Glaucoma image classification method and system based on multi-task learning

Abstract

The invention discloses a glaucoma image classification method and system based on multi-task learning, and the method comprises the following steps: S1, obtaining a multi-mode glaucoma fundus image data set of a target object, and carrying out the normalization and registration preprocessing of the data set, and obtaining a standardized glaucoma image sample set; s2, constructing a multi-task learning classification model containing a shared feature extraction sub-network and a task exclusive branch sub-network based on a dual-task target of glaucoma disease grading classification and focus region segmentation; s3, introducing a multi-task joint loss constraint mechanism, carrying out weighted fusion on the prediction loss of double tasks, and carrying out iterative training on the multi-task learning classification model by adopting a self-adaptive gradient descent algorithm until the model is converged; and S4, inputting a to-be-classified glaucoma image into the trained multi-task learning classification model, synchronously outputting a glaucoma disease classification result and a focus region segmentation classification result, and completing glaucoma image multi-task classification.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a glaucoma image classification method and system based on multi-task learning. Background Technology

[0002] Glaucoma, a blinding eye disease, relies on fundus image analysis for early diagnosis. Existing solutions suffer from problems such as fragmented single-task approaches and insufficient accuracy, making it difficult to simultaneously meet the needs of disease grading and lesion segmentation. The current mainstream solution is a single-task deep learning model, such as a CNN-based glaucoma grading model and a U-Net-based optic disc / cup segmentation model. However, these models have significant drawbacks: single-task models cannot reuse features, resulting in low training efficiency and weak generalization ability. Furthermore, multimodal image registration and normalization processing are coarse and easily affected by image quality.

[0003] Current solutions struggle to adaptively balance loss weights in multi-task collaborative optimization, and the fusion of shared features and task-specific features is insufficient. Effective fusion of multimodal fundus images and improvement of model robustness in small-sample scenarios are needed, while also considering classification accuracy and inference efficiency to meet actual clinical diagnostic needs.

[0004] Therefore, there is an urgent need to propose a glaucoma image classification method and system that integrates few-sample multi-task learning. Summary of the Invention

[0005] This application provides a glaucoma image classification method and system based on multi-task learning to solve the above-mentioned problems.

[0006] On the one hand, this application provides a glaucoma image classification method based on multi-task learning, the method comprising the following steps: Step S1: Obtain the multimodal glaucoma fundus image dataset of the target object, and perform normalization and registration preprocessing on the dataset to obtain a standardized glaucoma image sample set; Step S2: Based on the dual objectives of glaucoma disease classification and lesion region segmentation, construct a multi-task learning classification model that includes a shared feature extraction subnetwork and a task-specific branch subnetwork; Step S3: Introduce a multi-task joint loss constraint mechanism, weightedly fuse the prediction losses of the two tasks, and use an adaptive gradient descent algorithm to iteratively train the multi-task learning classification model until the model converges; Step S4: Input the glaucoma image to be classified into the trained multi-task learning classification model, and simultaneously output the glaucoma disease grading result and the lesion region segmentation classification result to complete the multi-task classification of glaucoma images.

[0007] In one implementation of this application, the normalization and registration preprocessing in step S1 specifically includes: linearly normalizing the pixel values ​​of the multimodal fundus images to the [0,1] interval, using a rigid registration algorithm based on SIFT feature points to eliminate spatial offset between different modal images, and simultaneously removing blurry or overexposed invalid samples through sharpness scoring.

[0008] In one implementation of this application, the calculation formula for the multi-task joint loss function in step S3 is as follows:

[0009] Among them, L total For the total joint loss of multiple tasks, L cls For the cross-entropy loss of the glaucoma disease grading task, L seg Dice loss for optic disc / optic cup lesion segmentation task; For the classification task loss weight coefficient, , , The loss weight coefficients for the segmentation task are given, and satisfy the following conditions are met. .

[0010] In one implementation of this application, the feature mapping calculation formula for the shared feature extraction sub-network in step S2 is as follows:

[0011] Among them, F shared The output is the shared depth feature map, X is the input of the normalized glaucoma image sample obtained in step S1, and W is the output. conv Let b be the weight matrix of the convolutional layer. conv This is the bias vector for the convolutional layer. For ReLU non-linear activation function, * is the convolution operator.

[0012] In one implementation of this application, a task-specific branch subnetwork embeds a task attention mechanism, and the formula for calculating its attention weight allocation is as follows:

[0013] Among them, A task W is a task-specific attention weight matrix. att Let b be the weight parameter matrix of the attention layer. att Here, represents the bias parameter of the attention layer, and sigmoid is the activation function that normalizes the weights to the [0,1] interval.

[0014] In one implementation of this application, the formula for calculating the probability of glaucoma classification output by the model in step S4 is as follows:

[0015] Among them, P cls F represents the multi-class probability vector for glaucoma: normal, early / intermediate / late stage. task W is the task-specific feature vector output by the task-specific branch. fc Let b be the weight matrix of the fully connected layer. fc is the bias vector of the fully connected layer, and Softmax is the multi-class normalization exponential function.

[0016] In one implementation of this application, the adaptive gradient descent algorithm employs the AdamW optimizer, and a cosine learning rate decay strategy is added during training, with the initial learning rate set to 10. 4 After each round of full-sample training, the learning rate gradually decreases to 10% according to a cosine function. 6 .

[0017] In one implementation of this application, the multimodal glaucoma fundus image dataset in step S1 includes optical coherence tomography (OCT) images and color fundus photography (CFP) images; the dataset is augmented by random rotation, horizontal flipping, and Gaussian noise addition to expand the sample size to 3 times that of the original dataset.

[0018] In one implementation of this application, when preprocessing the glaucoma image to be classified in step S4, image pyramid multi-scale sampling is used to generate input features of three scales: 256×256, 512×512, and 1024×1024. After fusing the multi-scale features, the input features are fed into the model to complete the classification inference.

[0019] On the other hand, this application also provides a glaucoma image classification system based on multi-task learning. The system includes: a data preprocessing module, used to acquire a multimodal glaucoma fundus image dataset of the target object, and perform normalization and registration preprocessing to obtain a standardized glaucoma image sample set. The model building module is used to construct a multi-task learning classification model based on the dual tasks of glaucoma disease classification and lesion region segmentation, which includes a shared feature extraction subnetwork and a task-specific branch subnetwork. The model training module is used to introduce a multi-task joint loss constraint mechanism, which weights and fuses the prediction losses of the two tasks, and uses an adaptive gradient descent algorithm to iteratively train the multi-task learning classification model until convergence. The classification reasoning module is used to input the glaucoma image to be classified into the trained multi-task learning classification model, and simultaneously output the glaucoma disease grading results and lesion region segmentation classification results.

[0020] This application provides a glaucoma image classification method and system based on multi-task learning, which has the following beneficial effects: 1. By constructing a multi-task learning classification model that includes a shared feature extraction subnetwork and a task-specific branch subnetwork, and combining the dual task objectives of glaucoma disease grading and lesion region segmentation, the grading results and lesion segmentation results are output simultaneously. This avoids the shortcomings of insufficient feature reuse and low training efficiency of single-task models, improves the comprehensiveness and practicality of glaucoma image classification, and adapts to the needs of simultaneous clinical diagnosis.

[0021] 2. By performing pixel normalization, SIFT feature point rigid registration preprocessing, and data augmentation such as random rotation and horizontal flipping on multimodal fundus images to expand the sample size to 3 times, we achieved unified image scale, eliminated spatial offset, removed invalid samples, alleviated the small sample dilemma, improved dataset quality and model robustness, and provided reliable data support for accurate classification.

[0022] 3. By introducing a multi-task joint loss constraint mechanism to weight and fuse dual-task losses, and using the AdamW optimizer and cosine learning rate decay strategy, combined with multi-scale feature fusion inference of image pyramid, the model achieves stable and efficient convergence, balances dual-task optimization conflicts, improves disease grading accuracy and lesion segmentation integrity, and enhances the model's adaptability to different imaging conditions. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of a glaucoma image classification method based on multi-task learning is provided for embodiments of this application; Figure 2 This is a diagram illustrating the components of a glaucoma image classification system based on multi-task learning, as provided in an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] This application provides a glaucoma image classification method and system based on multi-task learning. The technical solution proposed in this application will be described in detail below with reference to the accompanying drawings.

[0026] Figure 1 This document provides a flowchart of a glaucoma image classification method based on multi-task learning, as an embodiment of this application. Figure 1 As shown, the method mainly includes the following steps: Step S1: Obtain the multimodal glaucoma fundus image dataset of the target object, and perform normalization and registration preprocessing on the dataset to obtain a standardized glaucoma image sample set; Step S2: Based on the dual objectives of glaucoma disease classification and lesion region segmentation, construct a multi-task learning classification model that includes a shared feature extraction subnetwork and a task-specific branch subnetwork; Step S3: Introduce a multi-task joint loss constraint mechanism, weightedly fuse the prediction losses of the two tasks, and use an adaptive gradient descent algorithm to iteratively train the multi-task learning classification model until the model converges; Step S4: Input the glaucoma image to be classified into the trained multi-task learning classification model, and simultaneously output the glaucoma disease grading result and the lesion region segmentation classification result to complete the multi-task classification of glaucoma images.

[0027] Furthermore, the normalization and registration preprocessing described in step S1 specifically includes: linearly normalizing the pixel values ​​of the multimodal fundus images to the [0,1] interval, using a rigid registration algorithm based on SIFT feature points to eliminate spatial offsets between different modal images, and simultaneously removing blurry or overexposed invalid samples through sharpness scoring. The advantages of this approach are: 1. Pixel normalization to the [0,1] interval unifies image brightness and contrast scales, avoids numerical differences interfering with model training, and improves convergence efficiency; 2. Rigid SIFT feature point registration eliminates spatial offsets between multimodal images, ensuring alignment of data from different modalities and enhancing the effectiveness of feature fusion; 3. Removing blurry or overexposed invalid samples reduces noise interference, improves dataset quality, and lays a solid data foundation for accurate classification and segmentation in subsequent models.

[0028] Furthermore, the calculation formula for the multi-task joint loss function mentioned in step S3 is as follows:

[0029] Among them, L total For the total joint loss of multiple tasks, L cls For the cross-entropy loss of the glaucoma disease grading task, L seg Dice loss for optic disc / optic cup lesion segmentation task; For the classification task loss weight coefficient, , , The loss weight coefficients for the segmentation task are given, and satisfy the following conditions are met. The cross-entropy loss of the hierarchical task and the Dice loss of the segmentation task are weighted and fused, and the weight coefficients satisfy the following conditions: The system precisely balances the priorities of dual-task optimization. Cross-entropy loss adapts to the multi-classification requirements of disease grading, while Dice loss adapts to the region matching requirements of lesion segmentation. The two work together to constrain model parameter updates, avoiding biased loss in a single task, allowing the model to simultaneously improve grading accuracy and segmentation completeness, ensuring the consistency and reliability of multi-task outputs, and meeting the accuracy requirements of simultaneous clinical diagnosis.

[0030] Furthermore, the feature mapping calculation formula for the shared feature extraction sub-network in step S2 is as follows:

[0031] Among them, F shared The output is the shared depth feature map, X is the input of the normalized glaucoma image sample obtained in step S1, and W is the output. conv Let b be the weight matrix of the convolutional layer. conv This is the bias vector for the convolutional layer. The formula uses the ReLU non-linear activation function, with * representing the convolution operator. This formula combines standardized input X with convolution operations to efficiently extract common deep features from multimodal fundus images, enabling feature reuse, avoiding repetitive feature learning in dual tasks, and improving training efficiency. The ReLU non-linear activation function suppresses invalid features and alleviates gradient vanishing, allowing the model to focus on key lesion information; the weight matrix W... conv With bias vector b conv The learnability of the features is adapted to the diversity of glaucoma images, the generalization ability of shared features is enhanced, and high-quality basic features are provided for the subsequent dual-task dedicated branches, ensuring the accuracy of classification and segmentation.

[0032] Furthermore, the task-specific branch subnetwork embeds a task attention mechanism, and its attention weight allocation calculation formula is as follows:

[0033] Among them, A task W is a task-specific attention weight matrix. att Let b be the weight parameter matrix of the attention layer. att Here, W represents the bias parameters of the attention layer, and Sigmoid is the activation function that normalizes the weights to the [0,1] interval. This mechanism utilizes learnable W... att With b attThe parameters, combined with Sigmoid activation, normalize the weights to the [0,1] range, precisely allocating task-specific attention. This allows branch subnetworks to focus on the core features of grading and segmentation (e.g., grading focuses on the severity of lesions, segmentation focuses on the optic disc / optic cup region), suppressing interference from irrelevant information. Simultaneously, it adapts to the different needs of dual tasks, strengthening the transformation of shared features into task-specific features, improving the targeting and accuracy of dual-task outputs, and ensuring precise coordination of grading and segmentation results.

[0034] Furthermore, the formula for calculating the probability of glaucoma classification output by the model in step S4 is as follows:

[0035] Among them, P cls F represents the multi-class probability vector for glaucoma: normal, early / intermediate / late stage. task W is the task-specific feature vector output by the task-specific branch. fc Let b be the weight matrix of the fully connected layer. fc The bias vector is the fully connected layer bias vector, and Softmax is the multi-class normalization exponential function. This formula uses the Ftask output from the task-specific branch as the core input, ensuring the relevance of the hierarchical features and avoiding interference from irrelevant information. Learnable WFC and BFC optimize the mapping of features to probabilities, accurately adapting to normal, early / intermediate / late stage multi-class scenarios; Softmax normalizes the output to the [0,1] interval, and the probability sum is 1, intuitively presenting the probability of each level, which not only facilitates doctors to quickly judge the severity of the condition, but also improves the interpretability of the results, making the hierarchical output both accurate and practical, and adapting to the needs of clinical diagnostic decision-making.

[0036] Furthermore, the adaptive gradient descent algorithm employs the AdamW optimizer, and a cosine learning rate decay strategy is incorporated during training, with the initial learning rate set to 10. 4 After each round of full-sample training, the learning rate gradually decreases to 10% according to a cosine function. 6 The advantages of this setup are: the AdamW optimizer achieves effective regularization through weight decay, suppressing model overfitting and improving generalization ability; the cosine learning rate decay strategy aligns with training patterns, with an initial learning rate of 10. -4 Adapts to quickly explore the optimal solution, gradually decreasing to 10. -6 This approach avoids late-stage oscillations and accurately approximates the global optimum. The combination of these two methods makes multi-task model training more stable, balances the optimization conflicts between the two tasks, accelerates convergence while ensuring classification and segmentation accuracy, and adapts to the reliability requirements of clinical diagnosis.

[0037] Furthermore, the multimodal glaucoma fundus image dataset mentioned in step S1 includes optical coherence tomography (OCT) images and color fundus photography (CFP) images. Data augmentation is performed on this dataset using random rotation, horizontal flipping, and Gaussian noise addition to expand the sample size to three times that of the original dataset. The advantages of this setup are: OCT images and CFP images are complementary, providing information on the deep structure and surface morphology of the fundus, respectively, allowing the model to acquire more comprehensive lesion features; data augmentation expands the sample size to three times through three methods, addressing the challenge of small sample sizes and enriching data diversity, thus enhancing the model's adaptability to image variations. The combination of these two methods effectively avoids overfitting, improves the model's generalization ability and classification / segmentation accuracy, and adapts to the diagnostic needs under different imaging conditions in clinical practice.

[0038] Furthermore, in step S4, when preprocessing the glaucoma images to be classified, multi-scale sampling using an image pyramid is employed to generate input features at three scales: 256×256, 512×512, and 1024×1024. These multi-scale features are then fused and input into the model to complete the classification inference. The advantage of this setup is that the small-scale 256×256 captures the global structure, the medium-scale 512×512 balances the overall picture with local details, and the large-scale 1024×1024 focuses on subtle lesions. Multi-scale fusion compensates for the limitations of a single scale. This approach adapts to lesion regions of different sizes, comprehensively perceives image information, improves the accuracy of disease grading and the completeness of lesion segmentation, enhances the model's adaptability to changes in image scale, better meets the dual needs of clinical diagnosis for both detail and overall picture, and strengthens the reliability of classification inference. The above describes a glaucoma image classification method based on multi-task learning, as provided in this application. Based on the same inventive concept, this application also provides a glaucoma image classification system based on multi-task learning. Figure 2 A diagram illustrating the composition of a glaucoma image classification system based on multi-task learning, as provided in this application embodiment, is shown below. Figure 2 As shown, the system mainly includes: a data preprocessing module 201, used to acquire a multimodal glaucoma fundus image dataset of the target object, and perform normalization and registration preprocessing to obtain a standardized glaucoma image sample set; Model building module 202 is used to build a multi-task learning classification model based on the dual task objectives of glaucoma disease classification and lesion region segmentation, which includes a shared feature extraction subnetwork and a task-specific branch subnetwork. The model training module 203 is used to introduce a multi-task joint loss constraint mechanism, which weights and fuses the prediction losses of the two tasks, and uses an adaptive gradient descent algorithm to iteratively train the multi-task learning classification model until convergence. The classification reasoning module 204 is used to input the glaucoma image to be classified into the trained multi-task learning classification model and simultaneously output the glaucoma disease grading result and the lesion region segmentation classification result.

[0039] Below is an example of a specific application scenario for this application. A county-level hospital is conducting a free glaucoma screening program for people aged 45 and above. It needs to complete the diagnostic assessment of 500 high-risk individuals within one week, facing challenges such as a large sample size, limited medical resources, and difficulty in identifying early lesions. The method and system described in this application are implemented as follows: Data Acquisition and Preprocessing: 500 optical coherence tomography (OCT) images and 500 color fundus photography (CFP) images were acquired from 500 subjects using ophthalmology equipment, resulting in a total of 1000 images in the original dataset. The data preprocessing module performed the following steps: ① Pixel values ​​were linearly normalized to the [0,1] interval; ② The SIFT feature point rigid registration algorithm was used to align the dual-modal images; ③ 32 blurry or overexposed samples were removed using a sharpness score (threshold set at 0.7, values ​​below this were considered invalid), resulting in 968 valid original samples. Subsequent data augmentation was performed using random rotation, horizontal flipping, and Gaussian noise addition, expanding the dataset according to the rule of "original samples × 3," ultimately generating a standardized sample set of 968 × 3 = 2904 images, providing sufficient data support for model training.

[0040] Model Building and Training: The model building module constructs a multi-task learning model. A shared feature extraction sub-network extracts common features through convolutional operations, while task-specific branches embed attention mechanisms. The model training module employs the AdamW optimizer, combined with a cosine learning rate decay strategy. The learning rate calculation formula is as follows: (Where n is the current iteration round number, and the total number of iteration rounds is 80), for example, in the 1st round... Round 40 Round 80 During training, the multi-task joint loss is calculated according to the formula... .

[0041] Calculate (assuming) ,satisfy If Lcls=0.12 and Lseg=0.09 in a certain training round, then After 80 iterations, the total joint loss of the model stabilized below 0.08, meeting the convergence criterion.

[0042] Classification Reasoning and Clinical Application: Fundus images of the subjects to be screened are input into the classification reasoning module. Multi-scale sampling through image pyramids generates and fuses three scale features: 256×256, 512×512, and 1024×1024. Reasoning for a single image takes 0.3 seconds; for 500 subjects, the total reasoning time is 500 × 0.3 = 150 seconds (2.5 minutes). Traditional manual diagnosis requires an average of 3 minutes per person, totaling 25 hours for 500 subjects. This system is 3 times more efficient than the traditional method (25 hours ÷ 8.33 hours ≈ 3; based on an 8-hour workday, the system only needs 1.04 days to complete). System synchronous output results: 28 subjects were diagnosed with early-stage glaucoma, 12 with mid-to-late-stage glaucoma, and the optic disc / optic cup lesion segmentation results for 40 patients were verified using the Dice coefficient. For example, the intersection of a patient's segmented region A and the gold standard B is 120 pixels, and their union is 130 pixels. All models meet the high-precision requirement of Dice ≥ 0.92. Primary care physicians can quickly develop treatment plans by combining grading results with lesion segmentation maps, improving the early detection rate by 18% compared to single-task models, effectively addressing the shortage of diagnostic resources at the primary care level.

[0043] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0044] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0045] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A glaucoma image classification method based on multi-task learning, characterized in that, The method includes the following steps: Step S1: Obtain the multimodal glaucoma fundus image dataset of the target object, and perform normalization and registration preprocessing on the dataset to obtain a standardized glaucoma image sample set; Step S2: Based on the dual objectives of glaucoma disease classification and lesion region segmentation, construct a multi-task learning classification model that includes a shared feature extraction subnetwork and a task-specific branch subnetwork; Step S3: Introduce a multi-task joint loss constraint mechanism, weightedly fuse the prediction losses of the two tasks, and use an adaptive gradient descent algorithm to iteratively train the multi-task learning classification model until the model converges; Step S4: Input the glaucoma image to be classified into the trained multi-task learning classification model, and simultaneously output the glaucoma disease grading result and the lesion region segmentation classification result to complete the multi-task classification of glaucoma images.

2. The glaucoma image classification method based on multi-task learning according to claim 1, characterized in that, The normalization and registration preprocessing described in step S1 specifically includes: linearly normalizing the pixel values ​​of the multimodal fundus images to the [0,1] interval, using a rigid registration algorithm based on SIFT feature points to eliminate spatial offset between different modal images, and removing blurry or overexposed invalid samples through sharpness scoring.

3. The glaucoma image classification method based on multi-task learning according to claim 1, characterized in that, In step S3, the formula for calculating the loss function of the multi-task joint loss constraint mechanism is as follows: Among them, L total For the total joint loss of multiple tasks, L cls For the cross-entropy loss of the glaucoma disease grading task, L seg Dice loss for optic disc / optic cup lesion segmentation task; For the classification task loss weight coefficient, , , The loss weight coefficients for the segmentation task are given, and satisfy the following conditions are met. .

4. The glaucoma image classification method based on multi-task learning according to claim 1, characterized in that, The feature mapping calculation formula for the shared feature extraction sub-network in step S2 is as follows: Among them, F shared The output is the shared depth feature map, X is the input of the normalized glaucoma image sample obtained in step S1, and W is the output. conv Let b be the weight matrix of the convolutional layer. conv This is the bias vector for the convolutional layer. For ReLU non-linear activation function, * is the convolution operator.

5. The glaucoma image classification method based on multi-task learning according to claim 1, characterized in that, The task-specific branch subnetwork embeds a task attention mechanism, and its attention weight allocation calculation formula is as follows: Among them, A task W is a task-specific attention weight matrix. att Let b be the weight parameter matrix of the attention layer. att Here, represents the bias parameter of the attention layer, and sigmoid is the activation function that normalizes the weights to the [0,1] interval.

6. The glaucoma image classification method based on multi-task learning according to claim 1, characterized in that, The formula for calculating the probability of glaucoma classification output by the model in step S4 is as follows: Among them, P cls F represents the multi-class probability vector for glaucoma: normal, early / intermediate / late stage. task W is the task-specific feature vector output by the task-specific branch. fc Let b be the weight matrix of the fully connected layer. fc is the bias vector of the fully connected layer, and Softmax is the multi-class normalization exponential function.

7. The glaucoma image classification method based on multi-task learning according to claim 1, characterized in that, The adaptive gradient descent algorithm employs the AdamW optimizer, and a cosine learning rate decay strategy is incorporated during training, with the initial learning rate set to 10. 4 After each round of full-sample training, the learning rate gradually decreases to 10% according to a cosine function. 6 .

8. The glaucoma image classification method based on multi-task learning according to claim 1, characterized in that, The multimodal glaucoma fundus image dataset mentioned in step S1 includes optical coherence tomography (OCT) images and color fundus photography (CFP) images. The dataset is augmented by random rotation, horizontal flipping, and Gaussian noise addition to expand the sample size to 3 times that of the original dataset.

9. A glaucoma image classification method based on multi-task learning according to claim 1, characterized in that, Step S4: When preprocessing the glaucoma images to be classified, multi-scale sampling of the image pyramid is used to generate input features at three scales: 256×256, 512×512, and 1024×1024. After fusing the multi-scale features, the input is fed into the model to complete the classification inference.

10. A glaucoma image classification system based on multi-task learning, characterized in that, The system includes: a data preprocessing module, used to acquire a multimodal glaucoma fundus image dataset of the target object, and perform normalization and registration preprocessing to obtain a standardized glaucoma image sample set; The model building module is used to construct a multi-task learning classification model based on the dual tasks of glaucoma disease classification and lesion region segmentation, which includes a shared feature extraction subnetwork and a task-specific branch subnetwork. The model training module is used to introduce a multi-task joint loss constraint mechanism, which weights and fuses the prediction losses of the two tasks, and uses an adaptive gradient descent algorithm to iteratively train the multi-task learning classification model until convergence. The classification reasoning module is used to input the glaucoma image to be classified into the trained multi-task learning classification model, and simultaneously output the glaucoma disease grading results and lesion region segmentation classification results.

Images