to extract and learn glaucoma image features from fundus images. In some examples, the artificial neural network may be implemented by one or more processors (e.g., microprocessors, integrated circuits, field-programmable gate arrays, etc.). The training system 10 may be used to receive multiple fundus images as input and train the system on these images. The parameter values ​​of the artificial neural network are determined iteratively through the training dataset. The training dataset may consist of multiple fundus images. In some examples, during the training of the artificial neural network, preprocessed fundus images and labeled images are used as input, and the output of the artificial neural network is continuously optimized by manually setting the weights of the loss function for each region (i.e., each pixel) of the preprocessed fundus image.

[0045] In this embodiment, the training system 10 may include an acquisition module 100, an image preprocessing module 200, a blood vessel region detection module 300, a hybrid weight generation module 400, and a model training module 500 (see [link to documentation]). Figure 2 ).

[0046] In some examples, the acquisition module 100 can be used to acquire fundus images and labeled images. The image preprocessing module 200 can be used to preprocess the fundus images to obtain preprocessed fundus images. The blood vessel region detection module 300 can be used to detect blood vessel regions in the preprocessed fundus images to form blood vessel images. The mixed weight generation module 400 can be used to mix and weight the optic disc labeled images and blood vessel images to generate a mixed weight distribution map. The model training module 500 can train the artificial neural network based on the preprocessed fundus images, labeled images, and the mixed weight distribution map. In the above example, the training system 10 can obtain the mixed weight distribution map based on the optic disc labeled images and blood vessel images. In this case, the training system 10 can train the artificial neural network based on the preprocessed fundus images, labeled images, and the mixed weight distribution map, enabling the artificial neural network to take into account both blood vessel regions and optic disc regions during training, optimize the learning of small blood vessel course, and suppress positive and negative sample imbalance, thereby improving the accuracy of the artificial neural network in extracting features from glaucoma images.

[0047] Figure 3 A schematic diagram illustrating the annotation of fundus images to form an annotated image according to an embodiment of the present disclosure is shown, wherein, Figure 3 (a) shows fundus image P1. Figure 3 (b) shows the labeled image P20 on the visual disc. Figure 3 (c) shows the labeled image P30 of the visual cup.

[0048] In some examples, as described above, training system 10 may include acquisition module 100 (see...) Figure 2 The acquisition module 100 can be used to acquire fundus images and labeled images. The labeled images can be obtained by annotating the fundus images.

[0049] In some examples, the acquisition module 100 can be used to acquire fundus images. Fundus images can be images of the fundus captured by a fundus camera or other fundus imaging device. As an example of a fundus image, for example... Figure 3 (a) shows a fundus image P1 taken by a fundus camera. In some examples, the fundus image may include the area of ​​the optic disc and optic cup, but this embodiment is not limited to this. In some examples, the fundus image may only include the optic disc area.

[0050] In some examples, multiple fundus images can form a training dataset. The training dataset can include a training set and a test set. For example, 50,000 to 200,000 fundus images from collaborating hospitals, with patient information removed, could be selected as the training set, and 5,000 to 20,000 fundus images as the test set.

[0051] In some examples, the fundus image can be a color image. A color fundus image can clearly present rich information about the optic disc, optic cup, macula, blood vessels, etc. Additionally, the fundus image can be in RGB mode, CMYK mode, Lab mode, or grayscale mode.

[0052] In some examples, the acquisition module 100 can be used to acquire labeled images. The labeled images may include optic disc labeled images and optic cup labeled images. Medically, the optic disc and optic cup have clear anatomical definitions: the optic disc is defined as the edge of the posterior scleral foramen, bounded by the inner edge of the scleral ring; the optic cup is defined as the region between the lamina cribrosa and the retinal plane, with the course of small blood vessels being an important basis for identifying the optic cup region.

[0053] In other examples, the labeled images obtained by the acquisition module 100 may include only the view disk labeled images.

[0054] In some examples, the optic disc labeled image or the optic cup labeled image can be used as the ground truth for training the artificial neural network. In other examples, the optic disc labeled image and the optic cup labeled image can be combined into a single labeled image and used as the ground truth for training the artificial neural network.

[0055] In some examples, as described above, the labeled images can be images obtained by annotating fundus images. In this case, the optic disc labeled image can be an image obtained by annotating the optic disc in the fundus image. The optic cup labeled image can be an image obtained by annotating the optic cup in the fundus image.

[0056] Specifically, the optic disc region in fundus image P1 can be manually annotated to obtain optic disc annotated image P20 (see Figure 3 (b)). In some examples, manual annotation can be performed by experienced physicians, thereby improving the accuracy of optic cup region annotation. The optic cup region in fundus image P1 can be manually annotated to obtain optic cup annotated image P30 (see [link]). Figure 3 (c)). In some examples, manual annotation can be performed by experienced physicians, thereby improving the accuracy of optic cup region annotation.

[0057] In this embodiment, the image preprocessing module 200 can be used to preprocess the fundus image to obtain a preprocessed fundus image. Specifically, the image preprocessing module 200 can acquire the fundus image output by the acquisition module 100 and preprocess the fundus image to obtain a preprocessed fundus image.

[0058] In some examples, the image preprocessing module 200 can crop the fundus image. Generally, because the fundus images acquired by the acquisition module 100 may have different image formats and sizes, it is necessary to crop the fundus images to convert them into a fixed standard format. A fixed standard format can refer to images with the same format and consistent size. For example, in some examples, the preprocessed fundus images can be uniformly sized to 512×512 or 1024×1024 pixels.

[0059] In some examples, the image preprocessing module 200 can perform normalization processing on the fundus image. In some examples, normalization processing may include operations such as coordinate centering and scaling normalization of the fundus image. This overcomes the differences between various fundus images and improves the performance of the artificial neural network.

[0060] In some examples, the image preprocessing module 200 may include noise reduction and grayscale processing of the fundus image. This can highlight the features of the glaucoma image.

[0061] Additionally, in some examples, the image preprocessing module 200 may include scaling, flipping, and translating the fundus image. In this case, the amount of data for training the artificial neural network can be increased, thereby improving the generalization ability of the artificial neural network.

[0062] In some examples, fundus images can be used directly for training artificial neural networks without image preprocessing.

[0063] In other examples, the fundus image can be preprocessed first, and then the preprocessed fundus image can be annotated.

[0064] In some examples, the image preprocessing module 200 can acquire the labeled image output by the acquisition module 100. Preprocessing of the labeled image can be performed simultaneously with preprocessing of the fundus image. This ensures that the size of the labeled image and the preprocessed fundus image remain consistent, which is more beneficial for training artificial neural networks.

[0065] Figure 4 A schematic diagram is shown of a vascular image formed by vascular region detection of a preprocessed fundus image according to an embodiment of the present disclosure, wherein... Figure 4 (a) shows the preprocessed fundus image P10. Figure 4 (b) shows the vascular image P40.

[0066] In this embodiment, as described above, the training system 10 may include a blood vessel region detection module 300. The blood vessel region detection module 300 can be used to detect blood vessel regions in preprocessed fundus images to form blood vessel images.

[0067] As an example of blood vessel image formation, such as Figure 4 As shown, vascular region detection is performed on the preprocessed fundus image P10 to form a vascular image P40 containing the vascular region A4.

[0068] In some examples, vascular region detection can be performed on the preprocessed fundus image P10 based on Frangi (multi-scale linear) filtering to form a vascular image P40. Specifically, Frangi filtering is an edge detection enhancement filtering algorithm based on the Hessian matrix.

[0069] In Frangi filtering, firstly, the preprocessed fundus image P10 is converted into a grayscale image. Gaussian filtering is then used to denoise the preprocessed fundus image P10. Next, the Hessian matrix is ​​calculated. The Hessian matrix is ​​the square matrix of the second-order partial derivatives of a scalar function, which describes the local curvature of a multivariable function. The basic form of the Hessian matrix is ​​shown in equation (1) below:

[0070]

[0071] Wherein, the second-order partial differential in the x-direction is:

[0072]

[0073] Second-order partial differential in the y-direction:

[0074]

[0075] Mixed partial differentials in the x and y directions:

[0076]

[0077] Among them, f xy =f yx H is a real symmetric matrix, and enhancement filters can be constructed using two eigenvalues ​​λ1 and λ2. In the two-dimensional preprocessed fundus image, the eigenvalues ​​λ1 and λ2 can be calculated using the following formula:

[0078]

[0079]

[0080] Since second-order partial derivatives are sensitive to noise, Gaussian smoothing is performed before calculating the Hessian matrix. The response function of the p-vessel region at pixel P10 in the preprocessed fundus image is V(σ,p):

[0081]

[0082] Where σ is the scaling factor, which is the standard deviation of Gaussian smoothing when calculating the Hessian matrix. β can be set to 0.5 to distinguish between line and blocky objects. It is a parameter used to control the overall smoothness of linear objects. R B s is defined by the eigenvalues ​​λ1 and λ2.

[0083]

[0084] The filter output is maximized when the scale factor σ is closest to the actual width of the blood vessel. Therefore, the maximum response of each pixel p in the preprocessed fundus image P10 under different scale factors can be taken as the final blood vessel response. In this case, the scale factor σ is closest to the actual width of the blood vessel, and the final blood vessel response is shown in equation (9):

[0085]

[0086] Where, σ min The minimum value of the scaling factor σ, σ max This represents the maximum value of the scaling factor σ.

[0087] Finally, by setting a threshold T, the location where the vascular response is greater than T is identified as the detected vascular region A4 (see [link]). Figure 4 (b)).

[0088] In this disclosure, the Frangi algorithm is used to automatically detect vascular regions in preprocessed fundus images to obtain vascular images containing vascular regions. In this case, the vascular regions in the vascular image are more prominent compared to those in the preprocessed fundus image. This facilitates subsequent identification and processing of vascular regions.

[0089] However, the examples disclosed herein are not limited to these. In other examples, the detection of blood vessel regions in preprocessed fundus images can be achieved using matched filtering algorithms, adaptive contrast enhancement algorithms, two-dimensional Gabor filtering algorithms, or other types of artificial neural networks. Therefore, it is possible to select an appropriate algorithm or artificial neural network to achieve the detection of blood vessel regions according to different needs.

[0090] Figure 5 A schematic diagram illustrating the formation of a first hybrid weight distribution map according to an embodiment of this disclosure is shown. Wherein, Figure 5(a) shows a schematic diagram of the weighted visual disc annotation image P21. Figure 5 (b) shows a schematic diagram of the weighted vascular image P41. Figure 5 (c) shows a schematic diagram of generating the first mixed weight distribution map P3 based on the weighted optic disc labeled image P21 and the weighted blood vessel image P41.

[0091] In this embodiment, as described above, the training system 10 may include a mixed weight generation module 400. The mixed weight generation module 400 can be used to mix and weight the optic disc annotation image and the blood vessel image to generate a mixed weight distribution map.

[0092] In some examples, such as Figure 5 As shown, the mixed weight distribution map can be a first mixed weight distribution map P3. The first mixed weight distribution map P3 can be based on the visual disc labeled image P20 (see...). Figure 3 (b) and vascular images P40 (see Figure 4 (b) Obtaining the weighted image. For example, a weighted optic disc annotation image P21 can be generated by weighting the optic disc annotation image P20, and a weighted vascular image P41 can be generated by weighting the vascular image P40. A first mixed weight distribution map P3 is generated based on the weighted optic disc annotation image P21 and the weighted vascular image P41. In this case, using the mixed weight distribution map to train the artificial neural network enables the artificial neural network to take into account both the vascular region and the optic disc region during training, optimize the learning of small blood vessel course, and suppress the imbalance between positive and negative samples.

[0093] In some examples, the mixed weighting process involves assigning a first weight to the optic disc region, a second weight to the non-optic disc region, a third weight to the vascular region, and a fourth weight to the non-vascular region. Then, the weight of the intra-optic disc vascular region is the first weight multiplied by the third weight; the weight of the intra-optic disc non-vascular region is the first weight multiplied by the fourth weight; the weight of the extra-optic disc vascular region is the second weight multiplied by the third weight; and the weight of the extra-optic disc non-vascular region is the second weight multiplied by the fourth weight. The mixed weighting process is described in detail below with reference to the accompanying figures.

[0094] In some examples, such as Figure 5 As shown, in the mixed weighted processing, the weight of the viewing area A1 is set to the first weight w1, and the weight of the non-viewing area A1' is set to the second weight w2 (see [reference]). Figure 5 (a)). The weight of the vascular region A4 is the third weight v1, and the weight of the non-vascular region A4' is the fourth weight v2 (see [link]). Figure 5 (b)).

[0095] In some examples, when performing mixed weighting, the weight of the optic disc region can be greater than that of the non-optic disc region, and the weight of the vascular region can be greater than that of the non-vascular region. For example, as... Figure 5 As shown, when performing mixed weighting, the weight of the optic disc region A1 can be greater than the weight of the non-optic disc region A1', and the weight of the vascular region A4 can be greater than the weight of the non-vascular region A4'. That is, w1>w2 and v1>v2 can be set. In this disclosure, the optic disc region can also be referred to as the intra-optic disc region, and the non-optic disc region can also be referred to as the extra-optic disc region.

[0096] In some examples, the first mixed weight distribution map P3 obtained based on the optic disc labeled image P20 and the blood vessel image P40 includes four mixed regions: the blood vessel region A30 within the first optic disc, the non-vascular region A30' within the first optic disc, the blood vessel region A31 outside the first optic disc, and the non-vascular region A31' outside the first optic disc (see [link to relevant documentation]). Figure 5 (c)). In this case, for example, the weight of the vascular region A30 within the first optic disc can be obtained by multiplying the weight of region A1 within the optic disc and the weight of the vascular region A4. This allows for more accurate identification of glaucoma image features in different regions of the fundus image.

[0097] In some examples, as described above, when performing mixed weighting, the weight of the optic disc region A1 is set to the first weight w1, the weight of the non-optic disc region A1' to the second weight w2, the weight of the vascular region A4 to the third weight v1, and the weight of the non-vascular region A4' to the fourth weight v2. Then, the weight of the vascular region A30 within the first optic disc is the first weight multiplied by the third weight, i.e., w1v1. The weight of the non-vascular region A30' within the first optic disc is the first weight multiplied by the fourth weight, i.e., w1v2. The weight of the vascular region A31 outside the first optic disc is the second weight multiplied by the third weight, i.e., w2v1. The weight of the non-vascular region A31' outside the first optic disc is the second weight multiplied by the fourth weight, i.e., w2v2 (see [link to relevant documentation]). Figure 5 (c)). Thus, the weight values ​​of the intra-optic disc vascular region, intra-optic disc non-vascular region, extra-optic disc vascular region, and extra-optic disc non-vascular region can be obtained based on the weights of the optic disc region, non-optic disc region, vascular region, and non-vascular region, respectively.

[0098] In some examples, during the mixed weighting process, a mixed weight distribution map can be defined based on the eyeball contour. In this case, the mixed weight distribution map can include a fundus region and a background region. The fundus region can be the area within the eyeball contour. The fundus region can include four mixed regions: a vascular region within the first optic disc, a non-vascular region within the first optic disc, a vascular region outside the first optic disc, and a non-vascular region outside the first optic disc. The background region can be the area outside the eyeball contour. The background region can be a portion of the non-vascular region outside the first optic disc. When training the artificial neural network, the weight of the background region can be set to zero. This reduces the interference of the background region on the artificial neural network's extraction of glaucoma image features during training.

[0099] In some examples, as described above, the viewing disc region of the mixed weight distribution map has weights, and since the viewing disc region includes the viewing cup region, the viewing cup region of the mixed weight distribution map also has weights.

[0100] In other examples, the hybrid weight generation module 400 may perform weighting based solely on the viewport labeled image.

[0101] In other examples, the hybrid weight generation module 400 can perform weighting based solely on the blood vessel image.

[0102] Figure 6 A block diagram of the hybrid weight generation module of the training system according to an embodiment of the present disclosure is shown. Figure 7 A schematic diagram of a optic disc dilation image according to an embodiment of the present disclosure is shown, wherein, Figure 7 (a) shows the labeled image P20 on the visual disc. Figure 7 (b) shows the optic disc dilation image P23.

[0103] In some examples, the hybrid weight generation module 400 may include a display screen inflation module 410 (see [link to example]). Figure 6 ).

[0104] In some examples, in the optic disc dilation module 410, the optic disc region in the optic disc annotation image can be dilated to form an optic disc dilation image. The optic disc dilation image includes the optic disc dilation region. The optic disc dilation region may include the optic disc region and the region surrounding the optic disc.

[0105] For example, such as Figure 7 (a) and Figure 7 As shown in (b), the visual disk region in the visual disk annotation image P20 is dilated to form a dilated visual disk image P23. The dilated visual disk image P23 includes a dilated visual disk region A3. The dilated visual disk region A3 of the dilated visual disk image P23 corresponds to the visual disk region A1 of the visual disk annotation image P20. Wherein, Figure 7In (b), A3' represents the non-optic disc dilation region. The optic disc dilation region can include the optic disc region and the region near the optic disc. In this case, since the region near the optic disc affects the segmentation of the optic cup or optic disc, and thus affects the extraction of glaucoma image features, dilation processing is used to obtain an optic disc dilation image so that subsequent processing can be performed based on the optic disc dilation image, thereby improving the accuracy of glaucoma image feature extraction.

[0106] Figure 8 A schematic diagram of a vascular dilation image according to an embodiment of the present disclosure is shown, wherein, Figure 8 (a) shows the vascular image P40. Figure 8 (b) shows an image of vascular dilation on page 43.

[0107] In some examples, the hybrid weight generation module 400 may include a vascular expansion module 420 (see [link to example]). Figure 6 ).

[0108] In some examples, within the vascular expansion module 420, the vascular region in a vascular image can be expanded to form a vascular expansion image. The vascular expansion image includes the expanded vascular region. The expanded vascular region may include the vascular region and the region surrounding the vascular area. For example, as... Figure 8 (a) and Figure 8 As shown in (b), the blood vessel region A4 in the blood vessel image P40 can be dilated to form a dilated blood vessel image P43. The dilated blood vessel image P43 includes a dilated blood vessel region A5. The dilated blood vessel region A5 in the dilated blood vessel image P43 corresponds to the blood vessel region A4 in the blood vessel image P40. Wherein, Figure 8 In (b), A5' represents the non-vascular dilatation region. In this case, dilatation processing can reduce the error in detecting vessel boundaries based on vessel detection algorithms.

[0109] Figure 9 A schematic diagram illustrating the formation of a second hybrid weight distribution map according to an embodiment of this disclosure is shown, wherein, Figure 9 (a) shows a schematic diagram of weighting the optic disc dilation image P23. Figure 9 (b) shows a schematic diagram of weighting the vascular dilation image P43. Figure 9 (c) shows a schematic diagram of generating a second mixed weighted distribution map P4 based on the weighted optic disc dilatation image P23 and the weighted vascular dilatation image P43.

[0110] In other examples, the hybrid weight generation module 400 can be used to hybrid weight the optic disc dilatation image and the vascular dilatation image to generate a hybrid weight distribution map. That is, the hybrid weight distribution map can be generated by hybrid weighting the optic disc dilatation image and the vascular dilatation image. In this case, training the artificial neural network based on the hybrid weight distribution map generated from the dilatated optic disc image and the dilatated vascular image can reduce the error of the blood vessel detection algorithm at the blood vessel boundary.

[0111] For example, in some examples, such as Figure 9 As shown, the mixed weight distribution map can be the second mixed weight distribution map P4. Specifically, the optic disc dilatation image P23 and the vascular dilatation image P43 can be weighted separately (see...). Figure 9 (a) and Figure 9 (b) The weighted optic disc dilatation image P23 and the weighted vascular dilatation image P43 are mixed and weighted to generate a second mixed weight distribution map P4.

[0112] In some examples, such as Figure 9 As shown, based on the four basic regions of the expanded optic disc region A3, the non-optic disc region A3', the vascular region A5, and the non-vascular region A5', the second mixed weight distribution map P4 can include the vascular region A40 within the second optic disc, the non-vascular region A40' within the second optic disc, the vascular region A41 outside the second optic disc, and the non-vascular region A41' outside the second optic disc.

[0113] In some examples, the weight of the optic disc bulge region A3 can be set to the first weight w1', the weight of the non-optic disc bulge region A3' to the second weight w2', the weight of the vascular bulge region A5 to the third weight v1', and the weight of the non-vascular bulge region A5' to the fourth weight v2'. Then, the weight of the vascular region A40 within the second optic disc is the first weight multiplied by the third weight, i.e., w1'v1'. The weight of the non-vascular region A40' within the second optic disc is the first weight multiplied by the fourth weight, i.e., w'1v2'. The weight of the vascular region A41 outside the second optic disc is the second weight multiplied by the third weight, i.e., w2'v1'. The weight of the non-vascular region A41' outside the second optic disc is the second weight multiplied by the fourth weight, i.e., w2'v2' (see [reference]). Figure 9 (c)).

[0114] In this disclosure, the difference between the second mixed weight distribution map P4 and the first mixed weight distribution map P3 is that the first mixed weight distribution map P3 is a mixed weight distribution map generated based on the optic disc annotation image and the blood vessel image, while the second mixed weight distribution map P4 is a mixed weight distribution map generated based on the optic disc dilation image and the blood vessel dilation image. Therefore, the mixed weighting processing of the second mixed weight distribution map P3 can be referred to the mixed weighting processing of the first mixed weight distribution map P4, and will not be repeated here.

[0115] Examples of this disclosure are not limited to these; for example, a mixed weight distribution map can be generated based on optic disc labeled images and vascular dilatation images, or a mixed weight distribution map can be generated based on optic disc dilatation images and vascular images.

[0116] In this embodiment, as described above, the training system 10 may include a model training module 500. The model training module 500 may include an artificial neural network. The model training module 500 can train the artificial neural network based on a preprocessed fundus image, an annotated image, and a mixed weight distribution map. The preprocessed fundus image may be generated by the image preprocessing module 200. The annotated image may be generated by the acquisition module 100. The mixed weight distribution map may be generated by the mixed weight generation module 400.

[0117] Specifically, the labeled images, including the optic disc and / or optic cup, can be used as ground truth to predict each pixel in the preprocessed fundus image. A mixed weight distribution map is then used to assign loss function weights (i.e., coefficients of the loss function) to each pixel in the preprocessed fundus image. The artificial neural network is trained based on the loss function, and its output is optimized to obtain the optimal model. In this approach, the artificial neural network exhibits good segmentation accuracy and generalization ability and can automatically extract glaucoma image features.

[0118] In some examples, as mentioned above, the glaucoma image feature can be the cup-to-disc ratio (CDR). In this case, the cup-to-disc ratio of the fundus image can be predicted based on the optimal model, and the glaucoma lesions that may exist in the fundus image can be accurately identified.

[0119] However, the examples disclosed herein are not limited thereto, and the artificial neural network can be replaced by other image feature extraction models. Preferably, the other image feature extraction model can employ UNet or its improved types as the artificial neural network for glaucoma image feature extraction.

[0120] In this disclosure, the loss function can be used to calculate the loss and measure the quality of the model's predictions. Specifically, the difference between the predicted value and the true value of a single sample by an artificial neural network-based model can be called the loss. The smaller the loss, the better the model. In this invention, a single sample can refer to each pixel in a preprocessed fundus image.

[0121] In some examples, the loss function can be a predefined loss function; in others, it can be the cross-entropy loss function, the Dice loss function, etc. The cross-entropy loss function measures the difference between the true distribution and the predicted distribution, while the Dice loss function is a set similarity metric.

[0122] Specifically, taking the cross-entropy loss function as an example, the loss function for preprocessing each pixel in the fundus image is:

[0123]

[0124] Where c represents the predicted category for each pixel in the preprocessed fundus image, and the predicted category includes either optic cup or optic disc. (i,j) represents the coordinates of the pixel in the preprocessed fundus image. This represents the value of the pixel at coordinates (i,j) in the optic cup or optic disc labeled image, which is used as the true value of the pixel at coordinates (i,j) in the preprocessed fundus image. This represents the predicted value of the pixel at coordinates (i,j) in the preprocessed fundus image. c Weights for each category.

[0125] In some examples, a hybrid weight distribution map can be used to assign the weights of the loss function to each pixel in the preprocessed fundus image. As mentioned above, the hybrid weight distribution map can include a fundus region and a background region. The fundus region can be a region within the eyeball contour. The fundus region can include four hybrid regions: the vascular region within the first optic disc, the non-vascular region within the first optic disc, the vascular region outside the first optic disc, and the non-vascular region outside the first optic disc. The background region can be a region outside the eyeball contour. The background region can be a portion of the non-vascular region outside the first optic disc. In some examples, the weight of the optic disc region can be the first weight, the weight of the non-optic disc region can be the second weight, the weight of the vascular region can be the third weight, the weight of the non-vascular region can be the fourth weight, and the weight of the background region can be zero (that is, the weight of the portion of the non-vascular region outside the first optic disc that belongs to the background region is zero). Then the value of each pixel in the hybrid weight distribution map (that is, the weight of the loss function of each pixel in the preprocessed fundus image) can be as shown in the following equation (11):

[0126]

[0127] In the formula, w i,j p represents the weight of the pixel with coordinates (i,j). i,j Let (i,j) be the pixel, w1 be the weight of the pixel inside the optic disc (first weight), w2 be the weight of the pixel outside the optic disc (second weight), v1 be the weight of the pixel in the vascular region (third weight), v2 be the weight of the pixel in the non-vascular region (fourth weight), R1 be the set of pixels in the vascular region inside the optic disc, R2 be the set of pixels in the non-vascular region inside the optic disc, R3 be the set of pixels in the vascular region outside the optic disc, R4 be the set of pixels in the non-vascular region outside the optic disc, and R5 be the set of pixels in the background region.

[0128] In some examples, the loss function L of the artificial neural network can be obtained based on the weights of the loss function for each pixel:

[0129] L=∑ i,j (w i,j *loss i,j ...Equation (12)

[0130] Among them, w i,j The weights for the pixel at coordinates (i,j), loss i,j Let be the loss function for the pixel at coordinates (i,j). Therefore, we can train the artificial neural network based on the loss function to optimize its output and obtain the optimal model.

[0131] In some examples, minimum gradient descent can be used to optimize the parameters of an artificial neural network, adjusting them according to the direction of the fastest descent of the loss function. This allows for optimization of the training system using the coefficients in the loss function. In other examples, stochastic gradient descent can be used for parameter optimization.

[0132] In some examples, the model training module 500 can use the optimal model to extract glaucoma image features from fundus images, thereby predicting possible lesions in the fundus images. In some examples, the trained artificial neural network can be used to identify fundus images in the test set, achieving an average recognition accuracy of, for example, over 90%. Therefore, the training system 10 of this embodiment can achieve improved accuracy in glaucoma lesion diagnosis while taking into account clinical fundus conditions.

[0133] The following, combined with Figure 10 This disclosure describes in detail the training method for glaucoma image feature extraction based on artificial neural networks. The training method for glaucoma image feature extraction based on artificial neural networks disclosed herein can be simply referred to as the training method. The training method disclosed herein is applied to the aforementioned training system 10. Figure 10A flowchart is shown of a training method for glaucoma image feature extraction based on an artificial neural network, according to an embodiment of this disclosure.

[0134] In this embodiment, the training method for glaucoma image feature extraction based on an artificial neural network may include the following steps: preparing a fundus image and a labeled image (step S100); preprocessing the fundus image to form a preprocessed fundus image (step S200); detecting blood vessel regions in the preprocessed fundus image to form a blood vessel image (step S300); forming a hybrid weight distribution map based on the labeled image and the blood vessel image (step S400); and training the artificial neural network based on the preprocessed fundus image, the labeled image, and the hybrid weight distribution map (step S500). In this case, training the artificial neural network based on the preprocessed fundus image, the labeled image, and the hybrid weight distribution map enables the artificial neural network to take into account both the blood vessel region and the optic disc region during training, optimize the learning of small blood vessel course, and suppress positive and negative sample imbalance. Therefore, the accuracy of the artificial neural network in extracting glaucoma image features can be improved.

[0135] In step S100, a fundus image can be prepared. The fundus image can be an image of the fundus captured by a fundus camera or other fundus imaging device. It can be a color fundus image, or an image in RGB mode, CMYK mode, Lab mode, or grayscale mode, etc. For a detailed description, please refer to the acquisition module 100, which will not be repeated here.

[0136] In step S100, labeled images can be prepared. In some examples, the labeled images may include optic disc labeled images and optic cup labeled images. The labeled images can be obtained by experienced physicians manually labeling the optic disc and optic cup regions in the fundus image. This improves the accuracy of the optic disc and optic cup region labeling. For a detailed description, please refer to the acquisition module 100, which will not be repeated here.

[0137] In step S200, the fundus image can be preprocessed to obtain a preprocessed fundus image. In some examples, the fundus image can be cropped, normalized, or otherwise processed during preprocessing. This allows the fundus image to be converted into a fixed standard format and overcomes the differences between different fundus images, improving the performance of the artificial neural network. In some examples, the fundus image can be denoised and converted to grayscale during preprocessing. This highlights the characteristics of the glaucoma image. In some examples, the fundus image can be scaled, flipped, or translated during preprocessing, thereby increasing the amount of data for training the artificial neural network and improving its generalization ability. For a detailed description, please refer to the image preprocessing module 200, which will not be repeated here.

[0138] In step S300, vascular region detection can be performed on the preprocessed fundus image to form a vascular image including vascular regions. In some examples, vascular region detection can be performed based on Frangi filtering to form a vascular image. In this case, the vascular regions in the vascular image are more prominent compared to the vascular regions in the preprocessed fundus image. This facilitates the subsequent identification and processing of vascular regions. In other examples, matched filtering algorithms, adaptive contrast enhancement algorithms, two-dimensional Gabor filtering algorithms, or other types of artificial neural networks can also be used. Thus, appropriate algorithms or artificial neural networks can be selected according to different needs to achieve vascular region detection. For details on vascular region detection, please refer to the vascular region detection module 300, which will not be elaborated here.

[0139] In step S400, the optic disc annotation image obtained in step S100 and the blood vessel image obtained in step S300 can be mixed and weighted to generate a mixed weight distribution map. In this case, using the mixed weight distribution map to train the artificial neural network enables the artificial neural network to take into account both the blood vessel region and the optic disc region during training, optimize the learning of small blood vessel course, and suppress the imbalance between positive and negative samples.

[0140] In step S400, the mixed weight distribution map may include four mixed regions: the intradiscal vascular region, the intradiscal non-vascular region, the extradiscal vascular region, and the extradiscal non-vascular region. In some examples, the mixed weight distribution map may include the fundus region and the background region, and the weight of the background region may be set to zero. This reduces the interference of the background region on the artificial neural network's extraction of glaucoma image features during training. For details, please refer to the mixed weight generation module 400, which will not be repeated here.

[0141] In step S400, in some examples, the weight of the optic disc region can be set as a first weight, the weight of the non-optic disc region as a second weight, the weight of the vascular region as a third weight, and the weight of the non-vascular region as a fourth weight. Then, the weight of the vascular region within the optic disc is the first weight multiplied by the third weight. The weight of the non-vascular region within the optic disc is the first weight multiplied by the fourth weight. The weight of the vascular region outside the optic disc is the second weight multiplied by the third weight. The weight of the non-vascular region outside the optic disc is the second weight multiplied by the fourth weight. Thus, the weight values ​​of the vascular region within the optic disc, the non-vascular region within the optic disc, the vascular region outside the optic disc, and the non-vascular region outside the optic disc can be derived based on their respective weights. In some examples, the weight of the optic disc region can be greater than the weight of the non-optic disc region, and the weight of the vascular region can be greater than the weight of the non-vascular region. For details, please refer to the mixed weight generation module 400, which will not be elaborated here.

[0142] In step S400, in some other examples, the optic disc region in the optic disc annotation image can be dilated to form an optic disc dilated image including the optic disc region. Since the region near the optic disc affects the segmentation of the optic cup or optic disc, and thus affects the extraction of glaucoma image features, dilation processing is used to obtain an optic disc dilated image so that subsequent processing can be performed based on the optic disc dilated image, thereby improving the accuracy of glaucoma image feature extraction. For details, please refer to the optic disc dilation module 410, which will not be elaborated here.

[0143] In step S400, in some other examples, the vascular region in the vascular image can be dilated to form a vascular dilated image including the vascular region. In this case, the dilation process can reduce the error in detecting vascular boundaries based on the vascular detection algorithm. For details, please refer to the vascular dilation module 420, which will not be repeated here.

[0144] In step S400, in some other examples, the aforementioned optic disc dilation image and vascular dilation image can be mixed and weighted to generate a mixed weight distribution map. In this case, using the mixed weight distribution map generated based on the dilated optic disc image and the dilated vascular image to train the artificial neural network can reduce the error of the vascular detection algorithm at the vascular boundary. For details, please refer to the mixed weight generation module 400, which will not be repeated here.

[0145] In step S500, the artificial neural network can be trained based on the preprocessed fundus image, the labeled image, and the mixed weight distribution map. The preprocessed fundus image can be generated in step S200. The labeled image can be generated in step S100. The mixed weight distribution map can be generated in step S400. In some examples, the loss function weights, i.e., the coefficients of the loss function, can be assigned to each pixel in the preprocessed fundus image using the mixed weight distribution map. The artificial neural network is trained based on the loss function, and the output of the artificial neural network is optimized to obtain the optimal model of the artificial neural network. In this case, the artificial neural network has good segmentation accuracy and generalization ability and can automatically extract glaucoma image features. For details, please refer to the model training module 500, which will not be repeated here.

[0146] While the present disclosure has been specifically described above in conjunction with the accompanying drawings and embodiments, it is to be understood that the above description does not limit the present disclosure in any way. Those skilled in the art can make modifications and variations to the present disclosure as needed without departing from its essential spirit and scope, and all such modifications and variations fall within the scope of the present disclosure.

Claims

1. A training method for glaucoma image feature extraction based on artificial neural networks, characterized in that, include: Prepare fundus images and labeled images, wherein the labeled images include optic disc labeled images of the optic disc region and optic cup labeled images of the optic cup region; The fundus image is preprocessed to obtain a preprocessed fundus image, and a vascular image containing vascular regions is generated based on the vascular detection results. The optic disc annotation image and the vascular image are mixed and weighted to generate a mixed weight distribution map. This mixed weight distribution map includes intra-optic disc vascular regions, intra-optic disc non-vascular regions, extra-optic disc vascular regions, and extra-optic disc non-vascular regions. During the mixed weighting, the weight of the optic disc region is set greater than the weight of the non-optic disc region, and the weight of the vascular region is set greater than the weight of the non-vascular region. The artificial neural network is trained based on the preprocessed fundus image, the labeled image, and the hybrid weight distribution map. During training, the labeled image is used as the ground truth for training the artificial neural network, each pixel in the preprocessed fundus image is predicted, and the loss function weights of each pixel in the preprocessed fundus image are assigned using the hybrid weight distribution map.

2. The training method according to claim 1, characterized in that, When performing the mixed weighting, let the weight of the optic disc region be the first weight, the weight of the non-optic disc region be the second weight, the weight of the vascular region be the third weight, and the weight of the non-vascular region be the fourth weight. Then, the weight of the intra-optic disc vascular region is the first weight multiplied by the third weight, the weight of the intra-optic disc non-vascular region is the first weight multiplied by the fourth weight, the weight of the extra-optic disc vascular region is the second weight multiplied by the third weight, and the weight of the extra-optic disc non-vascular region is the second weight multiplied by the fourth weight.

3. The training method according to claim 1, characterized in that, The mixed weight distribution map is divided based on the eyeball contour. The mixed weight distribution map includes a fundus region and a background region. The fundus region is the region within the eyeball contour, and the background region is the region outside the eyeball contour. The fundus region includes the intraoptic disc vascular region, the intraoptic disc non-vascular region, the extraoptic disc vascular region, and the extraoptic disc non-vascular region. The background region includes a portion of the extraoptic disc non-vascular region.

4. The training method according to claim 3, characterized in that, When training the artificial neural network, the weight of the background region is set to zero.

5. The training method according to claim 1, characterized in that, The visual disc region in the visual disc annotation image is dilated to form a dilated visual disc image; The blood vessel region in the blood vessel image is dilated to form a blood vessel dilation image; The optic disc dilatation image and the vascular dilatation image are mixed and weighted to generate the mixed weight distribution map.

6. The training method according to claim 5, characterized in that, The optic disc dilation image includes an optic disc dilation region, which includes the optic disc region and a region near the optic disc. The optic disc dilation region corresponds to the optic disc region in the optic disc annotation image. The vascular dilation image includes a vascular dilation region, which includes the vascular region and a region near the vascular region. The vascular dilation region corresponds to the vascular region in the vascular image.

7. The training method according to claim 5, characterized in that, The optic disc dilatation image includes an optic disc dilatation region and a non-optic disc dilatation region, and the vascular dilatation image includes a vascular dilatation region and a non-vascular dilatation region. Let the weight of the optic disc dilatation region be a first weight, the weight of the non-optic disc dilatation region be a second weight, the weight of the vascular dilatation region be a third weight, and the weight of the non-vascular dilatation region be a fourth weight. The weight of the intra-optic disc vascular region is the first weight multiplied by the third weight. The weight of the intra-optic disc non-vascular region is the first weight multiplied by the fourth weight. The weight of the extra-optic disc vascular region is the second weight multiplied by the third weight, and the weight of the extra-optic disc non-vascular region is the second weight multiplied by the fourth weight.

8. The training method according to claim 1, characterized in that, During training, the coefficients of the loss function in the artificial neural network are obtained based on the hybrid weight distribution map, and the artificial neural network is trained based on the loss function.

9. The training method according to claim 1, characterized in that, The blood vessel image is formed by detecting blood vessel regions using at least one of the following algorithms: Frangi filtering, matched filtering, adaptive contrast enhancement, and two-dimensional Gabor filtering.

10. A training system for glaucoma image feature extraction based on artificial neural networks, characterized in that, It includes an acquisition module, an image preprocessing module, a blood vessel region detection module, a hybrid weight generation module, and a model training module; The acquisition module is configured to acquire fundus images and labeled images, wherein the labeled images include optic disc labeled images of the optic disc region and optic cup labeled images of the optic cup region; The image preprocessing module is configured to preprocess the fundus image to obtain a preprocessed fundus image; The blood vessel region detection module is configured to perform blood vessel region detection on the preprocessed fundus image to form a blood vessel image; The hybrid weighting generation module is configured to perform a hybrid weighting of the optic disc annotation image and the vascular image to generate a hybrid weighting distribution map. The hybrid weighting distribution map includes an intra-optic disc vascular region, an intra-optic disc non-vascular region, an extra-optic disc vascular region, and an extra-optic disc non-vascular region. During the hybrid weighting, the weight of the optic disc region is set greater than the weight of the non-optic disc region, and the weight of the vascular region is set greater than the weight of the non-vascular region. The model training module is configured to train the artificial neural network based on the preprocessed fundus image, the labeled image, and the hybrid weight distribution map. During training, the labeled image is used as the ground truth for training the artificial neural network, each pixel in the preprocessed fundus image is predicted, and the loss function weights of each pixel in the preprocessed fundus image are assigned using the hybrid weight distribution map.

Images