Preterm infant retinal fundus image quantitative index and staging method based on meta learning

Abstract

The application discloses a premature infant retinal fundus image quantitative index and staging method based on meta learning, and when facing new application scenes (such as ROP images of different hospitals or different imaging equipment), a model with better performance can be trained by using a small amount of data and time. The quantitative analysis result of the segmentation model combined by the original prototype segmentation network ADNet and GAN is applied to ROP lesion staging, a better effect can be achieved by using a small amount of data and faster speed, and the classification accuracy of the first and second stages of the retinopathy of prematurity is improved.

Description

Technical Field

[0001] The present invention relates to a meta-learning-based quantitative index and staging method for premature infant retinal fundus images, and belongs to the technical field of computer vision processing. Background Art

[0002] Retinopathy of prematurity (ROP) is a disease associated with retinal vascular development and is generally believed to be related to oxygen concentration. At birth, premature infants have immature retinal blood vessels, with peripheral avascular zones and developing vascular terminals as undifferentiated neovessels. Immature retinal vessels are highly sensitive to oxygen. High oxygen concentrations can damage retinal capillary endothelial cells, cause vasoconstriction, and occlusion, leading to retinal hypoxia and stimulating the proliferation of fibrovascular tissue. The incidence of ROP increases with gestational age and birth weight. The incidence rate in premature infants in my country is approximately 6%-7%. ROP can cause abnormal vascular rupture and bleeding, fibrosis, and, in severe cases, fundus lesions and severe vision loss. If left unchecked, ROP can lead to serious late-stage complications such as secondary glaucoma, cataracts, corneal and scleral staphylomas, and ocular atrophy, and can even cause blindness, causing lifelong suffering for the affected child and placing a burden on their family and society.

[0003] With the development of artificial intelligence (AI) technology, automated screening and staging for ROP (retinopathy of prematurity) has become possible, significantly alleviating the problem of insufficient medical resources. However, the application of current technologies still faces certain challenges. First, ROP data comes from a wide range of sources, and datasets vary between different devices or hospitals, making direct application impossible. Targeted data collection and model training are required, which consumes significant human, material, financial, and time resources. Second, ROP data annotation is difficult. Because ROP lesions are unclear, specialized physicians are required for annotation, making it difficult to obtain sufficient data images for dataset construction. Finally, ROP staging is challenging. Generally speaking, ROP is clinically divided into five stages: demarcation stage, ridge formation stage, proliferative stage, subtotal retinal detachment stage, and total retinal detachment stage. However, stages I and II are too similar to each other to be classified. These issues have led to significant difficulties in the clinical application of ROP staging. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a meta-learning-based quantitative index and staging method for retinal fundus images of premature infants. By introducing the meta-learning method, the quantitative analysis results of the segmentation model combining the prototype segmentation network ADNet and GAN are applied to ROP lesion staging. Better results can be achieved with a small amount of data and faster speed, and the classification accuracy of stage I and stage II retinopathy of prematurity is improved.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A meta-learning-based quantitative index and staging method for retinal fundus images of premature infants includes the following steps:

[0007] (1) Metadata collection: Two sets of ROP datasets from different sources are collected as meta-training data for the meta-learning training process and meta-testing data for the meta-learning testing process, respectively. The meta-training data is the source domain data, and the meta-testing data is the target domain data. The data are placed in different folders according to the six stages: normal, first phase, second phase, third phase, fourth phase, and fifth phase.

[0008] (2) Metadataset annotation: Labelme software was used to annotate the ROP dataset and obtain the segmentation mask of the ROP lesions, which is the gold standard;

[0009] (3) Metadata set preprocessing: A fast algorithm is used to remove patient information from the ROP image and convert the ROP image into YCrCb space. Adaptive histogram equalization is then used to enhance the contrast between the lesion and the background while preserving the ROP details. The image and the annotated image are then amplified.

[0010] (4) Establishing a meta-learning model: A segmentation model combining the prototype network ADNet and GAN was used as the network model for the ROP lesion segmentation task, and Resnet101 was used as the backbone network;

[0011] (5) Meta-learning model pre-training: Pre-train the meta-learning model on the COCO dataset using deep learning and save the pre-training weights;

[0012] (6) Meta-training stage of meta-learning model: Samples are collected from the source domain data to form several training tasks for model training. Each training task contains a support set and a query set. The support set is used to generate the lesion prototype, and the query set is used to generate the segmentation result with the lesion prototype. The construction of the meta-training dataset adopts 6-way and 5-shot, that is, the training task contains six categories (from normal and one to five phases), and each category in the support set has 5 training images. The default support set has 15 images for each category. In addition, the loss function uses the cross entropy loss function, and the loss here is recorded as L1.

[0013] (7) Training of the GAN discriminator: Only the discriminator part of the generative adversarial network is used. The discriminator is a binary convolutional neural network. The training of the discriminator requires the query set image and its segmentation result image and annotated image in the meta-training stage. The two categories of the discriminator are: (a) ROP original image and its segmentation result image; (b) ROP original image and its annotated image. Therefore, there are two images input into the discriminator at the same time. The label for category (a) is 0 (i.e. fake), and the label for category (b) is 1 (i.e. true). In addition, the loss function here uses the cross entropy loss function, and the loss here is recorded as L2.

[0014] (8) Network parameter update: The total network update loss L is the sum of L1 and L2, and backpropagation updates the meta-learner and discriminator.

[0015] (9) Meta-testing phase of meta-learning model: The meta-testing phase is similar to the meta-training phase. Samples are collected from the target domain data to generate test tasks. Each test task also contains a support set and a test set. The support set is used to generate lesion prototypes, and the test set is used to generate segmentation results with the lesion prototypes.

[0016] (10) Post-processing of segmentation results: The output segmentation results are post-processed using the dilation-erosion method. The purpose is to connect the segmented lesion areas to facilitate the calculation of subsequent quantitative indicators, including length, average width, area, and degree of conspicuity, to assist in ROP staging and classification. In addition, it is necessary to first perform an indicator analysis on the meta-test dataset and statistically analyze the length, average width, area, and degree of conspicuity of lesions in different stages. Then, the quantitative indicators of the segmentation results of the query set in the meta-test are compared with the statistical results to perform ROP staging.

[0017] The aforementioned meta-learning-based quantitative index and staging method for premature infant retinal fundus images is characterized in that the calculation of the length index requires the use of a skeleton extraction algorithm, the specific algorithm is as follows:

[0018]

[0019] Where N represents the number of skeleton pixels; P i bone Represents the i-th skeleton pixel.

[0020] The aforementioned meta-learning-based quantitative index and staging method for premature infant retinal fundus images is characterized by removing the lesion area on one side of the skeleton, calculating the distance between each skeleton pixel and the nearest lesion boundary pixel, and taking the average of all distances as the average width, as shown in the following formula:

[0021]

[0022] Where N represents the number of skeleton pixels, i.e., the length of the lesion; P i represents the i-th skeleton pixel; P i ′ represents the lesion boundary pixel closest to the i-th skeleton pixel.

[0023] The aforementioned meta-learning-based quantitative index and staging method for premature infant retinal fundus images is characterized by the calculation of the proportion of area, that is, the ratio of the number of pixels in the lesion area to the number of pixels in the entire image, as shown in the following formula:

[0024]

[0025] Among them, N represents the number of lesion pixels, and K represents the number of pixels in the entire image; represents the i-th lesion pixel, Represents the i-th image pixel.

[0026] The aforementioned meta-learning-based quantitative index and staging method for premature infant retinal fundus images is characterized by the following calculation formula for the degree of significance:

[0027] Obvious=AVE sur -AVE seg

[0028] Among them, AVE seg Indicates the average pixel value of the lesion area segmented by the network on the original image, AVE sur It represents the average pixel value of the area around the lesion area in the original image.

[0029] The aforementioned quantitative index and staging method for retinal fundus images of premature infants based on meta-learning is characterized in that: the calculation process of L1 is: directly compare the segmentation result with the gold standard, and use the nn.CrossEntropyLoss() function to calculate L1. The specific calculation process is as follows: take the log of each pixel value of the segmentation result, then multiply it with the gold standard, and finally add each element to take the average value, and then take the inverse to obtain the final L1; the calculation process of L2 is: use the nn.CrossEntropyLoss() function to calculate L 2. The specific process is as follows: the discriminator outputs a two-dimensional vector, the two-dimensional output vector is processed by the Softmax function, and then the two-dimensional output vector is logarithmized. The label (either 0 or 1) is one-hot processed to obtain a two-dimensional label vector. The two-dimensional output vector is multiplied element by element with the two-dimensional label vector, and the sum is averaged to obtain L2; L1 and L2 are added together to obtain the total loss L. Each parameter in Resnet101 and the discriminator can be chain-derived with respect to the total loss L to obtain a gradient value, and each parameter is updated by gradient descent. The beneficial effects of the present invention are:

[0030] 1. When faced with ROP images in new application scenarios (such as different hospitals or different imaging equipment), a model with good performance can be trained with a small amount of data and time;

[0031] 2. By using some quantitative indicators to analyze the segmentation results, we can assist in the classification of the first and second stages of premature infants' retinal images and achieve a better classification effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of a meta-learning-based quantitative index and staging method for retinal fundus images of premature infants in the present invention;

[0033] Figure 2 This is a network model diagram of a meta-learning-based quantitative index and staging method for retinal fundus images of premature infants in the present invention;

[0034] Figure 3 This is the original image taken by the ROP device in the meta-learning-based quantitative index and staging method for retinal fundus images of premature infants of the present invention;

[0035] Figure 4 This is a schematic diagram of the desensitized original image taken by the ROP device in a meta-learning-based quantitative index and staging method for retinal fundus images of premature infants of the present invention;

[0036] Figure 5This is a schematic diagram of the adaptive histogram equalization of the desensitized image in a meta-learning-based quantitative index and staging method for premature infant retinal fundus images of the present invention;

[0037] Figure 6 This is a schematic diagram of the lesion area segmented in a meta-learning-based quantitative index and staging method for retinal fundus images of premature infants of the present invention;

[0038] Figure 7 This is a schematic diagram of the connected lesion areas segmented in a meta-learning-based quantitative index and staging method for premature infant retinal fundus images of the present invention;

[0039] Figure 8 This is a schematic diagram of the length of the connected lesion area in a meta-learning-based quantitative index and staging method for retinal fundus images of premature infants of the present invention;

[0040] Figure 9 This is a schematic diagram of one side of the connected lesion skeleton in a meta-learning-based quantitative index and staging method for retinal fundus images of premature infants of the present invention;

[0041] Figure 10 This is a schematic diagram of the degree of conspicuity of the lesion area in the quantitative indicators and staging method of retinal fundus images of premature infants based on meta-learning in the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to specific embodiments.

[0043] like Figure 1 As shown in FIG, a meta-learning-based quantitative index and staging method for retinal fundus images of premature infants includes the following steps:

[0044] (1) Metadata Collection: Two sets of ROP datasets from different sources are collected as meta-training data for the meta-learning training process and meta-testing data for the meta-learning testing process. The meta-training data is the source domain data, and the meta-testing data is the target domain data. The data are placed in different folders according to the six stages: normal, stage 1, stage 2, stage 3, stage 4, and stage 5. When dividing the data, the severity of the disease is determined. Normal means no disease, and as the stage 1 to stage 5 increases, the disease becomes more and more serious. This data division method facilitates the construction of learning tasks in the later stage. Constructing the learning task requires sampling a number of images in each of the six folders.

[0045] (2) Metadataset annotation: Labelme software is used to annotate the ROP dataset to obtain the segmentation mask of the ROP lesion, which is the gold standard. The lesions are marked on the ROP image to obtain the segmentation mask, which is the gold standard. This is so that the subsequent model can output the segmentation results and compare them with the manually annotated gold standard. Only after the comparison can the model be updated.

[0046] (3) Metadata set preprocessing: Use a fast algorithm to remove patient information on the ROP image, and convert the ROP image to YCrCb space. Then use adaptive histogram equalization to enhance the contrast between the lesion and the background while retaining the ROP details. Figure 5 As shown, the pictures and annotation images are then amplified, including rotation, cutting, color change, etc. The purpose of amplification is to enhance the generalization ability of the model. For example, we can rotate the ROP image by a certain angle to obtain a new ROP image, and then input it into the model for segmentation processing. In this way, if some ROP images are also at a certain angle in the future, the model can also perform good segmentation. The purpose of amplification is to enhance the generalization ability of the model. Since some ROP devices take images (such as Figure 3 As shown in Figure 2, there is some sensitive information of patients on the ROP image, so the Fast Marching Method (FMM) is first used to remove the patient information on the ROP image, as shown in Figure 2. Figure 4 As shown;

[0047] (4) Establishing a meta-learning model: A segmentation model combining the prototype network ADNet and GAN was used as the network model for the ROP lesion segmentation task, and Resnet101 was used as the backbone network, such as Figure 2 As shown in the figure, the meta-learning model includes a base model (resnet101) and a discriminator (four convolutional layers and two fully connected layers). First, the support set and the query set are input into resnet101 to obtain the feature map of the support set and the feature map of the query set. The feature map of the support set is compared with that of the support set, and the masked average pooling method is used to calculate the lesion prototype. The lesion prototype and the feature map of the query set are compared using the cosine similarity calculation method to obtain the segmentation map of the query set. The segmentation map of the query set is then activated by the sigmoid function to obtain the final segmentation map of the query set. The query set segmentation map and the original query set map are input into the discriminator at the same time. Then, the query set gold standard and the original query set map are input into the discriminator at the same time.

[0048] The segmentation model that combines the prototype network ADNet and GAN has the following advantages for the ROP lesion segmentation task: ADNet's meta-learning method can be trained using small sample data (because labeled data is still relatively expensive) and still obtain good segmentation results; and when encountering ROP images from other hospitals or equipment, only a small number of images are needed to adapt to the new usage environment; the discriminator in GAN can improve segmentation accuracy. The advantage of using ResNet101 is that due to its deeper network, the network performance is stronger, and the ready-made pre-trained model parameter weights can be directly downloaded;

[0049] (5) Meta-learning model pre-training: The meta-learning model is pre-trained on the COCO dataset using deep learning. The purpose of pre-training is to obtain pre-training weights, that is, pre-training model parameters. Save the pre-trained weights. You can read the trained weights when performing meta-training later. Specifically, use the conventional deep learning training method to pre-train the model parameters of resnet101. Input the image into resnet101 to directly obtain the segmentation result. Then calculate the loss with the gold standard and update the resnet101 model parameters. (6) Meta-learning model meta-training stage: Collect samples from the source domain data to form several training tasks for model training. Each training task contains a support set and a query set. The support set is used to generate the lesion prototype, and the query set is used to generate the segmentation result with the lesion prototype. The construction of the meta-training dataset adopts 6-way and 5-shot, that is, the training task contains six categories (from normal and one to five phases). Each category in the support set has 5 training images. By default, each category in the support set has 15 images. This construction task is a requirement for meta-learning construction tasks. The 6-way indicates that there are six categories, that is, six prototypes will be generated. If the losses of the six categories are added up and then updated, better results can be achieved. In fact, it can also be viewed as six 1-way tasks, except that the losses for all six 1-way tasks are summed and averaged before updating the model. Furthermore, the cross-entropy loss function is used, and the loss here is denoted as L1. Regarding how the support set generates lesion prototypes, let's first assume that the output of a ResNet101 image in the support set has dimensions of 1×64×224×224, meaning that the feature map has 64 channels. A gold standard image of this image (e.g., dimensions of 1×224×224) is used, meaning that the gold standard has only one channel and is a binary image, meaning that the pixel values ​​for the lesion are 1 and the pixels for the background are 0. Multiply each channel of the gold standard by the feature map, resulting in a feature map of 1×64×224×224. Next, sum the height and width of the feature map, resulting in a feature map of 1×64. Finally, sum the height and width of the gold standard, resulting in a gold standard of 1×64. Finally, the feature map (1×64) is divided by the gold standard (1×64) to obtain the lesion prototype with a size of 1×64. Since the actual support set has multiple images, multiple prototypes will be generated. Simply taking the average of multiple prototypes can get a lesion prototype. Regarding how the query set and the lesion prototype produce segmentation results, first, assume that the feature map size of a picture in the query set after resnet101 output is 1×64×224×224, and the lesion prototype size is 1×64. The lesion prototype is expanded to 1×64×224×224 using the pytorch broadcast mechanism, and the cosine similarity between the feature map and the prototype is calculated. The segmentation result map is 1×224×224.

[0050] The calculation process of L1 is: directly compare the segmentation result with the gold standard, using nn The .CrossEntropyLoss() function calculates L1. The specific calculation process is as follows: take the log of each pixel value of the segmentation result, then multiply it by the gold standard, and finally add each element to get the average value, and then invert it to get the final L1;

[0051] (7) Training of the GAN discriminator: Only the discriminator part of the generative adversarial network is used. The discriminator is a binary convolutional neural network. The training of the discriminator requires the query set image and its segmentation result image and the annotated image in the meta-training phase. The two categories of the discriminator are: (a) ROP original image and its segmentation result image; (b) ROP original image and its annotated image. Therefore, there are two images input to the discriminator at the same time. The label for category (a) is 0 (i.e. fake), and the label for category 2 is (b) (i.e. true). The first category The input of the first category is the ROP original image in the query set and the corresponding segmentation result image, and the output is 0; the input of the second category is the ROP original image in the query set and the corresponding labeled image, and the output is 1; it is expected to enhance the discriminator's ability to distinguish between true labels and segmentation result images to promote the segmentation effect, and to reversely promote the segmentation accuracy of ADNet by enabling the discriminator to learn to distinguish whether a certain segmentation result is the gold standard or the result of model processing; in addition, the loss function here uses the cross entropy loss function, and the loss here is recorded as L2, where L2 is directly calculated using the classification cross entropy, using nn The CrossEntropyLoss() function calculates L2. The specific process is as follows: the discriminator outputs a two-dimensional vector, which is processed by the Softmax function. Then the log of the two-dimensional output vector is taken. The label (either 0 or 1) is one-hot processed to obtain a two-dimensional label vector. The two-dimensional output vector is multiplied element-by-element by the two-dimensional label vector, and the sum is averaged to obtain L2.

[0052] (8) Network parameter update: The total loss L of the network update is the sum of L1 and L2. Backpropagation is used to update the meta-learner and discriminator. L1 and L2 are directly added together to get the total loss L. Each parameter in Resnet101 and the discriminator can be chain-derived with respect to the total loss L to obtain a gradient value. Each parameter is then updated by gradient descent.

[0053] (9) Meta-testing stage of meta-learning model: The meta-testing stage is similar to the meta-training stage. Samples are collected from the target domain data to generate test tasks. Each test task also contains a support set and a test set. The support set is used to generate lesion prototypes, and the test set is used to generate segmentation results with the lesion prototypes, such as Figure 6 As shown;

[0054] (10) Post-processing of segmentation results: The output segmentation results are post-processed by dilation and corrosion, with the purpose of connecting the segmented lesion areas, such as Figure 7 As shown, this facilitates the calculation of subsequent quantitative indicators, including length, average width, area, and degree of conspicuity, to assist in ROP staging and classification. Furthermore, it is necessary to first perform indicator analysis on the meta-test dataset and statistically analyze the length, average width, area, and degree of conspicuity of lesions in different stages. The quantitative indicators of the segmentation results of the query set in the meta-test are then compared with the statistical results to perform ROP staging. The method of connecting the segmented lesion regions using dilation and erosion is as follows: First, the segmentation result is a binary image, with the pixel value of the lesion portion being 1 and the background being 0. Dilation then changes the pixel value around the lesion portion from 0 to 1, and erosion changes the pixel value at the edge of the lesion from 1 to 0. After n dilation operations, the separated lesion regions are connected. After being connected, n erosion operations are required to restore the original lesion size, but the lesion remains connected (the dilation-then-erosion operation is also called a closing operation).

[0055] like Figure 8 As shown, the calculation of the length index requires the use of a skeleton extraction algorithm, the specific algorithm is as follows:

[0056]

[0057] Where N represents the number of skeleton pixels; P i bone Represents the i-th skeleton pixel point. The skeleton extraction algorithm can effectively obtain the length of the lesion area. The lesion area is generally curved, and the skeleton extraction algorithm can calculate the length of the curved lesion.

[0058] like Figure 9 As shown, the lesion area on one side of the skeleton is removed, the distance between each skeleton pixel and the nearest lesion boundary pixel is calculated, and the average of all distances is taken as the average width, as shown in the following formula:

[0059]

[0060] Where N represents the number of skeleton pixels, i.e., the length of the lesion; P i represents the i-th skeleton pixel; P i ′ represents the lesion boundary pixel closest to the i-th skeleton pixel. Because the skeleton lines generated by the previous skeleton extraction algorithm divide the lesion into two parts, we only need to remove the lesion area on one side of the skeleton and calculate the closest distance from the skeleton to the lesion edge on the other side to obtain the single-side width.

[0061] The calculation of the area ratio is the ratio of the number of pixels in the lesion area to the number of pixels in the entire image, as shown in the following formula:

[0062]

[0063] Among them, N represents the number of lesion pixels, and K represents the number of pixels in the entire image; represents the i-th lesion pixel, Represents the i-th image pixel. This method is simple and clear, which is to add up the number of pixels in the segmented lesion area and divide it by the number of pixels in the entire image.

[0064] The formula for calculating the degree of significance is as follows:

[0065] Obvious=dVE sur -dVE seg

[0066] Among them, AVE seg Indicates the average pixel value of the lesion area segmented by the network on the original image, AVE sur It represents the average pixel value of the area around the lesion area on the original image. The result is as follows Figure 10 Because the ROP lesion is a distinct line or area, the color change between the lesion and the lesion edge can be calculated to determine the degree of lesion visibility.

[0067] In summary, the present invention provides a meta-learning-based quantitative index and staging method for retinal fundus images of premature infants. By introducing the meta-learning method, the quantitative analysis results of the segmentation model combining the prototype segmentation network ADNet and GAN are applied to ROP lesion staging. It can achieve better results with a small amount of data and faster speed, and improve the classification accuracy of stage I and stage II retinopathy of prematurity.

[0068] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A meta-learning-based quantitative index and staging method for retinal fundus images of premature infants, characterized by: The steps include: (1) Metadata collection: Two sets of ROP datasets from different sources are collected as meta-training data for the meta-learning training process and meta-testing data for the meta-learning testing process, respectively. The meta-training data is the source domain data, and the meta-testing data is the target domain data. The data are placed in different folders according to the six stages: normal, first phase, second phase, third phase, fourth phase, and fifth phase. (2) Metadataset annotation: Labelme software was used to annotate the ROP dataset and obtain the segmentation mask of the ROP lesions, which is the gold standard; (3) Metadata set preprocessing: A fast algorithm is used to remove patient information from the ROP image and convert the ROP image into YCrCb space. Adaptive histogram equalization is then used to enhance the contrast between the lesion and the background while preserving the ROP details. The image and the annotated image are then amplified. (4) Establishing a meta-learning model: A segmentation model combining the prototype network ADNet and GAN was used as the network model for the ROP lesion segmentation task, and Resnet101 was used as the backbone network; (5) Meta-learning model pre-training: Pre-train the meta-learning model on the COCO dataset using deep learning and save the pre-training weights; (6) Meta-training stage of meta-learning model: Samples are collected from the source domain data to form several training tasks for model training. Each training task contains a support set and a query set. The support set is used to generate the lesion prototype, and the query set is used to generate the segmentation result with the lesion prototype. The construction of the meta-training dataset adopts 6-way and 5-shot, that is, the training task contains six categories (from normal and one to five phases), and each category in the support set has 5 training images. The default support set has 15 images for each category. In addition, the loss function uses the cross entropy loss function, and the loss here is recorded as L1; (7) Training of the GAN discriminator: Only the discriminator part of the generative adversarial network is used. The discriminator is a binary convolutional neural network. The training of the discriminator requires the query set image and its segmentation result image and annotated image in the meta-training stage. The two categories of the discriminator are: (a) ROP original image and its segmentation result image; (b) ROP original image and its annotated image. Therefore, there are two images input into the discriminator at the same time. The label for category (a) is 0 (i.e. fake), and the label for category (b) is 1 (i.e. true). In addition, the loss function here uses the cross entropy loss function, and the loss here is recorded as L2; (8) Network parameter update: The total network update loss L is the sum of L1 and L2, and backpropagation updates the meta-learner and discriminator; (9) Meta-testing phase of meta-learning model: The meta-testing phase is similar to the meta-training phase. Samples are collected from the target domain data to generate test tasks. Each test task also contains a support set and a test set. The support set is used to generate lesion prototypes, and the test set is used to generate segmentation results with the lesion prototypes. (10) Post-processing of segmentation results: The output segmentation results are post-processed by dilation and corrosion. The purpose is to connect the segmented lesion areas to facilitate the calculation of subsequent quantitative indicators, including length, average width, area share and degree of conspicuity, which are used to assist in ROP staging and classification. In addition, it is necessary to first perform indicator analysis on the meta-test data set and statistically analyze the length, average width, area share and degree of conspicuity of lesions in different stages; then, the quantitative indicators of the segmentation results of the query set in the meta-test are compared with the statistical results to perform ROP staging.

2. The meta-learning-based quantitative index and staging method for retinal fundus images of premature infants according to claim 1, characterized in that: The calculation of the length index requires the use of a skeleton extraction algorithm, the specific algorithm is as follows: Where N represents the number of skeleton pixels; P i bone Represents the i-th skeleton pixel.

3. The meta-learning-based quantitative index and staging method for retinal fundus images of premature infants according to claim 2, characterized in that: Remove the lesion area on one side of the skeleton, calculate the distance between each skeleton pixel and the nearest lesion boundary pixel, and take the average of all distances as the average width, as shown in the following formula: Where N represents the number of skeleton pixels, i.e., the length of the lesion; P i represents the i-th skeleton pixel; P i ′ represents the lesion boundary pixel closest to the i-th skeleton pixel.

4. The meta-learning-based quantitative index and staging method for premature infant retinal fundus images according to claim 1, characterized in that: The calculation of the area ratio is the ratio of the number of pixels in the lesion area to the number of pixels in the entire image, as shown in the following formula: Among them, N represents the number of lesion pixels, and K represents the number of pixels in the entire image; represents the i-th lesion pixel, Represents the i-th image pixel.

5. The meta-learning-based quantitative index and staging method for premature infant retinal fundus images according to claim 1, characterized in that: The formula for calculating the degree of significance is as follows: Among them, AVE seg Indicates the average pixel value of the lesion area segmented by the network on the original image, AVE sur It represents the average pixel value of the area around the lesion area in the original image.

6. The meta-learning-based quantitative index and staging method for retinal fundus images of premature infants according to claim 1, characterized in that: The calculation process of L1 is: directly compare the segmentation result with the gold standard, and use the nn.CrossEntropyLoss() function to calculate L1. The specific calculation process is as follows: take the log of each pixel value of the segmentation result, then multiply it with the gold standard, and finally add each element to get the average, and then invert it to get the final L1; the calculation process of L2 is: use the nn.CrossEntropyLoss() function to calculate L2. The specific process is as follows: the discriminator outputs a two-dimensional vector, the two-dimensional output vector is processed by the Softmax function, and then the two-dimensional output vector is log, the label (either 0 or 1) is one-hot processed to get a two-dimensional label vector, the two-dimensional output vector is multiplied element by element with the two-dimensional label vector, and the sum is averaged to get L2; add L1 and L2 to get the total loss L, each parameter in Resnet101 and the discriminator can be chain-derived with respect to the total loss L to get a gradient value, and each parameter is updated by gradient descent.

Images